Add files via upload

utils/__init__.py

@@ -0,0 +1,80 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+utils/initialization
+"""
+
+import contextlib
+import platform
+import threading
+
+
+def emojis(str=''):
+    # Return platform-dependent emoji-safe version of string
+    return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str
+
+
+class TryExcept(contextlib.ContextDecorator):
+    # YOLOv5 TryExcept class. Usage: @TryExcept() decorator or 'with TryExcept():' context manager
+    def __init__(self, msg=''):
+        self.msg = msg
+
+    def __enter__(self):
+        pass
+
+    def __exit__(self, exc_type, value, traceback):
+        if value:
+            print(emojis(f"{self.msg}{': ' if self.msg else ''}{value}"))
+        return True
+
+
+def threaded(func):
+    # Multi-threads a target function and returns thread. Usage: @threaded decorator
+    def wrapper(*args, **kwargs):
+        thread = threading.Thread(target=func, args=args, kwargs=kwargs, daemon=True)
+        thread.start()
+        return thread
+
+    return wrapper
+
+
+def join_threads(verbose=False):
+    # Join all daemon threads, i.e. atexit.register(lambda: join_threads())
+    main_thread = threading.current_thread()
+    for t in threading.enumerate():
+        if t is not main_thread:
+            if verbose:
+                print(f'Joining thread {t.name}')
+            t.join()
+
+
+def notebook_init(verbose=True):
+    # Check system software and hardware
+    print('Checking setup...')
+
+    import os
+    import shutil
+
+    from utils.general import check_font, check_requirements, is_colab
+    from utils.torch_utils import select_device  # imports
+
+    check_font()
+
+    import psutil
+    from IPython import display  # to display images and clear console output
+
+    if is_colab():
+        shutil.rmtree('/content/sample_data', ignore_errors=True)  # remove colab /sample_data directory
+
+    # System info
+    if verbose:
+        gb = 1 << 30  # bytes to GiB (1024 ** 3)
+        ram = psutil.virtual_memory().total
+        total, used, free = shutil.disk_usage("/")
+        display.clear_output()
+        s = f'({os.cpu_count()} CPUs, {ram / gb:.1f} GB RAM, {(total - free) / gb:.1f}/{total / gb:.1f} GB disk)'
+    else:
+        s = ''
+
+    select_device(newline=False)
+    print(emojis(f'Setup complete ✅ {s}'))
+    return display

utils/activations.py

@@ -0,0 +1,103 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Activation functions
+"""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class SiLU(nn.Module):
+    # SiLU activation https://arxiv.org/pdf/1606.08415.pdf
+    @staticmethod
+    def forward(x):
+        return x * torch.sigmoid(x)
+
+
+class Hardswish(nn.Module):
+    # Hard-SiLU activation
+    @staticmethod
+    def forward(x):
+        # return x * F.hardsigmoid(x)  # for TorchScript and CoreML
+        return x * F.hardtanh(x + 3, 0.0, 6.0) / 6.0  # for TorchScript, CoreML and ONNX
+
+
+class Mish(nn.Module):
+    # Mish activation https://github.com/digantamisra98/Mish
+    @staticmethod
+    def forward(x):
+        return x * F.softplus(x).tanh()
+
+
+class MemoryEfficientMish(nn.Module):
+    # Mish activation memory-efficient
+    class F(torch.autograd.Function):
+
+        @staticmethod
+        def forward(ctx, x):
+            ctx.save_for_backward(x)
+            return x.mul(torch.tanh(F.softplus(x)))  # x * tanh(ln(1 + exp(x)))
+
+        @staticmethod
+        def backward(ctx, grad_output):
+            x = ctx.saved_tensors[0]
+            sx = torch.sigmoid(x)
+            fx = F.softplus(x).tanh()
+            return grad_output * (fx + x * sx * (1 - fx * fx))
+
+    def forward(self, x):
+        return self.F.apply(x)
+
+
+class FReLU(nn.Module):
+    # FReLU activation https://arxiv.org/abs/2007.11824
+    def __init__(self, c1, k=3):  # ch_in, kernel
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False)
+        self.bn = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        return torch.max(x, self.bn(self.conv(x)))
+
+
+class AconC(nn.Module):
+    r""" ACON activation (activate or not)
+    AconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is a learnable parameter
+    according to "Activate or Not: Learning Customized Activation" <https://arxiv.org/pdf/2009.04759.pdf>.
+    """
+
+    def __init__(self, c1):
+        super().__init__()
+        self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.beta = nn.Parameter(torch.ones(1, c1, 1, 1))
+
+    def forward(self, x):
+        dpx = (self.p1 - self.p2) * x
+        return dpx * torch.sigmoid(self.beta * dpx) + self.p2 * x
+
+
+class MetaAconC(nn.Module):
+    r""" ACON activation (activate or not)
+    MetaAconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is generated by a small network
+    according to "Activate or Not: Learning Customized Activation" <https://arxiv.org/pdf/2009.04759.pdf>.
+    """
+
+    def __init__(self, c1, k=1, s=1, r=16):  # ch_in, kernel, stride, r
+        super().__init__()
+        c2 = max(r, c1 // r)
+        self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1))
+        self.fc1 = nn.Conv2d(c1, c2, k, s, bias=True)
+        self.fc2 = nn.Conv2d(c2, c1, k, s, bias=True)
+        # self.bn1 = nn.BatchNorm2d(c2)
+        # self.bn2 = nn.BatchNorm2d(c1)
+
+    def forward(self, x):
+        y = x.mean(dim=2, keepdims=True).mean(dim=3, keepdims=True)
+        # batch-size 1 bug/instabilities https://github.com/ultralytics/yolov5/issues/2891
+        # beta = torch.sigmoid(self.bn2(self.fc2(self.bn1(self.fc1(y)))))  # bug/unstable
+        beta = torch.sigmoid(self.fc2(self.fc1(y)))  # bug patch BN layers removed
+        dpx = (self.p1 - self.p2) * x
+        return dpx * torch.sigmoid(beta * dpx) + self.p2 * x

utils/augmentations.py

@@ -0,0 +1,397 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Image augmentation functions
+"""
+
+import math
+import random
+
+import cv2
+import numpy as np
+import torch
+import torchvision.transforms as T
+import torchvision.transforms.functional as TF
+
+from utils.general import LOGGER, check_version, colorstr, resample_segments, segment2box, xywhn2xyxy
+from utils.metrics import bbox_ioa
+
+IMAGENET_MEAN = 0.485, 0.456, 0.406  # RGB mean
+IMAGENET_STD = 0.229, 0.224, 0.225  # RGB standard deviation
+
+
+class Albumentations:
+    # YOLOv5 Albumentations class (optional, only used if package is installed)
+    def __init__(self, size=640):
+        self.transform = None
+        prefix = colorstr('albumentations: ')
+        try:
+            import albumentations as A
+            check_version(A.__version__, '1.0.3', hard=True)  # version requirement
+
+            T = [
+                A.RandomResizedCrop(height=size, width=size, scale=(0.8, 1.0), ratio=(0.9, 1.11), p=0.0),
+                A.Blur(p=0.01),
+                A.MedianBlur(p=0.01),
+                A.ToGray(p=0.01),
+                A.CLAHE(p=0.01),
+                A.RandomBrightnessContrast(p=0.0),
+                A.RandomGamma(p=0.0),
+                A.ImageCompression(quality_lower=75, p=0.0)]  # transforms
+            self.transform = A.Compose(T, bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))
+
+            LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))
+        except ImportError:  # package not installed, skip
+            pass
+        except Exception as e:
+            LOGGER.info(f'{prefix}{e}')
+
+    def __call__(self, im, labels, p=1.0):
+        if self.transform and random.random() < p:
+            new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0])  # transformed
+            im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
+        return im, labels
+
+
+def normalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD, inplace=False):
+    # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = (x - mean) / std
+    return TF.normalize(x, mean, std, inplace=inplace)
+
+
+def denormalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD):
+    # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = x * std + mean
+    for i in range(3):
+        x[:, i] = x[:, i] * std[i] + mean[i]
+    return x
+
+
+def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):
+    # HSV color-space augmentation
+    if hgain or sgain or vgain:
+        r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
+        hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))
+        dtype = im.dtype  # uint8
+
+        x = np.arange(0, 256, dtype=r.dtype)
+        lut_hue = ((x * r[0]) % 180).astype(dtype)
+        lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
+        lut_val = np.clip(x * r[2], 0, 255).astype(dtype)
+
+        im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
+        cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im)  # no return needed
+
+
+def hist_equalize(im, clahe=True, bgr=False):
+    # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-255
+    yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)
+    if clahe:
+        c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
+        yuv[:, :, 0] = c.apply(yuv[:, :, 0])
+    else:
+        yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0])  # equalize Y channel histogram
+    return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB)  # convert YUV image to RGB
+
+
+def replicate(im, labels):
+    # Replicate labels
+    h, w = im.shape[:2]
+    boxes = labels[:, 1:].astype(int)
+    x1, y1, x2, y2 = boxes.T
+    s = ((x2 - x1) + (y2 - y1)) / 2  # side length (pixels)
+    for i in s.argsort()[:round(s.size * 0.5)]:  # smallest indices
+        x1b, y1b, x2b, y2b = boxes[i]
+        bh, bw = y2b - y1b, x2b - x1b
+        yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw))  # offset x, y
+        x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]
+        im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b]  # im4[ymin:ymax, xmin:xmax]
+        labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)
+
+    return im, labels
+
+
+def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):
+    # Resize and pad image while meeting stride-multiple constraints
+    shape = im.shape[:2]  # current shape [height, width]
+    if isinstance(new_shape, int):
+        new_shape = (new_shape, new_shape)
+
+    # Scale ratio (new / old)
+    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+    if not scaleup:  # only scale down, do not scale up (for better val mAP)
+        r = min(r, 1.0)
+
+    # Compute padding
+    ratio = r, r  # width, height ratios
+    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
+    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
+    if auto:  # minimum rectangle
+        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
+    elif scaleFill:  # stretch
+        dw, dh = 0.0, 0.0
+        new_unpad = (new_shape[1], new_shape[0])
+        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios
+
+    dw /= 2  # divide padding into 2 sides
+    dh /= 2
+
+    if shape[::-1] != new_unpad:  # resize
+        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
+    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+    im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
+    return im, ratio, (dw, dh)
+
+
+def random_perspective(im,
+                       targets=(),
+                       segments=(),
+                       degrees=10,
+                       translate=.1,
+                       scale=.1,
+                       shear=10,
+                       perspective=0.0,
+                       border=(0, 0)):
+    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 10))
+    # targets = [cls, xyxy]
+
+    height = im.shape[0] + border[0] * 2  # shape(h,w,c)
+    width = im.shape[1] + border[1] * 2
+
+    # Center
+    C = np.eye(3)
+    C[0, 2] = -im.shape[1] / 2  # x translation (pixels)
+    C[1, 2] = -im.shape[0] / 2  # y translation (pixels)
+
+    # Perspective
+    P = np.eye(3)
+    P[2, 0] = random.uniform(-perspective, perspective)  # x perspective (about y)
+    P[2, 1] = random.uniform(-perspective, perspective)  # y perspective (about x)
+
+    # Rotation and Scale
+    R = np.eye(3)
+    a = random.uniform(-degrees, degrees)
+    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
+    s = random.uniform(1 - scale, 1 + scale)
+    # s = 2 ** random.uniform(-scale, scale)
+    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
+
+    # Shear
+    S = np.eye(3)
+    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
+    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)
+
+    # Translation
+    T = np.eye(3)
+    T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width  # x translation (pixels)
+    T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height  # y translation (pixels)
+
+    # Combined rotation matrix
+    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
+    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
+        if perspective:
+            im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))
+        else:  # affine
+            im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
+
+    # Visualize
+    # import matplotlib.pyplot as plt
+    # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
+    # ax[0].imshow(im[:, :, ::-1])  # base
+    # ax[1].imshow(im2[:, :, ::-1])  # warped
+
+    # Transform label coordinates
+    n = len(targets)
+    if n:
+        use_segments = any(x.any() for x in segments)
+        new = np.zeros((n, 4))
+        if use_segments:  # warp segments
+            segments = resample_segments(segments)  # upsample
+            for i, segment in enumerate(segments):
+                xy = np.ones((len(segment), 3))
+                xy[:, :2] = segment
+                xy = xy @ M.T  # transform
+                xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]  # perspective rescale or affine
+
+                # clip
+                new[i] = segment2box(xy, width, height)
+
+        else:  # warp boxes
+            xy = np.ones((n * 4, 3))
+            xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
+            xy = xy @ M.T  # transform
+            xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine
+
+            # create new boxes
+            x = xy[:, [0, 2, 4, 6]]
+            y = xy[:, [1, 3, 5, 7]]
+            new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T
+
+            # clip
+            new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)
+            new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)
+
+        # filter candidates
+        i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)
+        targets = targets[i]
+        targets[:, 1:5] = new[i]
+
+    return im, targets
+
+
+def copy_paste(im, labels, segments, p=0.5):
+    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
+    n = len(segments)
+    if p and n:
+        h, w, c = im.shape  # height, width, channels
+        im_new = np.zeros(im.shape, np.uint8)
+        for j in random.sample(range(n), k=round(p * n)):
+            l, s = labels[j], segments[j]
+            box = w - l[3], l[2], w - l[1], l[4]
+            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
+            if (ioa < 0.30).all():  # allow 30% obscuration of existing labels
+                labels = np.concatenate((labels, [[l[0], *box]]), 0)
+                segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))
+                cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (1, 1, 1), cv2.FILLED)
+
+        result = cv2.flip(im, 1)  # augment segments (flip left-right)
+        i = cv2.flip(im_new, 1).astype(bool)
+        im[i] = result[i]  # cv2.imwrite('debug.jpg', im)  # debug
+
+    return im, labels, segments
+
+
+def cutout(im, labels, p=0.5):
+    # Applies image cutout augmentation https://arxiv.org/abs/1708.04552
+    if random.random() < p:
+        h, w = im.shape[:2]
+        scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16  # image size fraction
+        for s in scales:
+            mask_h = random.randint(1, int(h * s))  # create random masks
+            mask_w = random.randint(1, int(w * s))
+
+            # box
+            xmin = max(0, random.randint(0, w) - mask_w // 2)
+            ymin = max(0, random.randint(0, h) - mask_h // 2)
+            xmax = min(w, xmin + mask_w)
+            ymax = min(h, ymin + mask_h)
+
+            # apply random color mask
+            im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]
+
+            # return unobscured labels
+            if len(labels) and s > 0.03:
+                box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
+                ioa = bbox_ioa(box, xywhn2xyxy(labels[:, 1:5], w, h))  # intersection over area
+                labels = labels[ioa < 0.60]  # remove >60% obscured labels
+
+    return labels
+
+
+def mixup(im, labels, im2, labels2):
+    # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
+    r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
+    im = (im * r + im2 * (1 - r)).astype(np.uint8)
+    labels = np.concatenate((labels, labels2), 0)
+    return im, labels
+
+
+def box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16):  # box1(4,n), box2(4,n)
+    # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio
+    w1, h1 = box1[2] - box1[0], box1[3] - box1[1]
+    w2, h2 = box2[2] - box2[0], box2[3] - box2[1]
+    ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps))  # aspect ratio
+    return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr)  # candidates
+
+
+def classify_albumentations(
+        augment=True,
+        size=224,
+        scale=(0.08, 1.0),
+        ratio=(0.75, 1.0 / 0.75),  # 0.75, 1.33
+        hflip=0.5,
+        vflip=0.0,
+        jitter=0.4,
+        mean=IMAGENET_MEAN,
+        std=IMAGENET_STD,
+        auto_aug=False):
+    # YOLOv5 classification Albumentations (optional, only used if package is installed)
+    prefix = colorstr('albumentations: ')
+    try:
+        import albumentations as A
+        from albumentations.pytorch import ToTensorV2
+        check_version(A.__version__, '1.0.3', hard=True)  # version requirement
+        if augment:  # Resize and crop
+            T = [A.RandomResizedCrop(height=size, width=size, scale=scale, ratio=ratio)]
+            if auto_aug:
+                # TODO: implement AugMix, AutoAug & RandAug in albumentation
+                LOGGER.info(f'{prefix}auto augmentations are currently not supported')
+            else:
+                if hflip > 0:
+                    T += [A.HorizontalFlip(p=hflip)]
+                if vflip > 0:
+                    T += [A.VerticalFlip(p=vflip)]
+                if jitter > 0:
+                    color_jitter = (float(jitter),) * 3  # repeat value for brightness, contrast, satuaration, 0 hue
+                    T += [A.ColorJitter(*color_jitter, 0)]
+        else:  # Use fixed crop for eval set (reproducibility)
+            T = [A.SmallestMaxSize(max_size=size), A.CenterCrop(height=size, width=size)]
+        T += [A.Normalize(mean=mean, std=std), ToTensorV2()]  # Normalize and convert to Tensor
+        LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))
+        return A.Compose(T)
+
+    except ImportError:  # package not installed, skip
+        LOGGER.warning(f'{prefix}⚠️ not found, install with `pip install albumentations` (recommended)')
+    except Exception as e:
+        LOGGER.info(f'{prefix}{e}')
+
+
+def classify_transforms(size=224):
+    # Transforms to apply if albumentations not installed
+    assert isinstance(size, int), f'ERROR: classify_transforms size {size} must be integer, not (list, tuple)'
+    # T.Compose([T.ToTensor(), T.Resize(size), T.CenterCrop(size), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])
+    return T.Compose([CenterCrop(size), ToTensor(), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])
+
+
+class LetterBox:
+    # YOLOv5 LetterBox class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])
+    def __init__(self, size=(640, 640), auto=False, stride=32):
+        super().__init__()
+        self.h, self.w = (size, size) if isinstance(size, int) else size
+        self.auto = auto  # pass max size integer, automatically solve for short side using stride
+        self.stride = stride  # used with auto
+
+    def __call__(self, im):  # im = np.array HWC
+        imh, imw = im.shape[:2]
+        r = min(self.h / imh, self.w / imw)  # ratio of new/old
+        h, w = round(imh * r), round(imw * r)  # resized image
+        hs, ws = (math.ceil(x / self.stride) * self.stride for x in (h, w)) if self.auto else self.h, self.w
+        top, left = round((hs - h) / 2 - 0.1), round((ws - w) / 2 - 0.1)
+        im_out = np.full((self.h, self.w, 3), 114, dtype=im.dtype)
+        im_out[top:top + h, left:left + w] = cv2.resize(im, (w, h), interpolation=cv2.INTER_LINEAR)
+        return im_out
+
+
+class CenterCrop:
+    # YOLOv5 CenterCrop class for image preprocessing, i.e. T.Compose([CenterCrop(size), ToTensor()])
+    def __init__(self, size=640):
+        super().__init__()
+        self.h, self.w = (size, size) if isinstance(size, int) else size
+
+    def __call__(self, im):  # im = np.array HWC
+        imh, imw = im.shape[:2]
+        m = min(imh, imw)  # min dimension
+        top, left = (imh - m) // 2, (imw - m) // 2
+        return cv2.resize(im[top:top + m, left:left + m], (self.w, self.h), interpolation=cv2.INTER_LINEAR)
+
+
+class ToTensor:
+    # YOLOv5 ToTensor class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])
+    def __init__(self, half=False):
+        super().__init__()
+        self.half = half
+
+    def __call__(self, im):  # im = np.array HWC in BGR order
+        im = np.ascontiguousarray(im.transpose((2, 0, 1))[::-1])  # HWC to CHW -> BGR to RGB -> contiguous
+        im = torch.from_numpy(im)  # to torch
+        im = im.half() if self.half else im.float()  # uint8 to fp16/32
+        im /= 255.0  # 0-255 to 0.0-1.0
+        return im

utils/autoanchor.py

@@ -0,0 +1,169 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+AutoAnchor utils
+"""
+
+import random
+
+import numpy as np
+import torch
+import yaml
+from tqdm import tqdm
+
+from utils import TryExcept
+from utils.general import LOGGER, TQDM_BAR_FORMAT, colorstr
+
+PREFIX = colorstr('AutoAnchor: ')
+
+
+def check_anchor_order(m):
+    # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
+    a = m.anchors.prod(-1).mean(-1).view(-1)  # mean anchor area per output layer
+    da = a[-1] - a[0]  # delta a
+    ds = m.stride[-1] - m.stride[0]  # delta s
+    if da and (da.sign() != ds.sign()):  # same order
+        LOGGER.info(f'{PREFIX}Reversing anchor order')
+        m.anchors[:] = m.anchors.flip(0)
+
+
+@TryExcept(f'{PREFIX}ERROR')
+def check_anchors(dataset, model, thr=4.0, imgsz=640):
+    # Check anchor fit to data, recompute if necessary
+    m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1]  # Detect()
+    shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1))  # augment scale
+    wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float()  # wh
+
+    def metric(k):  # compute metric
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1 / r).min(2)[0]  # ratio metric
+        best = x.max(1)[0]  # best_x
+        aat = (x > 1 / thr).float().sum(1).mean()  # anchors above threshold
+        bpr = (best > 1 / thr).float().mean()  # best possible recall
+        return bpr, aat
+
+    stride = m.stride.to(m.anchors.device).view(-1, 1, 1)  # model strides
+    anchors = m.anchors.clone() * stride  # current anchors
+    bpr, aat = metric(anchors.cpu().view(-1, 2))
+    s = f'\n{PREFIX}{aat:.2f} anchors/target, {bpr:.3f} Best Possible Recall (BPR). '
+    if bpr > 0.98:  # threshold to recompute
+        LOGGER.info(f'{s}Current anchors are a good fit to dataset ✅')
+    else:
+        LOGGER.info(f'{s}Anchors are a poor fit to dataset ⚠️, attempting to improve...')
+        na = m.anchors.numel() // 2  # number of anchors
+        anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False)
+        new_bpr = metric(anchors)[0]
+        if new_bpr > bpr:  # replace anchors
+            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)
+            m.anchors[:] = anchors.clone().view_as(m.anchors)
+            check_anchor_order(m)  # must be in pixel-space (not grid-space)
+            m.anchors /= stride
+            s = f'{PREFIX}Done ✅ (optional: update model *.yaml to use these anchors in the future)'
+        else:
+            s = f'{PREFIX}Done ⚠️ (original anchors better than new anchors, proceeding with original anchors)'
+        LOGGER.info(s)
+
+
+def kmean_anchors(dataset='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True):
+    """ Creates kmeans-evolved anchors from training dataset
+
+        Arguments:
+            dataset: path to data.yaml, or a loaded dataset
+            n: number of anchors
+            img_size: image size used for training
+            thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0
+            gen: generations to evolve anchors using genetic algorithm
+            verbose: print all results
+
+        Return:
+            k: kmeans evolved anchors
+
+        Usage:
+            from utils.autoanchor import *; _ = kmean_anchors()
+    """
+    from scipy.cluster.vq import kmeans
+
+    npr = np.random
+    thr = 1 / thr
+
+    def metric(k, wh):  # compute metrics
+        r = wh[:, None] / k[None]
+        x = torch.min(r, 1 / r).min(2)[0]  # ratio metric
+        # x = wh_iou(wh, torch.tensor(k))  # iou metric
+        return x, x.max(1)[0]  # x, best_x
+
+    def anchor_fitness(k):  # mutation fitness
+        _, best = metric(torch.tensor(k, dtype=torch.float32), wh)
+        return (best * (best > thr).float()).mean()  # fitness
+
+    def print_results(k, verbose=True):
+        k = k[np.argsort(k.prod(1))]  # sort small to large
+        x, best = metric(k, wh0)
+        bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n  # best possible recall, anch > thr
+        s = f'{PREFIX}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr\n' \
+            f'{PREFIX}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' \
+            f'past_thr={x[x > thr].mean():.3f}-mean: '
+        for x in k:
+            s += '%i,%i, ' % (round(x[0]), round(x[1]))
+        if verbose:
+            LOGGER.info(s[:-2])
+        return k
+
+    if isinstance(dataset, str):  # *.yaml file
+        with open(dataset, errors='ignore') as f:
+            data_dict = yaml.safe_load(f)  # model dict
+        from utils.dataloaders import LoadImagesAndLabels
+        dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True)
+
+    # Get label wh
+    shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True)
+    wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)])  # wh
+
+    # Filter
+    i = (wh0 < 3.0).any(1).sum()
+    if i:
+        LOGGER.info(f'{PREFIX}WARNING ⚠️ Extremely small objects found: {i} of {len(wh0)} labels are <3 pixels in size')
+    wh = wh0[(wh0 >= 2.0).any(1)].astype(np.float32)  # filter > 2 pixels
+    # wh = wh * (npr.rand(wh.shape[0], 1) * 0.9 + 0.1)  # multiply by random scale 0-1
+
+    # Kmeans init
+    try:
+        LOGGER.info(f'{PREFIX}Running kmeans for {n} anchors on {len(wh)} points...')
+        assert n <= len(wh)  # apply overdetermined constraint
+        s = wh.std(0)  # sigmas for whitening
+        k = kmeans(wh / s, n, iter=30)[0] * s  # points
+        assert n == len(k)  # kmeans may return fewer points than requested if wh is insufficient or too similar
+    except Exception:
+        LOGGER.warning(f'{PREFIX}WARNING ⚠️ switching strategies from kmeans to random init')
+        k = np.sort(npr.rand(n * 2)).reshape(n, 2) * img_size  # random init
+    wh, wh0 = (torch.tensor(x, dtype=torch.float32) for x in (wh, wh0))
+    k = print_results(k, verbose=False)
+
+    # Plot
+    # k, d = [None] * 20, [None] * 20
+    # for i in tqdm(range(1, 21)):
+    #     k[i-1], d[i-1] = kmeans(wh / s, i)  # points, mean distance
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True)
+    # ax = ax.ravel()
+    # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.')
+    # fig, ax = plt.subplots(1, 2, figsize=(14, 7))  # plot wh
+    # ax[0].hist(wh[wh[:, 0]<100, 0],400)
+    # ax[1].hist(wh[wh[:, 1]<100, 1],400)
+    # fig.savefig('wh.png', dpi=200)
+
+    # Evolve
+    f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1  # fitness, generations, mutation prob, sigma
+    pbar = tqdm(range(gen), bar_format=TQDM_BAR_FORMAT)  # progress bar
+    for _ in pbar:
+        v = np.ones(sh)
+        while (v == 1).all():  # mutate until a change occurs (prevent duplicates)
+            v = ((npr.random(sh) < mp) * random.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0)
+        kg = (k.copy() * v).clip(min=2.0)
+        fg = anchor_fitness(kg)
+        if fg > f:
+            f, k = fg, kg.copy()
+            pbar.desc = f'{PREFIX}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}'
+            if verbose:
+                print_results(k, verbose)
+
+    return print_results(k).astype(np.float32)

utils/autobatch.py

@@ -0,0 +1,72 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Auto-batch utils
+"""
+
+from copy import deepcopy
+
+import numpy as np
+import torch
+
+from utils.general import LOGGER, colorstr
+from utils.torch_utils import profile
+
+
+def check_train_batch_size(model, imgsz=640, amp=True):
+    # Check YOLOv5 training batch size
+    with torch.cuda.amp.autocast(amp):
+        return autobatch(deepcopy(model).train(), imgsz)  # compute optimal batch size
+
+
+def autobatch(model, imgsz=640, fraction=0.8, batch_size=16):
+    # Automatically estimate best YOLOv5 batch size to use `fraction` of available CUDA memory
+    # Usage:
+    #     import torch
+    #     from utils.autobatch import autobatch
+    #     model = torch.hub.load('ultralytics/yolov5', 'yolov5s', autoshape=False)
+    #     print(autobatch(model))
+
+    # Check device
+    prefix = colorstr('AutoBatch: ')
+    LOGGER.info(f'{prefix}Computing optimal batch size for --imgsz {imgsz}')
+    device = next(model.parameters()).device  # get model device
+    if device.type == 'cpu':
+        LOGGER.info(f'{prefix}CUDA not detected, using default CPU batch-size {batch_size}')
+        return batch_size
+    if torch.backends.cudnn.benchmark:
+        LOGGER.info(f'{prefix} ⚠️ Requires torch.backends.cudnn.benchmark=False, using default batch-size {batch_size}')
+        return batch_size
+
+    # Inspect CUDA memory
+    gb = 1 << 30  # bytes to GiB (1024 ** 3)
+    d = str(device).upper()  # 'CUDA:0'
+    properties = torch.cuda.get_device_properties(device)  # device properties
+    t = properties.total_memory / gb  # GiB total
+    r = torch.cuda.memory_reserved(device) / gb  # GiB reserved
+    a = torch.cuda.memory_allocated(device) / gb  # GiB allocated
+    f = t - (r + a)  # GiB free
+    LOGGER.info(f'{prefix}{d} ({properties.name}) {t:.2f}G total, {r:.2f}G reserved, {a:.2f}G allocated, {f:.2f}G free')
+
+    # Profile batch sizes
+    batch_sizes = [1, 2, 4, 8, 16]
+    try:
+        img = [torch.empty(b, 3, imgsz, imgsz) for b in batch_sizes]
+        results = profile(img, model, n=3, device=device)
+    except Exception as e:
+        LOGGER.warning(f'{prefix}{e}')
+
+    # Fit a solution
+    y = [x[2] for x in results if x]  # memory [2]
+    p = np.polyfit(batch_sizes[:len(y)], y, deg=1)  # first degree polynomial fit
+    b = int((f * fraction - p[1]) / p[0])  # y intercept (optimal batch size)
+    if None in results:  # some sizes failed
+        i = results.index(None)  # first fail index
+        if b >= batch_sizes[i]:  # y intercept above failure point
+            b = batch_sizes[max(i - 1, 0)]  # select prior safe point
+    if b < 1 or b > 1024:  # b outside of safe range
+        b = batch_size
+        LOGGER.warning(f'{prefix}WARNING ⚠️ CUDA anomaly detected, recommend restart environment and retry command.')
+
+    fraction = (np.polyval(p, b) + r + a) / t  # actual fraction predicted
+    LOGGER.info(f'{prefix}Using batch-size {b} for {d} {t * fraction:.2f}G/{t:.2f}G ({fraction * 100:.0f}%) ✅')
+    return b

utils/callbacks.py

@@ -0,0 +1,76 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Callback utils
+"""
+
+import threading
+
+
+class Callbacks:
+    """"
+    Handles all registered callbacks for YOLOv5 Hooks
+    """
+
+    def __init__(self):
+        # Define the available callbacks
+        self._callbacks = {
+            'on_pretrain_routine_start': [],
+            'on_pretrain_routine_end': [],
+            'on_train_start': [],
+            'on_train_epoch_start': [],
+            'on_train_batch_start': [],
+            'optimizer_step': [],
+            'on_before_zero_grad': [],
+            'on_train_batch_end': [],
+            'on_train_epoch_end': [],
+            'on_val_start': [],
+            'on_val_batch_start': [],
+            'on_val_image_end': [],
+            'on_val_batch_end': [],
+            'on_val_end': [],
+            'on_fit_epoch_end': [],  # fit = train + val
+            'on_model_save': [],
+            'on_train_end': [],
+            'on_params_update': [],
+            'teardown': [],}
+        self.stop_training = False  # set True to interrupt training
+
+    def register_action(self, hook, name='', callback=None):
+        """
+        Register a new action to a callback hook
+
+        Args:
+            hook: The callback hook name to register the action to
+            name: The name of the action for later reference
+            callback: The callback to fire
+        """
+        assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}"
+        assert callable(callback), f"callback '{callback}' is not callable"
+        self._callbacks[hook].append({'name': name, 'callback': callback})
+
+    def get_registered_actions(self, hook=None):
+        """"
+        Returns all the registered actions by callback hook
+
+        Args:
+            hook: The name of the hook to check, defaults to all
+        """
+        return self._callbacks[hook] if hook else self._callbacks
+
+    def run(self, hook, *args, thread=False, **kwargs):
+        """
+        Loop through the registered actions and fire all callbacks on main thread
+
+        Args:
+            hook: The name of the hook to check, defaults to all
+            args: Arguments to receive from YOLOv5
+            thread: (boolean) Run callbacks in daemon thread
+            kwargs: Keyword Arguments to receive from YOLOv5
+        """
+
+        assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}"
+        for logger in self._callbacks[hook]:
+            if thread:
+                threading.Thread(target=logger['callback'], args=args, kwargs=kwargs, daemon=True).start()
+            else:
+                logger['callback'](*args, **kwargs)

utils/downloads.py

@@ -0,0 +1,108 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Download utils
+"""
+
+import logging
+import os
+import subprocess
+import urllib
+from pathlib import Path
+
+import requests
+import torch
+
+
+def is_url(url, check=True):
+    # Check if string is URL and check if URL exists
+    try:
+        url = str(url)
+        result = urllib.parse.urlparse(url)
+        assert all([result.scheme, result.netloc])  # check if is url
+        return (urllib.request.urlopen(url).getcode() == 200) if check else True  # check if exists online
+    except (AssertionError, urllib.request.HTTPError):
+        return False
+
+
+def gsutil_getsize(url=''):
+    # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du
+    s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8')
+    return eval(s.split(' ')[0]) if len(s) else 0  # bytes
+
+
+def url_getsize(url='https://ultralytics.com/images/bus.jpg'):
+    # Return downloadable file size in bytes
+    response = requests.head(url, allow_redirects=True)
+    return int(response.headers.get('content-length', -1))
+
+
+def safe_download(file, url, url2=None, min_bytes=1E0, error_msg=''):
+    # Attempts to download file from url or url2, checks and removes incomplete downloads < min_bytes
+    from utils.general import LOGGER
+
+    file = Path(file)
+    assert_msg = f"Downloaded file '{file}' does not exist or size is < min_bytes={min_bytes}"
+    try:  # url1
+        LOGGER.info(f'Downloading {url} to {file}...')
+        torch.hub.download_url_to_file(url, str(file), progress=LOGGER.level <= logging.INFO)
+        assert file.exists() and file.stat().st_size > min_bytes, assert_msg  # check
+    except Exception as e:  # url2
+        if file.exists():
+            file.unlink()  # remove partial downloads
+        LOGGER.info(f'ERROR: {e}\nRe-attempting {url2 or url} to {file}...')
+        os.system(f"curl -# -L '{url2 or url}' -o '{file}' --retry 3 -C -")  # curl download, retry and resume on fail
+    finally:
+        if not file.exists() or file.stat().st_size < min_bytes:  # check
+            if file.exists():
+                file.unlink()  # remove partial downloads
+            LOGGER.info(f"ERROR: {assert_msg}\n{error_msg}")
+        LOGGER.info('')
+
+
+def attempt_download(file, repo='ultralytics/yolov5', release='v7.0'):
+    # Attempt file download from GitHub release assets if not found locally. release = 'latest', 'v7.0', etc.
+    from utils.general import LOGGER
+
+    def github_assets(repository, version='latest'):
+        # Return GitHub repo tag (i.e. 'v7.0') and assets (i.e. ['yolov5s.pt', 'yolov5m.pt', ...])
+        if version != 'latest':
+            version = f'tags/{version}'  # i.e. tags/v7.0
+        response = requests.get(f'https://api.github.com/repos/{repository}/releases/{version}').json()  # github api
+        return response['tag_name'], [x['name'] for x in response['assets']]  # tag, assets
+
+    file = Path(str(file).strip().replace("'", ''))
+    if not file.exists():
+        # URL specified
+        name = Path(urllib.parse.unquote(str(file))).name  # decode '%2F' to '/' etc.
+        if str(file).startswith(('http:/', 'https:/')):  # download
+            url = str(file).replace(':/', '://')  # Pathlib turns :// -> :/
+            file = name.split('?')[0]  # parse authentication https://url.com/file.txt?auth...
+            if Path(file).is_file():
+                LOGGER.info(f'Found {url} locally at {file}')  # file already exists
+            else:
+                safe_download(file=file, url=url, min_bytes=1E5)
+            return file
+
+        # GitHub assets
+        assets = [f'yolov5{size}{suffix}.pt' for size in 'nsmlx' for suffix in ('', '6', '-cls', '-seg')]  # default
+        try:
+            tag, assets = github_assets(repo, release)
+        except Exception:
+            try:
+                tag, assets = github_assets(repo)  # latest release
+            except Exception:
+                try:
+                    tag = subprocess.check_output('git tag', shell=True, stderr=subprocess.STDOUT).decode().split()[-1]
+                except Exception:
+                    tag = release
+
+        file.parent.mkdir(parents=True, exist_ok=True)  # make parent dir (if required)
+        if name in assets:
+            url3 = 'https://drive.google.com/drive/folders/1EFQTEUeXWSFww0luse2jB9M1QNZQGwNl'  # backup gdrive mirror
+            safe_download(
+                file,
+                url=f'https://github.com/{repo}/releases/download/{tag}/{name}',
+                min_bytes=1E5,
+                error_msg=f'{file} missing, try downloading from https://github.com/{repo}/releases/{tag} or {url3}')
+
+    return str(file)

utils/loss.py

@@ -0,0 +1,234 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Loss functions
+"""
+
+import torch
+import torch.nn as nn
+
+from utils.metrics import bbox_iou
+from utils.torch_utils import de_parallel
+
+
+def smooth_BCE(eps=0.1):  # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441
+    # return positive, negative label smoothing BCE targets
+    return 1.0 - 0.5 * eps, 0.5 * eps
+
+
+class BCEBlurWithLogitsLoss(nn.Module):
+    # BCEwithLogitLoss() with reduced missing label effects.
+    def __init__(self, alpha=0.05):
+        super().__init__()
+        self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none')  # must be nn.BCEWithLogitsLoss()
+        self.alpha = alpha
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+        pred = torch.sigmoid(pred)  # prob from logits
+        dx = pred - true  # reduce only missing label effects
+        # dx = (pred - true).abs()  # reduce missing label and false label effects
+        alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4))
+        loss *= alpha_factor
+        return loss.mean()
+
+
+class FocalLoss(nn.Module):
+    # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
+    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
+        super().__init__()
+        self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = loss_fcn.reduction
+        self.loss_fcn.reduction = 'none'  # required to apply FL to each element
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+        # p_t = torch.exp(-loss)
+        # loss *= self.alpha * (1.000001 - p_t) ** self.gamma  # non-zero power for gradient stability
+
+        # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py
+        pred_prob = torch.sigmoid(pred)  # prob from logits
+        p_t = true * pred_prob + (1 - true) * (1 - pred_prob)
+        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
+        modulating_factor = (1.0 - p_t) ** self.gamma
+        loss *= alpha_factor * modulating_factor
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:  # 'none'
+            return loss
+
+
+class QFocalLoss(nn.Module):
+    # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5)
+    def __init__(self, loss_fcn, gamma=1.5, alpha=0.25):
+        super().__init__()
+        self.loss_fcn = loss_fcn  # must be nn.BCEWithLogitsLoss()
+        self.gamma = gamma
+        self.alpha = alpha
+        self.reduction = loss_fcn.reduction
+        self.loss_fcn.reduction = 'none'  # required to apply FL to each element
+
+    def forward(self, pred, true):
+        loss = self.loss_fcn(pred, true)
+
+        pred_prob = torch.sigmoid(pred)  # prob from logits
+        alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha)
+        modulating_factor = torch.abs(true - pred_prob) ** self.gamma
+        loss *= alpha_factor * modulating_factor
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:  # 'none'
+            return loss
+
+
+class ComputeLoss:
+    sort_obj_iou = False
+
+    # Compute losses
+    def __init__(self, model, autobalance=False):
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        m = de_parallel(model).model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(m.nl, [4.0, 1.0, 0.25, 0.06, 0.02])  # P3-P7
+        self.ssi = list(m.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance
+        self.na = m.na  # number of anchors
+        self.nc = m.nc  # number of classes
+        self.nl = m.nl  # number of layers
+        self.anchors = m.anchors
+        self.device = device
+
+    def __call__(self, p, targets):  # predictions, targets
+        lcls = torch.zeros(1, device=self.device)  # class loss
+        lbox = torch.zeros(1, device=self.device)  # box loss
+        lobj = torch.zeros(1, device=self.device)  # object loss
+        tcls, tbox, indices, anchors = self.build_targets(p, targets)  # targets
+
+        # Losses
+        for i, pi in enumerate(p):  # layer index, layer predictions
+            b, a, gj, gi = indices[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros(pi.shape[:4], dtype=pi.dtype, device=self.device)  # target obj
+
+            n = b.shape[0]  # number of targets
+            if n:
+                # pxy, pwh, _, pcls = pi[b, a, gj, gi].tensor_split((2, 4, 5), dim=1)  # faster, requires torch 1.8.0
+                pxy, pwh, _, pcls = pi[b, a, gj, gi].split((2, 2, 1, self.nc), 1)  # target-subset of predictions
+
+                # Regression
+                pxy = pxy.sigmoid() * 2 - 0.5
+                pwh = (pwh.sigmoid() * 2) ** 2 * anchors[i]
+                pbox = torch.cat((pxy, pwh), 1)  # predicted box
+                iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze()  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Objectness
+                iou = iou.detach().clamp(0).type(tobj.dtype)
+                if self.sort_obj_iou:
+                    j = iou.argsort()
+                    b, a, gj, gi, iou = b[j], a[j], gj[j], gi[j], iou[j]
+                if self.gr < 1:
+                    iou = (1.0 - self.gr) + self.gr * iou
+                tobj[b, a, gj, gi] = iou  # iou ratio
+
+                # Classification
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(pcls, self.cn, device=self.device)  # targets
+                    t[range(n), tcls[i]] = self.cp
+                    lcls += self.BCEcls(pcls, t)  # BCE
+
+                # Append targets to text file
+                # with open('targets.txt', 'a') as file:
+                #     [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)]
+
+            obji = self.BCEobj(pi[..., 4], tobj)
+            lobj += obji * self.balance[i]  # obj loss
+            if self.autobalance:
+                self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
+
+        if self.autobalance:
+            self.balance = [x / self.balance[self.ssi] for x in self.balance]
+        lbox *= self.hyp['box']
+        lobj *= self.hyp['obj']
+        lcls *= self.hyp['cls']
+        bs = tobj.shape[0]  # batch size
+
+        return (lbox + lobj + lcls) * bs, torch.cat((lbox, lobj, lcls)).detach()
+
+    def build_targets(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        tcls, tbox, indices, anch = [], [], [], []
+        gain = torch.ones(7, device=self.device)  # normalized to gridspace gain
+        ai = torch.arange(na, device=self.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[..., None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor(
+            [
+                [0, 0],
+                [1, 0],
+                [0, 1],
+                [-1, 0],
+                [0, -1],  # j,k,l,m
+                # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+            ],
+            device=self.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors, shape = self.anchors[i], p[i].shape
+            gain[2:6] = torch.tensor(shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain  # shape(3,n,7)
+            if nt:
+                # Matches
+                r = t[..., 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1 < g) & (gxy > 1)).T
+                l, m = ((gxi % 1 < g) & (gxi > 1)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            bc, gxy, gwh, a = t.chunk(4, 1)  # (image, class), grid xy, grid wh, anchors
+            a, (b, c) = a.long().view(-1), bc.long().T  # anchors, image, class
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid indices
+
+            # Append
+            indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1)))  # image, anchor, grid
+            tbox.append(torch.cat((gxy - gij, gwh), 1))  # box
+            anch.append(anchors[a])  # anchors
+            tcls.append(c)  # class
+
+        return tcls, tbox, indices, anch

utils/metrics.py

@@ -0,0 +1,360 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Model validation metrics
+"""
+
+import math
+import warnings
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import numpy as np
+import torch
+
+from utils import TryExcept, threaded
+
+
+def fitness(x):
+    # Model fitness as a weighted combination of metrics
+    w = [0.0, 0.0, 0.1, 0.9]  # weights for [P, R, mAP@0.5, mAP@0.5:0.95]
+    return (x[:, :4] * w).sum(1)
+
+
+def smooth(y, f=0.05):
+    # Box filter of fraction f
+    nf = round(len(y) * f * 2) // 2 + 1  # number of filter elements (must be odd)
+    p = np.ones(nf // 2)  # ones padding
+    yp = np.concatenate((p * y[0], y, p * y[-1]), 0)  # y padded
+    return np.convolve(yp, np.ones(nf) / nf, mode='valid')  # y-smoothed
+
+
+def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16, prefix=""):
+    """ Compute the average precision, given the recall and precision curves.
+    Source: https://github.com/rafaelpadilla/Object-Detection-Metrics.
+    # Arguments
+        tp:  True positives (nparray, nx1 or nx10).
+        conf:  Objectness value from 0-1 (nparray).
+        pred_cls:  Predicted object classes (nparray).
+        target_cls:  True object classes (nparray).
+        plot:  Plot precision-recall curve at mAP@0.5
+        save_dir:  Plot save directory
+    # Returns
+        The average precision as computed in py-faster-rcnn.
+    """
+
+    # Sort by objectness
+    i = np.argsort(-conf)
+    tp, conf, pred_cls = tp[i], conf[i], pred_cls[i]
+
+    # Find unique classes
+    unique_classes, nt = np.unique(target_cls, return_counts=True)
+    nc = unique_classes.shape[0]  # number of classes, number of detections
+
+    # Create Precision-Recall curve and compute AP for each class
+    px, py = np.linspace(0, 1, 1000), []  # for plotting
+    ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000))
+    for ci, c in enumerate(unique_classes):
+        i = pred_cls == c
+        n_l = nt[ci]  # number of labels
+        n_p = i.sum()  # number of predictions
+        if n_p == 0 or n_l == 0:
+            continue
+
+        # Accumulate FPs and TPs
+        fpc = (1 - tp[i]).cumsum(0)
+        tpc = tp[i].cumsum(0)
+
+        # Recall
+        recall = tpc / (n_l + eps)  # recall curve
+        r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0)  # negative x, xp because xp decreases
+
+        # Precision
+        precision = tpc / (tpc + fpc)  # precision curve
+        p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1)  # p at pr_score
+
+        # AP from recall-precision curve
+        for j in range(tp.shape[1]):
+            ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j])
+            if plot and j == 0:
+                py.append(np.interp(px, mrec, mpre))  # precision at mAP@0.5
+
+    # Compute F1 (harmonic mean of precision and recall)
+    f1 = 2 * p * r / (p + r + eps)
+    names = [v for k, v in names.items() if k in unique_classes]  # list: only classes that have data
+    names = dict(enumerate(names))  # to dict
+    if plot:
+        plot_pr_curve(px, py, ap, Path(save_dir) / f'{prefix}PR_curve.png', names)
+        plot_mc_curve(px, f1, Path(save_dir) / f'{prefix}F1_curve.png', names, ylabel='F1')
+        plot_mc_curve(px, p, Path(save_dir) / f'{prefix}P_curve.png', names, ylabel='Precision')
+        plot_mc_curve(px, r, Path(save_dir) / f'{prefix}R_curve.png', names, ylabel='Recall')
+
+    i = smooth(f1.mean(0), 0.1).argmax()  # max F1 index
+    p, r, f1 = p[:, i], r[:, i], f1[:, i]
+    tp = (r * nt).round()  # true positives
+    fp = (tp / (p + eps) - tp).round()  # false positives
+    return tp, fp, p, r, f1, ap, unique_classes.astype(int)
+
+
+def compute_ap(recall, precision):
+    """ Compute the average precision, given the recall and precision curves
+    # Arguments
+        recall:    The recall curve (list)
+        precision: The precision curve (list)
+    # Returns
+        Average precision, precision curve, recall curve
+    """
+
+    # Append sentinel values to beginning and end
+    mrec = np.concatenate(([0.0], recall, [1.0]))
+    mpre = np.concatenate(([1.0], precision, [0.0]))
+
+    # Compute the precision envelope
+    mpre = np.flip(np.maximum.accumulate(np.flip(mpre)))
+
+    # Integrate area under curve
+    method = 'interp'  # methods: 'continuous', 'interp'
+    if method == 'interp':
+        x = np.linspace(0, 1, 101)  # 101-point interp (COCO)
+        ap = np.trapz(np.interp(x, mrec, mpre), x)  # integrate
+    else:  # 'continuous'
+        i = np.where(mrec[1:] != mrec[:-1])[0]  # points where x axis (recall) changes
+        ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1])  # area under curve
+
+    return ap, mpre, mrec
+
+
+class ConfusionMatrix:
+    # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix
+    def __init__(self, nc, conf=0.25, iou_thres=0.45):
+        self.matrix = np.zeros((nc + 1, nc + 1))
+        self.nc = nc  # number of classes
+        self.conf = conf
+        self.iou_thres = iou_thres
+
+    def process_batch(self, detections, labels):
+        """
+        Return intersection-over-union (Jaccard index) of boxes.
+        Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+        Arguments:
+            detections (Array[N, 6]), x1, y1, x2, y2, conf, class
+            labels (Array[M, 5]), class, x1, y1, x2, y2
+        Returns:
+            None, updates confusion matrix accordingly
+        """
+        if detections is None:
+            gt_classes = labels.int()
+            for gc in gt_classes:
+                self.matrix[self.nc, gc] += 1  # background FN
+            return
+
+        detections = detections[detections[:, 4] > self.conf]
+        gt_classes = labels[:, 0].int()
+        detection_classes = detections[:, 5].int()
+        iou = box_iou(labels[:, 1:], detections[:, :4])
+
+        x = torch.where(iou > self.iou_thres)
+        if x[0].shape[0]:
+            matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy()
+            if x[0].shape[0] > 1:
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 1], return_index=True)[1]]
+                matches = matches[matches[:, 2].argsort()[::-1]]
+                matches = matches[np.unique(matches[:, 0], return_index=True)[1]]
+        else:
+            matches = np.zeros((0, 3))
+
+        n = matches.shape[0] > 0
+        m0, m1, _ = matches.transpose().astype(int)
+        for i, gc in enumerate(gt_classes):
+            j = m0 == i
+            if n and sum(j) == 1:
+                self.matrix[detection_classes[m1[j]], gc] += 1  # correct
+            else:
+                self.matrix[self.nc, gc] += 1  # true background
+
+        if n:
+            for i, dc in enumerate(detection_classes):
+                if not any(m1 == i):
+                    self.matrix[dc, self.nc] += 1  # predicted background
+
+    def tp_fp(self):
+        tp = self.matrix.diagonal()  # true positives
+        fp = self.matrix.sum(1) - tp  # false positives
+        # fn = self.matrix.sum(0) - tp  # false negatives (missed detections)
+        return tp[:-1], fp[:-1]  # remove background class
+
+    @TryExcept('WARNING ⚠️ ConfusionMatrix plot failure')
+    def plot(self, normalize=True, save_dir='', names=()):
+        import seaborn as sn
+
+        array = self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1E-9) if normalize else 1)  # normalize columns
+        array[array < 0.005] = np.nan  # don't annotate (would appear as 0.00)
+
+        fig, ax = plt.subplots(1, 1, figsize=(12, 9), tight_layout=True)
+        nc, nn = self.nc, len(names)  # number of classes, names
+        sn.set(font_scale=1.0 if nc < 50 else 0.8)  # for label size
+        labels = (0 < nn < 99) and (nn == nc)  # apply names to ticklabels
+        ticklabels = (names + ['background']) if labels else "auto"
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress empty matrix RuntimeWarning: All-NaN slice encountered
+            sn.heatmap(array,
+                       ax=ax,
+                       annot=nc < 30,
+                       annot_kws={
+                           "size": 8},
+                       cmap='Blues',
+                       fmt='.2f',
+                       square=True,
+                       vmin=0.0,
+                       xticklabels=ticklabels,
+                       yticklabels=ticklabels).set_facecolor((1, 1, 1))
+        ax.set_ylabel('True')
+        ax.set_ylabel('Predicted')
+        ax.set_title('Confusion Matrix')
+        fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250)
+        plt.close(fig)
+
+    def print(self):
+        for i in range(self.nc + 1):
+            print(' '.join(map(str, self.matrix[i])))
+
+
+def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7):
+    # Returns Intersection over Union (IoU) of box1(1,4) to box2(n,4)
+
+    # Get the coordinates of bounding boxes
+    if xywh:  # transform from xywh to xyxy
+        (x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, -1), box2.chunk(4, -1)
+        w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2
+        b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_
+        b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_
+    else:  # x1, y1, x2, y2 = box1
+        b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, -1)
+        b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, -1)
+        w1, h1 = b1_x2 - b1_x1, (b1_y2 - b1_y1).clamp(eps)
+        w2, h2 = b2_x2 - b2_x1, (b2_y2 - b2_y1).clamp(eps)
+
+    # Intersection area
+    inter = (b1_x2.minimum(b2_x2) - b1_x1.maximum(b2_x1)).clamp(0) * \
+            (b1_y2.minimum(b2_y2) - b1_y1.maximum(b2_y1)).clamp(0)
+
+    # Union Area
+    union = w1 * h1 + w2 * h2 - inter + eps
+
+    # IoU
+    iou = inter / union
+    if CIoU or DIoU or GIoU:
+        cw = b1_x2.maximum(b2_x2) - b1_x1.minimum(b2_x1)  # convex (smallest enclosing box) width
+        ch = b1_y2.maximum(b2_y2) - b1_y1.minimum(b2_y1)  # convex height
+        if CIoU or DIoU:  # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1
+            c2 = cw ** 2 + ch ** 2 + eps  # convex diagonal squared
+            rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4  # center dist ** 2
+            if CIoU:  # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47
+                v = (4 / math.pi ** 2) * (torch.atan(w2 / h2) - torch.atan(w1 / h1)).pow(2)
+                with torch.no_grad():
+                    alpha = v / (v - iou + (1 + eps))
+                return iou - (rho2 / c2 + v * alpha)  # CIoU
+            return iou - rho2 / c2  # DIoU
+        c_area = cw * ch + eps  # convex area
+        return iou - (c_area - union) / c_area  # GIoU https://arxiv.org/pdf/1902.09630.pdf
+    return iou  # IoU
+
+
+def box_iou(box1, box2, eps=1e-7):
+    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+    Arguments:
+        box1 (Tensor[N, 4])
+        box2 (Tensor[M, 4])
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+
+    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
+    (a1, a2), (b1, b2) = box1.unsqueeze(1).chunk(2, 2), box2.unsqueeze(0).chunk(2, 2)
+    inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp(0).prod(2)
+
+    # IoU = inter / (area1 + area2 - inter)
+    return inter / ((a2 - a1).prod(2) + (b2 - b1).prod(2) - inter + eps)
+
+
+def bbox_ioa(box1, box2, eps=1e-7):
+    """ Returns the intersection over box2 area given box1, box2. Boxes are x1y1x2y2
+    box1:       np.array of shape(4)
+    box2:       np.array of shape(nx4)
+    returns:    np.array of shape(n)
+    """
+
+    # Get the coordinates of bounding boxes
+    b1_x1, b1_y1, b1_x2, b1_y2 = box1
+    b2_x1, b2_y1, b2_x2, b2_y2 = box2.T
+
+    # Intersection area
+    inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \
+                 (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0)
+
+    # box2 area
+    box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + eps
+
+    # Intersection over box2 area
+    return inter_area / box2_area
+
+
+def wh_iou(wh1, wh2, eps=1e-7):
+    # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2
+    wh1 = wh1[:, None]  # [N,1,2]
+    wh2 = wh2[None]  # [1,M,2]
+    inter = torch.min(wh1, wh2).prod(2)  # [N,M]
+    return inter / (wh1.prod(2) + wh2.prod(2) - inter + eps)  # iou = inter / (area1 + area2 - inter)
+
+
+# Plots ----------------------------------------------------------------------------------------------------------------
+
+
+@threaded
+def plot_pr_curve(px, py, ap, save_dir=Path('pr_curve.png'), names=()):
+    # Precision-recall curve
+    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
+    py = np.stack(py, axis=1)
+
+    if 0 < len(names) < 21:  # display per-class legend if < 21 classes
+        for i, y in enumerate(py.T):
+            ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}')  # plot(recall, precision)
+    else:
+        ax.plot(px, py, linewidth=1, color='grey')  # plot(recall, precision)
+
+    ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean())
+    ax.set_xlabel('Recall')
+    ax.set_ylabel('Precision')
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    ax.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
+    ax.set_title('Precision-Recall Curve')
+    fig.savefig(save_dir, dpi=250)
+    plt.close(fig)
+
+
+@threaded
+def plot_mc_curve(px, py, save_dir=Path('mc_curve.png'), names=(), xlabel='Confidence', ylabel='Metric'):
+    # Metric-confidence curve
+    fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True)
+
+    if 0 < len(names) < 21:  # display per-class legend if < 21 classes
+        for i, y in enumerate(py):
+            ax.plot(px, y, linewidth=1, label=f'{names[i]}')  # plot(confidence, metric)
+    else:
+        ax.plot(px, py.T, linewidth=1, color='grey')  # plot(confidence, metric)
+
+    y = smooth(py.mean(0), 0.05)
+    ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}')
+    ax.set_xlabel(xlabel)
+    ax.set_ylabel(ylabel)
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    ax.legend(bbox_to_anchor=(1.04, 1), loc="upper left")
+    ax.set_title(f'{ylabel}-Confidence Curve')
+    fig.savefig(save_dir, dpi=250)
+    plt.close(fig)

utils/plots.py

@@ -0,0 +1,559 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Plotting utils
+"""
+
+import contextlib
+import math
+import os
+from copy import copy
+from pathlib import Path
+from urllib.error import URLError
+
+import cv2
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import seaborn as sn
+import torch
+from PIL import Image, ImageDraw, ImageFont
+
+from utils import TryExcept, threaded
+from utils.general import (CONFIG_DIR, FONT, LOGGER, check_font, check_requirements, clip_boxes, increment_path,
+                           is_ascii, xywh2xyxy, xyxy2xywh)
+from utils.metrics import fitness
+from utils.segment.general import scale_image
+
+# Settings
+RANK = int(os.getenv('RANK', -1))
+matplotlib.rc('font', **{'size': 11})
+#matplotlib.use('Agg')  # for writing to files only
+
+
+class Colors:
+    # Ultralytics color palette https://ultralytics.com/
+    def __init__(self):
+        # hex = matplotlib.colors.TABLEAU_COLORS.values()
+        hexs = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB',
+                '2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7')
+        self.palette = [self.hex2rgb(f'#{c}') for c in hexs]
+        self.n = len(self.palette)
+
+    def __call__(self, i, bgr=False):
+        c = self.palette[int(i) % self.n]
+        return (c[2], c[1], c[0]) if bgr else c
+
+    @staticmethod
+    def hex2rgb(h):  # rgb order (PIL)
+        return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
+
+
+colors = Colors()  # create instance for 'from utils.plots import colors'
+
+
+def check_pil_font(font=FONT, size=10):
+    # Return a PIL TrueType Font, downloading to CONFIG_DIR if necessary
+    font = Path(font)
+    font = font if font.exists() else (CONFIG_DIR / font.name)
+    try:
+        return ImageFont.truetype(str(font) if font.exists() else font.name, size)
+    except Exception:  # download if missing
+        try:
+            check_font(font)
+            return ImageFont.truetype(str(font), size)
+        except TypeError:
+            check_requirements('Pillow>=8.4.0')  # known issue https://github.com/ultralytics/yolov5/issues/5374
+        except URLError:  # not online
+            return ImageFont.load_default()
+
+
+class Annotator:
+    # YOLOv5 Annotator for train/val mosaics and jpgs and detect/hub inference annotations
+    def __init__(self, im, line_width=None, font_size=None, font='Arial.ttf', pil=False, example='abc'):
+        assert im.data.contiguous, 'Image not contiguous. Apply np.ascontiguousarray(im) to Annotator() input images.'
+        non_ascii = not is_ascii(example)  # non-latin labels, i.e. asian, arabic, cyrillic
+        self.pil = pil or non_ascii
+        if self.pil:  # use PIL
+            self.im = im if isinstance(im, Image.Image) else Image.fromarray(im)
+            self.draw = ImageDraw.Draw(self.im)
+            self.font = check_pil_font(font='Arial.Unicode.ttf' if non_ascii else font,
+                                       size=font_size or max(round(sum(self.im.size) / 2 * 0.035), 12))
+        else:  # use cv2
+            self.im = im
+        self.lw = line_width or max(round(sum(im.shape) / 2 * 0.003), 2)  # line width
+
+    def box_label(self, box, label='', color=(128, 128, 128), txt_color=(255, 255, 255)):
+        # Add one xyxy box to image with label
+        if self.pil or not is_ascii(label):
+            self.draw.rectangle(box, width=self.lw, outline=color)  # box
+            if label:
+                w, h = self.font.getsize(label)  # text width, height
+                outside = box[1] - h >= 0  # label fits outside box
+                self.draw.rectangle(
+                    (box[0], box[1] - h if outside else box[1], box[0] + w + 1,
+                     box[1] + 1 if outside else box[1] + h + 1),
+                    fill=color,
+                )
+                # self.draw.text((box[0], box[1]), label, fill=txt_color, font=self.font, anchor='ls')  # for PIL>8.0
+                self.draw.text((box[0], box[1] - h if outside else box[1]), label, fill=txt_color, font=self.font)
+        else:  # cv2
+            p1, p2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3]))
+            cv2.rectangle(self.im, p1, p2, color, thickness=self.lw, lineType=cv2.LINE_AA)
+            if label:
+                tf = max(self.lw - 1, 1)  # font thickness
+                w, h = cv2.getTextSize(label, 0, fontScale=self.lw / 3, thickness=tf)[0]  # text width, height
+                outside = p1[1] - h >= 3
+                p2 = p1[0] + w, p1[1] - h - 3 if outside else p1[1] + h + 3
+                cv2.rectangle(self.im, p1, p2, color, -1, cv2.LINE_AA)  # filled
+                cv2.putText(self.im,
+                            label, (p1[0], p1[1] - 2 if outside else p1[1] + h + 2),
+                            0,
+                            self.lw / 3,
+                            txt_color,
+                            thickness=tf,
+                            lineType=cv2.LINE_AA)
+
+    def masks(self, masks, colors, im_gpu, alpha=0.5, retina_masks=False):
+        """Plot masks at once.
+        Args:
+            masks (tensor): predicted masks on cuda, shape: [n, h, w]
+            colors (List[List[Int]]): colors for predicted masks, [[r, g, b] * n]
+            im_gpu (tensor): img is in cuda, shape: [3, h, w], range: [0, 1]
+            alpha (float): mask transparency: 0.0 fully transparent, 1.0 opaque
+        """
+        if self.pil:
+            # convert to numpy first
+            self.im = np.asarray(self.im).copy()
+        if len(masks) == 0:
+            self.im[:] = im_gpu.permute(1, 2, 0).contiguous().cpu().numpy() * 255
+        colors = torch.tensor(colors, device=im_gpu.device, dtype=torch.float32) / 255.0
+        colors = colors[:, None, None]  # shape(n,1,1,3)
+        masks = masks.unsqueeze(3)  # shape(n,h,w,1)
+        masks_color = masks * (colors * alpha)  # shape(n,h,w,3)
+
+        inv_alph_masks = (1 - masks * alpha).cumprod(0)  # shape(n,h,w,1)
+        mcs = (masks_color * inv_alph_masks).sum(0) * 2  # mask color summand shape(n,h,w,3)
+
+        im_gpu = im_gpu.flip(dims=[0])  # flip channel
+        im_gpu = im_gpu.permute(1, 2, 0).contiguous()  # shape(h,w,3)
+        im_gpu = im_gpu * inv_alph_masks[-1] + mcs
+        im_mask = (im_gpu * 255).byte().cpu().numpy()
+        self.im[:] = im_mask if retina_masks else scale_image(im_gpu.shape, im_mask, self.im.shape)
+        if self.pil:
+            # convert im back to PIL and update draw
+            self.fromarray(self.im)
+
+    def rectangle(self, xy, fill=None, outline=None, width=1):
+        # Add rectangle to image (PIL-only)
+        self.draw.rectangle(xy, fill, outline, width)
+
+    def text(self, xy, text, txt_color=(255, 255, 255), anchor='top'):
+        # Add text to image (PIL-only)
+        if anchor == 'bottom':  # start y from font bottom
+            w, h = self.font.getsize(text)  # text width, height
+            xy[1] += 1 - h
+        self.draw.text(xy, text, fill=txt_color, font=self.font)
+
+    def fromarray(self, im):
+        # Update self.im from a numpy array
+        self.im = im if isinstance(im, Image.Image) else Image.fromarray(im)
+        self.draw = ImageDraw.Draw(self.im)
+
+    def result(self):
+        # Return annotated image as array
+        return np.asarray(self.im)
+
+
+def feature_visualization(x, module_type, stage, n=32, save_dir=Path('runs/detect/exp')):
+    """
+    x:              Features to be visualized
+    module_type:    Module type
+    stage:          Module stage within model
+    n:              Maximum number of feature maps to plot
+    save_dir:       Directory to save results
+    """
+    if 'Detect' not in module_type:
+        batch, channels, height, width = x.shape  # batch, channels, height, width
+        if height > 1 and width > 1:
+            f = save_dir / f"stage{stage}_{module_type.split('.')[-1]}_features.png"  # filename
+
+            blocks = torch.chunk(x[0].cpu(), channels, dim=0)  # select batch index 0, block by channels
+            n = min(n, channels)  # number of plots
+            fig, ax = plt.subplots(math.ceil(n / 8), 8, tight_layout=True)  # 8 rows x n/8 cols
+            ax = ax.ravel()
+            plt.subplots_adjust(wspace=0.05, hspace=0.05)
+            for i in range(n):
+                ax[i].imshow(blocks[i].squeeze())  # cmap='gray'
+                ax[i].axis('off')
+
+            LOGGER.info(f'Saving {f}... ({n}/{channels})')
+            plt.savefig(f, dpi=300, bbox_inches='tight')
+            plt.close()
+            np.save(str(f.with_suffix('.npy')), x[0].cpu().numpy())  # npy save
+
+
+def hist2d(x, y, n=100):
+    # 2d histogram used in labels.png and evolve.png
+    xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n)
+    hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges))
+    xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1)
+    yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1)
+    return np.log(hist[xidx, yidx])
+
+
+def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5):
+    from scipy.signal import butter, filtfilt
+
+    # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy
+    def butter_lowpass(cutoff, fs, order):
+        nyq = 0.5 * fs
+        normal_cutoff = cutoff / nyq
+        return butter(order, normal_cutoff, btype='low', analog=False)
+
+    b, a = butter_lowpass(cutoff, fs, order=order)
+    return filtfilt(b, a, data)  # forward-backward filter
+
+
+def output_to_target(output, max_det=300):
+    # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] for plotting
+    targets = []
+    for i, o in enumerate(output):
+        box, conf, cls = o[:max_det, :6].cpu().split((4, 1, 1), 1)
+        j = torch.full((conf.shape[0], 1), i)
+        targets.append(torch.cat((j, cls, xyxy2xywh(box), conf), 1))
+    return torch.cat(targets, 0).numpy()
+
+
+@threaded
+def plot_images(images, targets, paths=None, fname='images.jpg', names=None):
+    # Plot image grid with labels
+    if isinstance(images, torch.Tensor):
+        images = images.cpu().float().numpy()
+    if isinstance(targets, torch.Tensor):
+        targets = targets.cpu().numpy()
+
+    max_size = 1920  # max image size
+    max_subplots = 16  # max image subplots, i.e. 4x4
+    bs, _, h, w = images.shape  # batch size, _, height, width
+    bs = min(bs, max_subplots)  # limit plot images
+    ns = np.ceil(bs ** 0.5)  # number of subplots (square)
+    if np.max(images[0]) <= 1:
+        images *= 255  # de-normalise (optional)
+
+    # Build Image
+    mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8)  # init
+    for i, im in enumerate(images):
+        if i == max_subplots:  # if last batch has fewer images than we expect
+            break
+        x, y = int(w * (i // ns)), int(h * (i % ns))  # block origin
+        im = im.transpose(1, 2, 0)
+        mosaic[y:y + h, x:x + w, :] = im
+
+    # Resize (optional)
+    scale = max_size / ns / max(h, w)
+    if scale < 1:
+        h = math.ceil(scale * h)
+        w = math.ceil(scale * w)
+        mosaic = cv2.resize(mosaic, tuple(int(x * ns) for x in (w, h)))
+
+    # Annotate
+    fs = int((h + w) * ns * 0.01)  # font size
+    annotator = Annotator(mosaic, line_width=round(fs / 10), font_size=fs, pil=True, example=names)
+    for i in range(i + 1):
+        x, y = int(w * (i // ns)), int(h * (i % ns))  # block origin
+        annotator.rectangle([x, y, x + w, y + h], None, (255, 255, 255), width=2)  # borders
+        if paths:
+            annotator.text((x + 5, y + 5), text=Path(paths[i]).name[:40], txt_color=(220, 220, 220))  # filenames
+        if len(targets) > 0:
+            ti = targets[targets[:, 0] == i]  # image targets
+            boxes = xywh2xyxy(ti[:, 2:6]).T
+            classes = ti[:, 1].astype('int')
+            labels = ti.shape[1] == 6  # labels if no conf column
+            conf = None if labels else ti[:, 6]  # check for confidence presence (label vs pred)
+
+            if boxes.shape[1]:
+                if boxes.max() <= 1.01:  # if normalized with tolerance 0.01
+                    boxes[[0, 2]] *= w  # scale to pixels
+                    boxes[[1, 3]] *= h
+                elif scale < 1:  # absolute coords need scale if image scales
+                    boxes *= scale
+            boxes[[0, 2]] += x
+            boxes[[1, 3]] += y
+            for j, box in enumerate(boxes.T.tolist()):
+                cls = classes[j]
+                color = colors(cls)
+                cls = names[cls] if names else cls
+                if labels or conf[j] > 0.25:  # 0.25 conf thresh
+                    label = f'{cls}' if labels else f'{cls} {conf[j]:.1f}'
+                    annotator.box_label(box, label, color=color)
+    annotator.im.save(fname)  # save
+
+
+def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''):
+    # Plot LR simulating training for full epochs
+    optimizer, scheduler = copy(optimizer), copy(scheduler)  # do not modify originals
+    y = []
+    for _ in range(epochs):
+        scheduler.step()
+        y.append(optimizer.param_groups[0]['lr'])
+    plt.plot(y, '.-', label='LR')
+    plt.xlabel('epoch')
+    plt.ylabel('LR')
+    plt.grid()
+    plt.xlim(0, epochs)
+    plt.ylim(0)
+    plt.savefig(Path(save_dir) / 'LR.png', dpi=200)
+    plt.close()
+
+
+def plot_val_txt():  # from utils.plots import *; plot_val()
+    # Plot val.txt histograms
+    x = np.loadtxt('val.txt', dtype=np.float32)
+    box = xyxy2xywh(x[:, :4])
+    cx, cy = box[:, 0], box[:, 1]
+
+    fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True)
+    ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0)
+    ax.set_aspect('equal')
+    plt.savefig('hist2d.png', dpi=300)
+
+    fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True)
+    ax[0].hist(cx, bins=600)
+    ax[1].hist(cy, bins=600)
+    plt.savefig('hist1d.png', dpi=200)
+
+
+def plot_targets_txt():  # from utils.plots import *; plot_targets_txt()
+    # Plot targets.txt histograms
+    x = np.loadtxt('targets.txt', dtype=np.float32).T
+    s = ['x targets', 'y targets', 'width targets', 'height targets']
+    fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)
+    ax = ax.ravel()
+    for i in range(4):
+        ax[i].hist(x[i], bins=100, label=f'{x[i].mean():.3g} +/- {x[i].std():.3g}')
+        ax[i].legend()
+        ax[i].set_title(s[i])
+    plt.savefig('targets.jpg', dpi=200)
+
+
+def plot_val_study(file='', dir='', x=None):  # from utils.plots import *; plot_val_study()
+    # Plot file=study.txt generated by val.py (or plot all study*.txt in dir)
+    save_dir = Path(file).parent if file else Path(dir)
+    plot2 = False  # plot additional results
+    if plot2:
+        ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)[1].ravel()
+
+    fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True)
+    # for f in [save_dir / f'study_coco_{x}.txt' for x in ['yolov5n6', 'yolov5s6', 'yolov5m6', 'yolov5l6', 'yolov5x6']]:
+    for f in sorted(save_dir.glob('study*.txt')):
+        y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T
+        x = np.arange(y.shape[1]) if x is None else np.array(x)
+        if plot2:
+            s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_preprocess (ms/img)', 't_inference (ms/img)', 't_NMS (ms/img)']
+            for i in range(7):
+                ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8)
+                ax[i].set_title(s[i])
+
+        j = y[3].argmax() + 1
+        ax2.plot(y[5, 1:j],
+                 y[3, 1:j] * 1E2,
+                 '.-',
+                 linewidth=2,
+                 markersize=8,
+                 label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO'))
+
+    ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5],
+             'k.-',
+             linewidth=2,
+             markersize=8,
+             alpha=.25,
+             label='EfficientDet')
+
+    ax2.grid(alpha=0.2)
+    ax2.set_yticks(np.arange(20, 60, 5))
+    ax2.set_xlim(0, 57)
+    ax2.set_ylim(25, 55)
+    ax2.set_xlabel('GPU Speed (ms/img)')
+    ax2.set_ylabel('COCO AP val')
+    ax2.legend(loc='lower right')
+    f = save_dir / 'study.png'
+    print(f'Saving {f}...')
+    plt.savefig(f, dpi=300)
+
+
+@TryExcept()  # known issue https://github.com/ultralytics/yolov5/issues/5395
+def plot_labels(labels, names=(), save_dir=Path('')):
+    # plot dataset labels
+    LOGGER.info(f"Plotting labels to {save_dir / 'labels.jpg'}... ")
+    c, b = labels[:, 0], labels[:, 1:].transpose()  # classes, boxes
+    nc = int(c.max() + 1)  # number of classes
+    x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height'])
+
+    # seaborn correlogram
+    sn.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9))
+    plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200)
+    plt.close()
+
+    # matplotlib labels
+    matplotlib.use('svg')  # faster
+    ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel()
+    y = ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8)
+    with contextlib.suppress(Exception):  # color histogram bars by class
+        [y[2].patches[i].set_color([x / 255 for x in colors(i)]) for i in range(nc)]  # known issue #3195
+    ax[0].set_ylabel('instances')
+    if 0 < len(names) < 30:
+        ax[0].set_xticks(range(len(names)))
+        ax[0].set_xticklabels(list(names.values()), rotation=90, fontsize=10)
+    else:
+        ax[0].set_xlabel('classes')
+    sn.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9)
+    sn.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9)
+
+    # rectangles
+    labels[:, 1:3] = 0.5  # center
+    labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000
+    img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255)
+    for cls, *box in labels[:1000]:
+        ImageDraw.Draw(img).rectangle(box, width=1, outline=colors(cls))  # plot
+    ax[1].imshow(img)
+    ax[1].axis('off')
+
+    for a in [0, 1, 2, 3]:
+        for s in ['top', 'right', 'left', 'bottom']:
+            ax[a].spines[s].set_visible(False)
+
+    plt.savefig(save_dir / 'labels.jpg', dpi=200)
+    matplotlib.use('Agg')
+    plt.close()
+
+
+def imshow_cls(im, labels=None, pred=None, names=None, nmax=25, verbose=False, f=Path('images.jpg')):
+    # Show classification image grid with labels (optional) and predictions (optional)
+    from utils.augmentations import denormalize
+
+    names = names or [f'class{i}' for i in range(1000)]
+    blocks = torch.chunk(denormalize(im.clone()).cpu().float(), len(im),
+                         dim=0)  # select batch index 0, block by channels
+    n = min(len(blocks), nmax)  # number of plots
+    m = min(8, round(n ** 0.5))  # 8 x 8 default
+    fig, ax = plt.subplots(math.ceil(n / m), m)  # 8 rows x n/8 cols
+    ax = ax.ravel() if m > 1 else [ax]
+    # plt.subplots_adjust(wspace=0.05, hspace=0.05)
+    for i in range(n):
+        ax[i].imshow(blocks[i].squeeze().permute((1, 2, 0)).numpy().clip(0.0, 1.0))
+        ax[i].axis('off')
+        if labels is not None:
+            s = names[labels[i]] + (f'—{names[pred[i]]}' if pred is not None else '')
+            ax[i].set_title(s, fontsize=8, verticalalignment='top')
+    plt.savefig(f, dpi=300, bbox_inches='tight')
+    plt.close()
+    if verbose:
+        LOGGER.info(f"Saving {f}")
+        if labels is not None:
+            LOGGER.info('True:     ' + ' '.join(f'{names[i]:3s}' for i in labels[:nmax]))
+        if pred is not None:
+            LOGGER.info('Predicted:' + ' '.join(f'{names[i]:3s}' for i in pred[:nmax]))
+    return f
+
+
+def plot_evolve(evolve_csv='path/to/evolve.csv'):  # from utils.plots import *; plot_evolve()
+    # Plot evolve.csv hyp evolution results
+    evolve_csv = Path(evolve_csv)
+    data = pd.read_csv(evolve_csv)
+    keys = [x.strip() for x in data.columns]
+    x = data.values
+    f = fitness(x)
+    j = np.argmax(f)  # max fitness index
+    plt.figure(figsize=(10, 12), tight_layout=True)
+    matplotlib.rc('font', **{'size': 8})
+    print(f'Best results from row {j} of {evolve_csv}:')
+    for i, k in enumerate(keys[7:]):
+        v = x[:, 7 + i]
+        mu = v[j]  # best single result
+        plt.subplot(6, 5, i + 1)
+        plt.scatter(v, f, c=hist2d(v, f, 20), cmap='viridis', alpha=.8, edgecolors='none')
+        plt.plot(mu, f.max(), 'k+', markersize=15)
+        plt.title(f'{k} = {mu:.3g}', fontdict={'size': 9})  # limit to 40 characters
+        if i % 5 != 0:
+            plt.yticks([])
+        print(f'{k:>15}: {mu:.3g}')
+    f = evolve_csv.with_suffix('.png')  # filename
+    plt.savefig(f, dpi=200)
+    plt.close()
+    print(f'Saved {f}')
+
+
+def plot_results(file='path/to/results.csv', dir=''):
+    # Plot training results.csv. Usage: from utils.plots import *; plot_results('path/to/results.csv')
+    save_dir = Path(file).parent if file else Path(dir)
+    fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True)
+    ax = ax.ravel()
+    files = list(save_dir.glob('results*.csv'))
+    assert len(files), f'No results.csv files found in {save_dir.resolve()}, nothing to plot.'
+    for f in files:
+        try:
+            data = pd.read_csv(f)
+            s = [x.strip() for x in data.columns]
+            x = data.values[:, 0]
+            for i, j in enumerate([1, 2, 3, 4, 5, 8, 9, 10, 6, 7]):
+                y = data.values[:, j].astype('float')
+                # y[y == 0] = np.nan  # don't show zero values
+                ax[i].plot(x, y, marker='.', label=f.stem, linewidth=2, markersize=8)
+                ax[i].set_title(s[j], fontsize=12)
+                # if j in [8, 9, 10]:  # share train and val loss y axes
+                #     ax[i].get_shared_y_axes().join(ax[i], ax[i - 5])
+        except Exception as e:
+            LOGGER.info(f'Warning: Plotting error for {f}: {e}')
+    ax[1].legend()
+    fig.savefig(save_dir / 'results.png', dpi=200)
+    plt.close()
+
+
+def profile_idetection(start=0, stop=0, labels=(), save_dir=''):
+    # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection()
+    ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel()
+    s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS']
+    files = list(Path(save_dir).glob('frames*.txt'))
+    for fi, f in enumerate(files):
+        try:
+            results = np.loadtxt(f, ndmin=2).T[:, 90:-30]  # clip first and last rows
+            n = results.shape[1]  # number of rows
+            x = np.arange(start, min(stop, n) if stop else n)
+            results = results[:, x]
+            t = (results[0] - results[0].min())  # set t0=0s
+            results[0] = x
+            for i, a in enumerate(ax):
+                if i < len(results):
+                    label = labels[fi] if len(labels) else f.stem.replace('frames_', '')
+                    a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5)
+                    a.set_title(s[i])
+                    a.set_xlabel('time (s)')
+                    # if fi == len(files) - 1:
+                    #     a.set_ylim(bottom=0)
+                    for side in ['top', 'right']:
+                        a.spines[side].set_visible(False)
+                else:
+                    a.remove()
+        except Exception as e:
+            print(f'Warning: Plotting error for {f}; {e}')
+    ax[1].legend()
+    plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200)
+
+
+def save_one_box(xyxy, im, file=Path('im.jpg'), gain=1.02, pad=10, square=False, BGR=False, save=True):
+    # Save image crop as {file} with crop size multiple {gain} and {pad} pixels. Save and/or return crop
+    xyxy = torch.tensor(xyxy).view(-1, 4)
+    b = xyxy2xywh(xyxy)  # boxes
+    if square:
+        b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1)  # attempt rectangle to square
+    b[:, 2:] = b[:, 2:] * gain + pad  # box wh * gain + pad
+    xyxy = xywh2xyxy(b).long()
+    clip_boxes(xyxy, im.shape)
+    crop = im[int(xyxy[0, 1]):int(xyxy[0, 3]), int(xyxy[0, 0]):int(xyxy[0, 2]), ::(1 if BGR else -1)]
+    if save:
+        file.parent.mkdir(parents=True, exist_ok=True)  # make directory
+        f = str(increment_path(file).with_suffix('.jpg'))
+        # cv2.imwrite(f, crop)  # save BGR, https://github.com/ultralytics/yolov5/issues/7007 chroma subsampling issue
+        Image.fromarray(crop[..., ::-1]).save(f, quality=95, subsampling=0)  # save RGB
+    return crop

utils/segment/augmentations.py

@@ -0,0 +1,104 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Image augmentation functions
+"""
+
+import math
+import random
+
+import cv2
+import numpy as np
+
+from ..augmentations import box_candidates
+from ..general import resample_segments, segment2box
+
+
+def mixup(im, labels, segments, im2, labels2, segments2):
+    # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
+    r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
+    im = (im * r + im2 * (1 - r)).astype(np.uint8)
+    labels = np.concatenate((labels, labels2), 0)
+    segments = np.concatenate((segments, segments2), 0)
+    return im, labels, segments
+
+
+def random_perspective(im,
+                       targets=(),
+                       segments=(),
+                       degrees=10,
+                       translate=.1,
+                       scale=.1,
+                       shear=10,
+                       perspective=0.0,
+                       border=(0, 0)):
+    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
+    # targets = [cls, xyxy]
+
+    height = im.shape[0] + border[0] * 2  # shape(h,w,c)
+    width = im.shape[1] + border[1] * 2
+
+    # Center
+    C = np.eye(3)
+    C[0, 2] = -im.shape[1] / 2  # x translation (pixels)
+    C[1, 2] = -im.shape[0] / 2  # y translation (pixels)
+
+    # Perspective
+    P = np.eye(3)
+    P[2, 0] = random.uniform(-perspective, perspective)  # x perspective (about y)
+    P[2, 1] = random.uniform(-perspective, perspective)  # y perspective (about x)
+
+    # Rotation and Scale
+    R = np.eye(3)
+    a = random.uniform(-degrees, degrees)
+    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
+    s = random.uniform(1 - scale, 1 + scale)
+    # s = 2 ** random.uniform(-scale, scale)
+    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)
+
+    # Shear
+    S = np.eye(3)
+    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
+    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)
+
+    # Translation
+    T = np.eye(3)
+    T[0, 2] = (random.uniform(0.5 - translate, 0.5 + translate) * width)  # x translation (pixels)
+    T[1, 2] = (random.uniform(0.5 - translate, 0.5 + translate) * height)  # y translation (pixels)
+
+    # Combined rotation matrix
+    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
+    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
+        if perspective:
+            im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))
+        else:  # affine
+            im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))
+
+    # Visualize
+    # import matplotlib.pyplot as plt
+    # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
+    # ax[0].imshow(im[:, :, ::-1])  # base
+    # ax[1].imshow(im2[:, :, ::-1])  # warped
+
+    # Transform label coordinates
+    n = len(targets)
+    new_segments = []
+    if n:
+        new = np.zeros((n, 4))
+        segments = resample_segments(segments)  # upsample
+        for i, segment in enumerate(segments):
+            xy = np.ones((len(segment), 3))
+            xy[:, :2] = segment
+            xy = xy @ M.T  # transform
+            xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2])  # perspective rescale or affine
+
+            # clip
+            new[i] = segment2box(xy, width, height)
+            new_segments.append(xy)
+
+        # filter candidates
+        i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01)
+        targets = targets[i]
+        targets[:, 1:5] = new[i]
+        new_segments = np.array(new_segments)[i]
+
+    return im, targets, new_segments

utils/segment/dataloaders.py

@@ -0,0 +1,331 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Dataloaders
+"""
+
+import os
+import random
+
+import cv2
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, distributed
+
+from ..augmentations import augment_hsv, copy_paste, letterbox
+from ..dataloaders import InfiniteDataLoader, LoadImagesAndLabels, seed_worker
+from ..general import LOGGER, xyn2xy, xywhn2xyxy, xyxy2xywhn
+from ..torch_utils import torch_distributed_zero_first
+from .augmentations import mixup, random_perspective
+
+RANK = int(os.getenv('RANK', -1))
+
+
+def create_dataloader(path,
+                      imgsz,
+                      batch_size,
+                      stride,
+                      single_cls=False,
+                      hyp=None,
+                      augment=False,
+                      cache=False,
+                      pad=0.0,
+                      rect=False,
+                      rank=-1,
+                      workers=8,
+                      image_weights=False,
+                      quad=False,
+                      prefix='',
+                      shuffle=False,
+                      mask_downsample_ratio=1,
+                      overlap_mask=False):
+    if rect and shuffle:
+        LOGGER.warning('WARNING ⚠️ --rect is incompatible with DataLoader shuffle, setting shuffle=False')
+        shuffle = False
+    with torch_distributed_zero_first(rank):  # init dataset *.cache only once if DDP
+        dataset = LoadImagesAndLabelsAndMasks(
+            path,
+            imgsz,
+            batch_size,
+            augment=augment,  # augmentation
+            hyp=hyp,  # hyperparameters
+            rect=rect,  # rectangular batches
+            cache_images=cache,
+            single_cls=single_cls,
+            stride=int(stride),
+            pad=pad,
+            image_weights=image_weights,
+            prefix=prefix,
+            downsample_ratio=mask_downsample_ratio,
+            overlap=overlap_mask)
+
+    batch_size = min(batch_size, len(dataset))
+    nd = torch.cuda.device_count()  # number of CUDA devices
+    nw = min([os.cpu_count() // max(nd, 1), batch_size if batch_size > 1 else 0, workers])  # number of workers
+    sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle)
+    loader = DataLoader if image_weights else InfiniteDataLoader  # only DataLoader allows for attribute updates
+    generator = torch.Generator()
+    generator.manual_seed(6148914691236517205 + RANK)
+    return loader(
+        dataset,
+        batch_size=batch_size,
+        shuffle=shuffle and sampler is None,
+        num_workers=nw,
+        sampler=sampler,
+        pin_memory=True,
+        collate_fn=LoadImagesAndLabelsAndMasks.collate_fn4 if quad else LoadImagesAndLabelsAndMasks.collate_fn,
+        worker_init_fn=seed_worker,
+        generator=generator,
+    ), dataset
+
+
+class LoadImagesAndLabelsAndMasks(LoadImagesAndLabels):  # for training/testing
+
+    def __init__(
+        self,
+        path,
+        img_size=640,
+        batch_size=16,
+        augment=False,
+        hyp=None,
+        rect=False,
+        image_weights=False,
+        cache_images=False,
+        single_cls=False,
+        stride=32,
+        pad=0,
+        min_items=0,
+        prefix="",
+        downsample_ratio=1,
+        overlap=False,
+    ):
+        super().__init__(path, img_size, batch_size, augment, hyp, rect, image_weights, cache_images, single_cls,
+                         stride, pad, min_items, prefix)
+        self.downsample_ratio = downsample_ratio
+        self.overlap = overlap
+
+    def __getitem__(self, index):
+        index = self.indices[index]  # linear, shuffled, or image_weights
+
+        hyp = self.hyp
+        mosaic = self.mosaic and random.random() < hyp['mosaic']
+        masks = []
+        if mosaic:
+            # Load mosaic
+            img, labels, segments = self.load_mosaic(index)
+            shapes = None
+
+            # MixUp augmentation
+            if random.random() < hyp["mixup"]:
+                img, labels, segments = mixup(img, labels, segments, *self.load_mosaic(random.randint(0, self.n - 1)))
+
+        else:
+            # Load image
+            img, (h0, w0), (h, w) = self.load_image(index)
+
+            # Letterbox
+            shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size  # final letterboxed shape
+            img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
+            shapes = (h0, w0), ((h / h0, w / w0), pad)  # for COCO mAP rescaling
+
+            labels = self.labels[index].copy()
+            # [array, array, ....], array.shape=(num_points, 2), xyxyxyxy
+            segments = self.segments[index].copy()
+            if len(segments):
+                for i_s in range(len(segments)):
+                    segments[i_s] = xyn2xy(
+                        segments[i_s],
+                        ratio[0] * w,
+                        ratio[1] * h,
+                        padw=pad[0],
+                        padh=pad[1],
+                    )
+            if labels.size:  # normalized xywh to pixel xyxy format
+                labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])
+
+            if self.augment:
+                img, labels, segments = random_perspective(img,
+                                                           labels,
+                                                           segments=segments,
+                                                           degrees=hyp["degrees"],
+                                                           translate=hyp["translate"],
+                                                           scale=hyp["scale"],
+                                                           shear=hyp["shear"],
+                                                           perspective=hyp["perspective"])
+
+        nl = len(labels)  # number of labels
+        if nl:
+            labels[:, 1:5] = xyxy2xywhn(labels[:, 1:5], w=img.shape[1], h=img.shape[0], clip=True, eps=1e-3)
+            if self.overlap:
+                masks, sorted_idx = polygons2masks_overlap(img.shape[:2],
+                                                           segments,
+                                                           downsample_ratio=self.downsample_ratio)
+                masks = masks[None]  # (640, 640) -> (1, 640, 640)
+                labels = labels[sorted_idx]
+            else:
+                masks = polygons2masks(img.shape[:2], segments, color=1, downsample_ratio=self.downsample_ratio)
+
+        masks = (torch.from_numpy(masks) if len(masks) else torch.zeros(1 if self.overlap else nl, img.shape[0] //
+                                                                        self.downsample_ratio, img.shape[1] //
+                                                                        self.downsample_ratio))
+        # TODO: albumentations support
+        if self.augment:
+            # Albumentations
+            # there are some augmentation that won't change boxes and masks,
+            # so just be it for now.
+            img, labels = self.albumentations(img, labels)
+            nl = len(labels)  # update after albumentations
+
+            # HSV color-space
+            augment_hsv(img, hgain=hyp["hsv_h"], sgain=hyp["hsv_s"], vgain=hyp["hsv_v"])
+
+            # Flip up-down
+            if random.random() < hyp["flipud"]:
+                img = np.flipud(img)
+                if nl:
+                    labels[:, 2] = 1 - labels[:, 2]
+                    masks = torch.flip(masks, dims=[1])
+
+            # Flip left-right
+            if random.random() < hyp["fliplr"]:
+                img = np.fliplr(img)
+                if nl:
+                    labels[:, 1] = 1 - labels[:, 1]
+                    masks = torch.flip(masks, dims=[2])
+
+            # Cutouts  # labels = cutout(img, labels, p=0.5)
+
+        labels_out = torch.zeros((nl, 6))
+        if nl:
+            labels_out[:, 1:] = torch.from_numpy(labels)
+
+        # Convert
+        img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
+        img = np.ascontiguousarray(img)
+
+        return (torch.from_numpy(img), labels_out, self.im_files[index], shapes, masks)
+
+    def load_mosaic(self, index):
+        # YOLOv5 4-mosaic loader. Loads 1 image + 3 random images into a 4-image mosaic
+        labels4, segments4 = [], []
+        s = self.img_size
+        yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border)  # mosaic center x, y
+
+        # 3 additional image indices
+        indices = [index] + random.choices(self.indices, k=3)  # 3 additional image indices
+        for i, index in enumerate(indices):
+            # Load image
+            img, _, (h, w) = self.load_image(index)
+
+            # place img in img4
+            if i == 0:  # top left
+                img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
+                x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
+                x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
+            elif i == 1:  # top right
+                x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
+                x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
+            elif i == 2:  # bottom left
+                x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
+                x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
+            elif i == 3:  # bottom right
+                x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
+                x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)
+
+            img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
+            padw = x1a - x1b
+            padh = y1a - y1b
+
+            labels, segments = self.labels[index].copy(), self.segments[index].copy()
+
+            if labels.size:
+                labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh)  # normalized xywh to pixel xyxy format
+                segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
+            labels4.append(labels)
+            segments4.extend(segments)
+
+        # Concat/clip labels
+        labels4 = np.concatenate(labels4, 0)
+        for x in (labels4[:, 1:], *segments4):
+            np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
+        # img4, labels4 = replicate(img4, labels4)  # replicate
+
+        # Augment
+        img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp["copy_paste"])
+        img4, labels4, segments4 = random_perspective(img4,
+                                                      labels4,
+                                                      segments4,
+                                                      degrees=self.hyp["degrees"],
+                                                      translate=self.hyp["translate"],
+                                                      scale=self.hyp["scale"],
+                                                      shear=self.hyp["shear"],
+                                                      perspective=self.hyp["perspective"],
+                                                      border=self.mosaic_border)  # border to remove
+        return img4, labels4, segments4
+
+    @staticmethod
+    def collate_fn(batch):
+        img, label, path, shapes, masks = zip(*batch)  # transposed
+        batched_masks = torch.cat(masks, 0)
+        for i, l in enumerate(label):
+            l[:, 0] = i  # add target image index for build_targets()
+        return torch.stack(img, 0), torch.cat(label, 0), path, shapes, batched_masks
+
+
+def polygon2mask(img_size, polygons, color=1, downsample_ratio=1):
+    """
+    Args:
+        img_size (tuple): The image size.
+        polygons (np.ndarray): [N, M], N is the number of polygons,
+            M is the number of points(Be divided by 2).
+    """
+    mask = np.zeros(img_size, dtype=np.uint8)
+    polygons = np.asarray(polygons)
+    polygons = polygons.astype(np.int32)
+    shape = polygons.shape
+    polygons = polygons.reshape(shape[0], -1, 2)
+    cv2.fillPoly(mask, polygons, color=color)
+    nh, nw = (img_size[0] // downsample_ratio, img_size[1] // downsample_ratio)
+    # NOTE: fillPoly firstly then resize is trying the keep the same way
+    # of loss calculation when mask-ratio=1.
+    mask = cv2.resize(mask, (nw, nh))
+    return mask
+
+
+def polygons2masks(img_size, polygons, color, downsample_ratio=1):
+    """
+    Args:
+        img_size (tuple): The image size.
+        polygons (list[np.ndarray]): each polygon is [N, M],
+            N is the number of polygons,
+            M is the number of points(Be divided by 2).
+    """
+    masks = []
+    for si in range(len(polygons)):
+        mask = polygon2mask(img_size, [polygons[si].reshape(-1)], color, downsample_ratio)
+        masks.append(mask)
+    return np.array(masks)
+
+
+def polygons2masks_overlap(img_size, segments, downsample_ratio=1):
+    """Return a (640, 640) overlap mask."""
+    masks = np.zeros((img_size[0] // downsample_ratio, img_size[1] // downsample_ratio),
+                     dtype=np.int32 if len(segments) > 255 else np.uint8)
+    areas = []
+    ms = []
+    for si in range(len(segments)):
+        mask = polygon2mask(
+            img_size,
+            [segments[si].reshape(-1)],
+            downsample_ratio=downsample_ratio,
+            color=1,
+        )
+        ms.append(mask)
+        areas.append(mask.sum())
+    areas = np.asarray(areas)
+    index = np.argsort(-areas)
+    ms = np.array(ms)[index]
+    for i in range(len(segments)):
+        mask = ms[i] * (i + 1)
+        masks = masks + mask
+        masks = np.clip(masks, a_min=0, a_max=i + 1)
+    return masks, index

utils/segment/general.py

@@ -0,0 +1,160 @@
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+def crop_mask(masks, boxes):
+    """
+    "Crop" predicted masks by zeroing out everything not in the predicted bbox.
+    Vectorized by Chong (thanks Chong).
+
+    Args:
+        - masks should be a size [h, w, n] tensor of masks
+        - boxes should be a size [n, 4] tensor of bbox coords in relative point form
+    """
+
+    n, h, w = masks.shape
+    x1, y1, x2, y2 = torch.chunk(boxes[:, :, None], 4, 1)  # x1 shape(1,1,n)
+    r = torch.arange(w, device=masks.device, dtype=x1.dtype)[None, None, :]  # rows shape(1,w,1)
+    c = torch.arange(h, device=masks.device, dtype=x1.dtype)[None, :, None]  # cols shape(h,1,1)
+
+    return masks * ((r >= x1) * (r < x2) * (c >= y1) * (c < y2))
+
+
+def process_mask_upsample(protos, masks_in, bboxes, shape):
+    """
+    Crop after upsample.
+    proto_out: [mask_dim, mask_h, mask_w]
+    out_masks: [n, mask_dim], n is number of masks after nms
+    bboxes: [n, 4], n is number of masks after nms
+    shape:input_image_size, (h, w)
+
+    return: h, w, n
+    """
+
+    c, mh, mw = protos.shape  # CHW
+    masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)
+    masks = F.interpolate(masks[None], shape, mode='bilinear', align_corners=False)[0]  # CHW
+    masks = crop_mask(masks, bboxes)  # CHW
+    return masks.gt_(0.5)
+
+
+def process_mask(protos, masks_in, bboxes, shape, upsample=False):
+    """
+    Crop before upsample.
+    proto_out: [mask_dim, mask_h, mask_w]
+    out_masks: [n, mask_dim], n is number of masks after nms
+    bboxes: [n, 4], n is number of masks after nms
+    shape:input_image_size, (h, w)
+
+    return: h, w, n
+    """
+
+    c, mh, mw = protos.shape  # CHW
+    ih, iw = shape
+    masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)  # CHW
+
+    downsampled_bboxes = bboxes.clone()
+    downsampled_bboxes[:, 0] *= mw / iw
+    downsampled_bboxes[:, 2] *= mw / iw
+    downsampled_bboxes[:, 3] *= mh / ih
+    downsampled_bboxes[:, 1] *= mh / ih
+
+    masks = crop_mask(masks, downsampled_bboxes)  # CHW
+    if upsample:
+        masks = F.interpolate(masks[None], shape, mode='bilinear', align_corners=False)[0]  # CHW
+    return masks.gt_(0.5)
+
+
+def process_mask_native(protos, masks_in, bboxes, dst_shape):
+    """
+    Crop after upsample.
+    proto_out: [mask_dim, mask_h, mask_w]
+    out_masks: [n, mask_dim], n is number of masks after nms
+    bboxes: [n, 4], n is number of masks after nms
+    shape:input_image_size, (h, w)
+
+    return: h, w, n
+    """
+    c, mh, mw = protos.shape  # CHW
+    masks = (masks_in @ protos.float().view(c, -1)).sigmoid().view(-1, mh, mw)
+    gain = min(mh / dst_shape[0], mw / dst_shape[1])  # gain  = old / new
+    pad = (mw - dst_shape[1] * gain) / 2, (mh - dst_shape[0] * gain) / 2  # wh padding
+    top, left = int(pad[1]), int(pad[0])  # y, x
+    bottom, right = int(mh - pad[1]), int(mw - pad[0])
+    masks = masks[:, top:bottom, left:right]
+
+    masks = F.interpolate(masks[None], dst_shape, mode='bilinear', align_corners=False)[0]  # CHW
+    masks = crop_mask(masks, bboxes)  # CHW
+    return masks.gt_(0.5)
+
+
+def scale_image(im1_shape, masks, im0_shape, ratio_pad=None):
+    """
+    img1_shape: model input shape, [h, w]
+    img0_shape: origin pic shape, [h, w, 3]
+    masks: [h, w, num]
+    """
+    # Rescale coordinates (xyxy) from im1_shape to im0_shape
+    if ratio_pad is None:  # calculate from im0_shape
+        gain = min(im1_shape[0] / im0_shape[0], im1_shape[1] / im0_shape[1])  # gain  = old / new
+        pad = (im1_shape[1] - im0_shape[1] * gain) / 2, (im1_shape[0] - im0_shape[0] * gain) / 2  # wh padding
+    else:
+        pad = ratio_pad[1]
+    top, left = int(pad[1]), int(pad[0])  # y, x
+    bottom, right = int(im1_shape[0] - pad[1]), int(im1_shape[1] - pad[0])
+
+    if len(masks.shape) < 2:
+        raise ValueError(f'"len of masks shape" should be 2 or 3, but got {len(masks.shape)}')
+    masks = masks[top:bottom, left:right]
+    # masks = masks.permute(2, 0, 1).contiguous()
+    # masks = F.interpolate(masks[None], im0_shape[:2], mode='bilinear', align_corners=False)[0]
+    # masks = masks.permute(1, 2, 0).contiguous()
+    masks = cv2.resize(masks, (im0_shape[1], im0_shape[0]))
+
+    if len(masks.shape) == 2:
+        masks = masks[:, :, None]
+    return masks
+
+
+def mask_iou(mask1, mask2, eps=1e-7):
+    """
+    mask1: [N, n] m1 means number of predicted objects
+    mask2: [M, n] m2 means number of gt objects
+    Note: n means image_w x image_h
+
+    return: masks iou, [N, M]
+    """
+    intersection = torch.matmul(mask1, mask2.t()).clamp(0)
+    union = (mask1.sum(1)[:, None] + mask2.sum(1)[None]) - intersection  # (area1 + area2) - intersection
+    return intersection / (union + eps)
+
+
+def masks_iou(mask1, mask2, eps=1e-7):
+    """
+    mask1: [N, n] m1 means number of predicted objects
+    mask2: [N, n] m2 means number of gt objects
+    Note: n means image_w x image_h
+
+    return: masks iou, (N, )
+    """
+    intersection = (mask1 * mask2).sum(1).clamp(0)  # (N, )
+    union = (mask1.sum(1) + mask2.sum(1))[None] - intersection  # (area1 + area2) - intersection
+    return intersection / (union + eps)
+
+
+def masks2segments(masks, strategy='largest'):
+    # Convert masks(n,160,160) into segments(n,xy)
+    segments = []
+    for x in masks.int().cpu().numpy().astype('uint8'):
+        c = cv2.findContours(x, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[0]
+        if c:
+            if strategy == 'concat':  # concatenate all segments
+                c = np.concatenate([x.reshape(-1, 2) for x in c])
+            elif strategy == 'largest':  # select largest segment
+                c = np.array(c[np.array([len(x) for x in c]).argmax()]).reshape(-1, 2)
+        else:
+            c = np.zeros((0, 2))  # no segments found
+        segments.append(c.astype('float32'))
+    return segments

utils/segment/loss.py

@@ -0,0 +1,186 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..general import xywh2xyxy
+from ..loss import FocalLoss, smooth_BCE
+from ..metrics import bbox_iou
+from ..torch_utils import de_parallel
+from .general import crop_mask
+
+
+class ComputeLoss:
+    # Compute losses
+    def __init__(self, model, autobalance=False, overlap=False):
+        self.sort_obj_iou = False
+        self.overlap = overlap
+        device = next(model.parameters()).device  # get model device
+        h = model.hyp  # hyperparameters
+        self.device = device
+
+        # Define criteria
+        BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device))
+        BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device))
+
+        # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3
+        self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0))  # positive, negative BCE targets
+
+        # Focal loss
+        g = h['fl_gamma']  # focal loss gamma
+        if g > 0:
+            BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g)
+
+        m = de_parallel(model).model[-1]  # Detect() module
+        self.balance = {3: [4.0, 1.0, 0.4]}.get(m.nl, [4.0, 1.0, 0.25, 0.06, 0.02])  # P3-P7
+        self.ssi = list(m.stride).index(16) if autobalance else 0  # stride 16 index
+        self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance
+        self.na = m.na  # number of anchors
+        self.nc = m.nc  # number of classes
+        self.nl = m.nl  # number of layers
+        self.nm = m.nm  # number of masks
+        self.anchors = m.anchors
+        self.device = device
+
+    def __call__(self, preds, targets, masks):  # predictions, targets, model
+        p, proto = preds
+        bs, nm, mask_h, mask_w = proto.shape  # batch size, number of masks, mask height, mask width
+        lcls = torch.zeros(1, device=self.device)
+        lbox = torch.zeros(1, device=self.device)
+        lobj = torch.zeros(1, device=self.device)
+        lseg = torch.zeros(1, device=self.device)
+        tcls, tbox, indices, anchors, tidxs, xywhn = self.build_targets(p, targets)  # targets
+
+        # Losses
+        for i, pi in enumerate(p):  # layer index, layer predictions
+            b, a, gj, gi = indices[i]  # image, anchor, gridy, gridx
+            tobj = torch.zeros(pi.shape[:4], dtype=pi.dtype, device=self.device)  # target obj
+
+            n = b.shape[0]  # number of targets
+            if n:
+                pxy, pwh, _, pcls, pmask = pi[b, a, gj, gi].split((2, 2, 1, self.nc, nm), 1)  # subset of predictions
+
+                # Box regression
+                pxy = pxy.sigmoid() * 2 - 0.5
+                pwh = (pwh.sigmoid() * 2) ** 2 * anchors[i]
+                pbox = torch.cat((pxy, pwh), 1)  # predicted box
+                iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze()  # iou(prediction, target)
+                lbox += (1.0 - iou).mean()  # iou loss
+
+                # Objectness
+                iou = iou.detach().clamp(0).type(tobj.dtype)
+                if self.sort_obj_iou:
+                    j = iou.argsort()
+                    b, a, gj, gi, iou = b[j], a[j], gj[j], gi[j], iou[j]
+                if self.gr < 1:
+                    iou = (1.0 - self.gr) + self.gr * iou
+                tobj[b, a, gj, gi] = iou  # iou ratio
+
+                # Classification
+                if self.nc > 1:  # cls loss (only if multiple classes)
+                    t = torch.full_like(pcls, self.cn, device=self.device)  # targets
+                    t[range(n), tcls[i]] = self.cp
+                    lcls += self.BCEcls(pcls, t)  # BCE
+
+                # Mask regression
+                if tuple(masks.shape[-2:]) != (mask_h, mask_w):  # downsample
+                    masks = F.interpolate(masks[None], (mask_h, mask_w), mode="nearest")[0]
+                marea = xywhn[i][:, 2:].prod(1)  # mask width, height normalized
+                mxyxy = xywh2xyxy(xywhn[i] * torch.tensor([mask_w, mask_h, mask_w, mask_h], device=self.device))
+                for bi in b.unique():
+                    j = b == bi  # matching index
+                    if self.overlap:
+                        mask_gti = torch.where(masks[bi][None] == tidxs[i][j].view(-1, 1, 1), 1.0, 0.0)
+                    else:
+                        mask_gti = masks[tidxs[i]][j]
+                    lseg += self.single_mask_loss(mask_gti, pmask[j], proto[bi], mxyxy[j], marea[j])
+
+            obji = self.BCEobj(pi[..., 4], tobj)
+            lobj += obji * self.balance[i]  # obj loss
+            if self.autobalance:
+                self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item()
+
+        if self.autobalance:
+            self.balance = [x / self.balance[self.ssi] for x in self.balance]
+        lbox *= self.hyp["box"]
+        lobj *= self.hyp["obj"]
+        lcls *= self.hyp["cls"]
+        lseg *= self.hyp["box"] / bs
+
+        loss = lbox + lobj + lcls + lseg
+        return loss * bs, torch.cat((lbox, lseg, lobj, lcls)).detach()
+
+    def single_mask_loss(self, gt_mask, pred, proto, xyxy, area):
+        # Mask loss for one image
+        pred_mask = (pred @ proto.view(self.nm, -1)).view(-1, *proto.shape[1:])  # (n,32) @ (32,80,80) -> (n,80,80)
+        loss = F.binary_cross_entropy_with_logits(pred_mask, gt_mask, reduction="none")
+        return (crop_mask(loss, xyxy).mean(dim=(1, 2)) / area).mean()
+
+    def build_targets(self, p, targets):
+        # Build targets for compute_loss(), input targets(image,class,x,y,w,h)
+        na, nt = self.na, targets.shape[0]  # number of anchors, targets
+        tcls, tbox, indices, anch, tidxs, xywhn = [], [], [], [], [], []
+        gain = torch.ones(8, device=self.device)  # normalized to gridspace gain
+        ai = torch.arange(na, device=self.device).float().view(na, 1).repeat(1, nt)  # same as .repeat_interleave(nt)
+        if self.overlap:
+            batch = p[0].shape[0]
+            ti = []
+            for i in range(batch):
+                num = (targets[:, 0] == i).sum()  # find number of targets of each image
+                ti.append(torch.arange(num, device=self.device).float().view(1, num).repeat(na, 1) + 1)  # (na, num)
+            ti = torch.cat(ti, 1)  # (na, nt)
+        else:
+            ti = torch.arange(nt, device=self.device).float().view(1, nt).repeat(na, 1)
+        targets = torch.cat((targets.repeat(na, 1, 1), ai[..., None], ti[..., None]), 2)  # append anchor indices
+
+        g = 0.5  # bias
+        off = torch.tensor(
+            [
+                [0, 0],
+                [1, 0],
+                [0, 1],
+                [-1, 0],
+                [0, -1],  # j,k,l,m
+                # [1, 1], [1, -1], [-1, 1], [-1, -1],  # jk,jm,lk,lm
+            ],
+            device=self.device).float() * g  # offsets
+
+        for i in range(self.nl):
+            anchors, shape = self.anchors[i], p[i].shape
+            gain[2:6] = torch.tensor(shape)[[3, 2, 3, 2]]  # xyxy gain
+
+            # Match targets to anchors
+            t = targets * gain  # shape(3,n,7)
+            if nt:
+                # Matches
+                r = t[..., 4:6] / anchors[:, None]  # wh ratio
+                j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t']  # compare
+                # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t']  # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
+                t = t[j]  # filter
+
+                # Offsets
+                gxy = t[:, 2:4]  # grid xy
+                gxi = gain[[2, 3]] - gxy  # inverse
+                j, k = ((gxy % 1 < g) & (gxy > 1)).T
+                l, m = ((gxi % 1 < g) & (gxi > 1)).T
+                j = torch.stack((torch.ones_like(j), j, k, l, m))
+                t = t.repeat((5, 1, 1))[j]
+                offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j]
+            else:
+                t = targets[0]
+                offsets = 0
+
+            # Define
+            bc, gxy, gwh, at = t.chunk(4, 1)  # (image, class), grid xy, grid wh, anchors
+            (a, tidx), (b, c) = at.long().T, bc.long().T  # anchors, image, class
+            gij = (gxy - offsets).long()
+            gi, gj = gij.T  # grid indices
+
+            # Append
+            indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1)))  # image, anchor, grid
+            tbox.append(torch.cat((gxy - gij, gwh), 1))  # box
+            anch.append(anchors[a])  # anchors
+            tcls.append(c)  # class
+            tidxs.append(tidx)
+            xywhn.append(torch.cat((gxy, gwh), 1) / gain[2:6])  # xywh normalized
+
+        return tcls, tbox, indices, anch, tidxs, xywhn

utils/segment/metrics.py

@@ -0,0 +1,210 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Model validation metrics
+"""
+
+import numpy as np
+
+from ..metrics import ap_per_class
+
+
+def fitness(x):
+    # Model fitness as a weighted combination of metrics
+    w = [0.0, 0.0, 0.1, 0.9, 0.0, 0.0, 0.1, 0.9]
+    return (x[:, :8] * w).sum(1)
+
+
+def ap_per_class_box_and_mask(
+        tp_m,
+        tp_b,
+        conf,
+        pred_cls,
+        target_cls,
+        plot=False,
+        save_dir=".",
+        names=(),
+):
+    """
+    Args:
+        tp_b: tp of boxes.
+        tp_m: tp of masks.
+        other arguments see `func: ap_per_class`.
+    """
+    results_boxes = ap_per_class(tp_b,
+                                 conf,
+                                 pred_cls,
+                                 target_cls,
+                                 plot=plot,
+                                 save_dir=save_dir,
+                                 names=names,
+                                 prefix="Box")[2:]
+    results_masks = ap_per_class(tp_m,
+                                 conf,
+                                 pred_cls,
+                                 target_cls,
+                                 plot=plot,
+                                 save_dir=save_dir,
+                                 names=names,
+                                 prefix="Mask")[2:]
+
+    results = {
+        "boxes": {
+            "p": results_boxes[0],
+            "r": results_boxes[1],
+            "ap": results_boxes[3],
+            "f1": results_boxes[2],
+            "ap_class": results_boxes[4]},
+        "masks": {
+            "p": results_masks[0],
+            "r": results_masks[1],
+            "ap": results_masks[3],
+            "f1": results_masks[2],
+            "ap_class": results_masks[4]}}
+    return results
+
+
+class Metric:
+
+    def __init__(self) -> None:
+        self.p = []  # (nc, )
+        self.r = []  # (nc, )
+        self.f1 = []  # (nc, )
+        self.all_ap = []  # (nc, 10)
+        self.ap_class_index = []  # (nc, )
+
+    @property
+    def ap50(self):
+        """AP@0.5 of all classes.
+        Return:
+            (nc, ) or [].
+        """
+        return self.all_ap[:, 0] if len(self.all_ap) else []
+
+    @property
+    def ap(self):
+        """AP@0.5:0.95
+        Return:
+            (nc, ) or [].
+        """
+        return self.all_ap.mean(1) if len(self.all_ap) else []
+
+    @property
+    def mp(self):
+        """mean precision of all classes.
+        Return:
+            float.
+        """
+        return self.p.mean() if len(self.p) else 0.0
+
+    @property
+    def mr(self):
+        """mean recall of all classes.
+        Return:
+            float.
+        """
+        return self.r.mean() if len(self.r) else 0.0
+
+    @property
+    def map50(self):
+        """Mean AP@0.5 of all classes.
+        Return:
+            float.
+        """
+        return self.all_ap[:, 0].mean() if len(self.all_ap) else 0.0
+
+    @property
+    def map(self):
+        """Mean AP@0.5:0.95 of all classes.
+        Return:
+            float.
+        """
+        return self.all_ap.mean() if len(self.all_ap) else 0.0
+
+    def mean_results(self):
+        """Mean of results, return mp, mr, map50, map"""
+        return (self.mp, self.mr, self.map50, self.map)
+
+    def class_result(self, i):
+        """class-aware result, return p[i], r[i], ap50[i], ap[i]"""
+        return (self.p[i], self.r[i], self.ap50[i], self.ap[i])
+
+    def get_maps(self, nc):
+        maps = np.zeros(nc) + self.map
+        for i, c in enumerate(self.ap_class_index):
+            maps[c] = self.ap[i]
+        return maps
+
+    def update(self, results):
+        """
+        Args:
+            results: tuple(p, r, ap, f1, ap_class)
+        """
+        p, r, all_ap, f1, ap_class_index = results
+        self.p = p
+        self.r = r
+        self.all_ap = all_ap
+        self.f1 = f1
+        self.ap_class_index = ap_class_index
+
+
+class Metrics:
+    """Metric for boxes and masks."""
+
+    def __init__(self) -> None:
+        self.metric_box = Metric()
+        self.metric_mask = Metric()
+
+    def update(self, results):
+        """
+        Args:
+            results: Dict{'boxes': Dict{}, 'masks': Dict{}}
+        """
+        self.metric_box.update(list(results["boxes"].values()))
+        self.metric_mask.update(list(results["masks"].values()))
+
+    def mean_results(self):
+        return self.metric_box.mean_results() + self.metric_mask.mean_results()
+
+    def class_result(self, i):
+        return self.metric_box.class_result(i) + self.metric_mask.class_result(i)
+
+    def get_maps(self, nc):
+        return self.metric_box.get_maps(nc) + self.metric_mask.get_maps(nc)
+
+    @property
+    def ap_class_index(self):
+        # boxes and masks have the same ap_class_index
+        return self.metric_box.ap_class_index
+
+
+KEYS = [
+    "train/box_loss",
+    "train/seg_loss",  # train loss
+    "train/obj_loss",
+    "train/cls_loss",
+    "metrics/precision(B)",
+    "metrics/recall(B)",
+    "metrics/mAP_0.5(B)",
+    "metrics/mAP_0.5:0.95(B)",  # metrics
+    "metrics/precision(M)",
+    "metrics/recall(M)",
+    "metrics/mAP_0.5(M)",
+    "metrics/mAP_0.5:0.95(M)",  # metrics
+    "val/box_loss",
+    "val/seg_loss",  # val loss
+    "val/obj_loss",
+    "val/cls_loss",
+    "x/lr0",
+    "x/lr1",
+    "x/lr2",]
+
+BEST_KEYS = [
+    "best/epoch",
+    "best/precision(B)",
+    "best/recall(B)",
+    "best/mAP_0.5(B)",
+    "best/mAP_0.5:0.95(B)",
+    "best/precision(M)",
+    "best/recall(M)",
+    "best/mAP_0.5(M)",
+    "best/mAP_0.5:0.95(M)",]

utils/segment/plots.py

@@ -0,0 +1,143 @@
+import contextlib
+import math
+from pathlib import Path
+
+import cv2
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import torch
+
+from .. import threaded
+from ..general import xywh2xyxy
+from ..plots import Annotator, colors
+
+
+@threaded
+def plot_images_and_masks(images, targets, masks, paths=None, fname='images.jpg', names=None):
+    # Plot image grid with labels
+    if isinstance(images, torch.Tensor):
+        images = images.cpu().float().numpy()
+    if isinstance(targets, torch.Tensor):
+        targets = targets.cpu().numpy()
+    if isinstance(masks, torch.Tensor):
+        masks = masks.cpu().numpy().astype(int)
+
+    max_size = 1920  # max image size
+    max_subplots = 16  # max image subplots, i.e. 4x4
+    bs, _, h, w = images.shape  # batch size, _, height, width
+    bs = min(bs, max_subplots)  # limit plot images
+    ns = np.ceil(bs ** 0.5)  # number of subplots (square)
+    if np.max(images[0]) <= 1:
+        images *= 255  # de-normalise (optional)
+
+    # Build Image
+    mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8)  # init
+    for i, im in enumerate(images):
+        if i == max_subplots:  # if last batch has fewer images than we expect
+            break
+        x, y = int(w * (i // ns)), int(h * (i % ns))  # block origin
+        im = im.transpose(1, 2, 0)
+        mosaic[y:y + h, x:x + w, :] = im
+
+    # Resize (optional)
+    scale = max_size / ns / max(h, w)
+    if scale < 1:
+        h = math.ceil(scale * h)
+        w = math.ceil(scale * w)
+        mosaic = cv2.resize(mosaic, tuple(int(x * ns) for x in (w, h)))
+
+    # Annotate
+    fs = int((h + w) * ns * 0.01)  # font size
+    annotator = Annotator(mosaic, line_width=round(fs / 10), font_size=fs, pil=True, example=names)
+    for i in range(i + 1):
+        x, y = int(w * (i // ns)), int(h * (i % ns))  # block origin
+        annotator.rectangle([x, y, x + w, y + h], None, (255, 255, 255), width=2)  # borders
+        if paths:
+            annotator.text((x + 5, y + 5 + h), text=Path(paths[i]).name[:40], txt_color=(220, 220, 220))  # filenames
+        if len(targets) > 0:
+            idx = targets[:, 0] == i
+            ti = targets[idx]  # image targets
+
+            boxes = xywh2xyxy(ti[:, 2:6]).T
+            classes = ti[:, 1].astype('int')
+            labels = ti.shape[1] == 6  # labels if no conf column
+            conf = None if labels else ti[:, 6]  # check for confidence presence (label vs pred)
+
+            if boxes.shape[1]:
+                if boxes.max() <= 1.01:  # if normalized with tolerance 0.01
+                    boxes[[0, 2]] *= w  # scale to pixels
+                    boxes[[1, 3]] *= h
+                elif scale < 1:  # absolute coords need scale if image scales
+                    boxes *= scale
+            boxes[[0, 2]] += x
+            boxes[[1, 3]] += y
+            for j, box in enumerate(boxes.T.tolist()):
+                cls = classes[j]
+                color = colors(cls)
+                cls = names[cls] if names else cls
+                if labels or conf[j] > 0.25:  # 0.25 conf thresh
+                    label = f'{cls}' if labels else f'{cls} {conf[j]:.1f}'
+                    annotator.box_label(box, label, color=color)
+
+            # Plot masks
+            if len(masks):
+                if masks.max() > 1.0:  # mean that masks are overlap
+                    image_masks = masks[[i]]  # (1, 640, 640)
+                    nl = len(ti)
+                    index = np.arange(nl).reshape(nl, 1, 1) + 1
+                    image_masks = np.repeat(image_masks, nl, axis=0)
+                    image_masks = np.where(image_masks == index, 1.0, 0.0)
+                else:
+                    image_masks = masks[idx]
+
+                im = np.asarray(annotator.im).copy()
+                for j, box in enumerate(boxes.T.tolist()):
+                    if labels or conf[j] > 0.25:  # 0.25 conf thresh
+                        color = colors(classes[j])
+                        mh, mw = image_masks[j].shape
+                        if mh != h or mw != w:
+                            mask = image_masks[j].astype(np.uint8)
+                            mask = cv2.resize(mask, (w, h))
+                            mask = mask.astype(bool)
+                        else:
+                            mask = image_masks[j].astype(bool)
+                        with contextlib.suppress(Exception):
+                            im[y:y + h, x:x + w, :][mask] = im[y:y + h, x:x + w, :][mask] * 0.4 + np.array(color) * 0.6
+                annotator.fromarray(im)
+    annotator.im.save(fname)  # save
+
+
+def plot_results_with_masks(file="path/to/results.csv", dir="", best=True):
+    # Plot training results.csv. Usage: from utils.plots import *; plot_results('path/to/results.csv')
+    save_dir = Path(file).parent if file else Path(dir)
+    fig, ax = plt.subplots(2, 8, figsize=(18, 6), tight_layout=True)
+    ax = ax.ravel()
+    files = list(save_dir.glob("results*.csv"))
+    assert len(files), f"No results.csv files found in {save_dir.resolve()}, nothing to plot."
+    for f in files:
+        try:
+            data = pd.read_csv(f)
+            index = np.argmax(0.9 * data.values[:, 8] + 0.1 * data.values[:, 7] + 0.9 * data.values[:, 12] +
+                              0.1 * data.values[:, 11])
+            s = [x.strip() for x in data.columns]
+            x = data.values[:, 0]
+            for i, j in enumerate([1, 2, 3, 4, 5, 6, 9, 10, 13, 14, 15, 16, 7, 8, 11, 12]):
+                y = data.values[:, j]
+                # y[y == 0] = np.nan  # don't show zero values
+                ax[i].plot(x, y, marker=".", label=f.stem, linewidth=2, markersize=2)
+                if best:
+                    # best
+                    ax[i].scatter(index, y[index], color="r", label=f"best:{index}", marker="*", linewidth=3)
+                    ax[i].set_title(s[j] + f"\n{round(y[index], 5)}")
+                else:
+                    # last
+                    ax[i].scatter(x[-1], y[-1], color="r", label="last", marker="*", linewidth=3)
+                    ax[i].set_title(s[j] + f"\n{round(y[-1], 5)}")
+                # if j in [8, 9, 10]:  # share train and val loss y axes
+                #     ax[i].get_shared_y_axes().join(ax[i], ax[i - 5])
+        except Exception as e:
+            print(f"Warning: Plotting error for {f}: {e}")
+    ax[1].legend()
+    fig.savefig(save_dir / "results.png", dpi=200)
+    plt.close()

utils/torch_utils.py

@@ -0,0 +1,432 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+PyTorch utils
+"""
+
+import math
+import os
+import platform
+import subprocess
+import time
+import warnings
+from contextlib import contextmanager
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.parallel import DistributedDataParallel as DDP
+
+from utils.general import LOGGER, check_version, colorstr, file_date, git_describe
+
+LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
+RANK = int(os.getenv('RANK', -1))
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))
+
+try:
+    import thop  # for FLOPs computation
+except ImportError:
+    thop = None
+
+# Suppress PyTorch warnings
+warnings.filterwarnings('ignore', message='User provided device_type of \'cuda\', but CUDA is not available. Disabling')
+warnings.filterwarnings('ignore', category=UserWarning)
+
+
+def smart_inference_mode(torch_1_9=check_version(torch.__version__, '1.9.0')):
+    # Applies torch.inference_mode() decorator if torch>=1.9.0 else torch.no_grad() decorator
+    def decorate(fn):
+        return (torch.inference_mode if torch_1_9 else torch.no_grad)()(fn)
+
+    return decorate
+
+
+def smartCrossEntropyLoss(label_smoothing=0.0):
+    # Returns nn.CrossEntropyLoss with label smoothing enabled for torch>=1.10.0
+    if check_version(torch.__version__, '1.10.0'):
+        return nn.CrossEntropyLoss(label_smoothing=label_smoothing)
+    if label_smoothing > 0:
+        LOGGER.warning(f'WARNING ⚠️ label smoothing {label_smoothing} requires torch>=1.10.0')
+    return nn.CrossEntropyLoss()
+
+
+def smart_DDP(model):
+    # Model DDP creation with checks
+    assert not check_version(torch.__version__, '1.12.0', pinned=True), \
+        'torch==1.12.0 torchvision==0.13.0 DDP training is not supported due to a known issue. ' \
+        'Please upgrade or downgrade torch to use DDP. See https://github.com/ultralytics/yolov5/issues/8395'
+    if check_version(torch.__version__, '1.11.0'):
+        return DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK, static_graph=True)
+    else:
+        return DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)
+
+
+def reshape_classifier_output(model, n=1000):
+    # Update a TorchVision classification model to class count 'n' if required
+    from models.common import Classify
+    name, m = list((model.model if hasattr(model, 'model') else model).named_children())[-1]  # last module
+    if isinstance(m, Classify):  # YOLOv5 Classify() head
+        if m.linear.out_features != n:
+            m.linear = nn.Linear(m.linear.in_features, n)
+    elif isinstance(m, nn.Linear):  # ResNet, EfficientNet
+        if m.out_features != n:
+            setattr(model, name, nn.Linear(m.in_features, n))
+    elif isinstance(m, nn.Sequential):
+        types = [type(x) for x in m]
+        if nn.Linear in types:
+            i = types.index(nn.Linear)  # nn.Linear index
+            if m[i].out_features != n:
+                m[i] = nn.Linear(m[i].in_features, n)
+        elif nn.Conv2d in types:
+            i = types.index(nn.Conv2d)  # nn.Conv2d index
+            if m[i].out_channels != n:
+                m[i] = nn.Conv2d(m[i].in_channels, n, m[i].kernel_size, m[i].stride, bias=m[i].bias is not None)
+
+
+@contextmanager
+def torch_distributed_zero_first(local_rank: int):
+    # Decorator to make all processes in distributed training wait for each local_master to do something
+    if local_rank not in [-1, 0]:
+        dist.barrier(device_ids=[local_rank])
+    yield
+    if local_rank == 0:
+        dist.barrier(device_ids=[0])
+
+
+def device_count():
+    # Returns number of CUDA devices available. Safe version of torch.cuda.device_count(). Supports Linux and Windows
+    assert platform.system() in ('Linux', 'Windows'), 'device_count() only supported on Linux or Windows'
+    try:
+        cmd = 'nvidia-smi -L | wc -l' if platform.system() == 'Linux' else 'nvidia-smi -L | find /c /v ""'  # Windows
+        return int(subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1])
+    except Exception:
+        return 0
+
+
+def select_device(device='', batch_size=0, newline=True):
+    # device = None or 'cpu' or 0 or '0' or '0,1,2,3'
+    s = f'YOLOv5 🚀 {git_describe() or file_date()} Python-{platform.python_version()} torch-{torch.__version__} '
+    device = str(device).strip().lower().replace('cuda:', '').replace('none', '')  # to string, 'cuda:0' to '0'
+    cpu = device == 'cpu'
+    mps = device == 'mps'  # Apple Metal Performance Shaders (MPS)
+    if cpu or mps:
+        os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # force torch.cuda.is_available() = False
+    elif device:  # non-cpu device requested
+        os.environ['CUDA_VISIBLE_DEVICES'] = device  # set environment variable - must be before assert is_available()
+        assert torch.cuda.is_available() and torch.cuda.device_count() >= len(device.replace(',', '')), \
+            f"Invalid CUDA '--device {device}' requested, use '--device cpu' or pass valid CUDA device(s)"
+
+    if not cpu and not mps and torch.cuda.is_available():  # prefer GPU if available
+        devices = device.split(',') if device else '0'  # range(torch.cuda.device_count())  # i.e. 0,1,6,7
+        n = len(devices)  # device count
+        if n > 1 and batch_size > 0:  # check batch_size is divisible by device_count
+            assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}'
+        space = ' ' * (len(s) + 1)
+        for i, d in enumerate(devices):
+            p = torch.cuda.get_device_properties(i)
+            s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / (1 << 20):.0f}MiB)\n"  # bytes to MB
+        arg = 'cuda:0'
+    elif mps and getattr(torch, 'has_mps', False) and torch.backends.mps.is_available():  # prefer MPS if available
+        s += 'MPS\n'
+        arg = 'mps'
+    else:  # revert to CPU
+        s += 'CPU\n'
+        arg = 'cpu'
+
+    if not newline:
+        s = s.rstrip()
+    LOGGER.info(s)
+    return torch.device(arg)
+
+
+def time_sync():
+    # PyTorch-accurate time
+    if torch.cuda.is_available():
+        torch.cuda.synchronize()
+    return time.time()
+
+
+def profile(input, ops, n=10, device=None):
+    """ YOLOv5 speed/memory/FLOPs profiler
+    Usage:
+        input = torch.randn(16, 3, 640, 640)
+        m1 = lambda x: x * torch.sigmoid(x)
+        m2 = nn.SiLU()
+        profile(input, [m1, m2], n=100)  # profile over 100 iterations
+    """
+    results = []
+    if not isinstance(device, torch.device):
+        device = select_device(device)
+    print(f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}"
+          f"{'input':>24s}{'output':>24s}")
+
+    for x in input if isinstance(input, list) else [input]:
+        x = x.to(device)
+        x.requires_grad = True
+        for m in ops if isinstance(ops, list) else [ops]:
+            m = m.to(device) if hasattr(m, 'to') else m  # device
+            m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m
+            tf, tb, t = 0, 0, [0, 0, 0]  # dt forward, backward
+            try:
+                flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # GFLOPs
+            except Exception:
+                flops = 0
+
+            try:
+                for _ in range(n):
+                    t[0] = time_sync()
+                    y = m(x)
+                    t[1] = time_sync()
+                    try:
+                        _ = (sum(yi.sum() for yi in y) if isinstance(y, list) else y).sum().backward()
+                        t[2] = time_sync()
+                    except Exception:  # no backward method
+                        # print(e)  # for debug
+                        t[2] = float('nan')
+                    tf += (t[1] - t[0]) * 1000 / n  # ms per op forward
+                    tb += (t[2] - t[1]) * 1000 / n  # ms per op backward
+                mem = torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0  # (GB)
+                s_in, s_out = (tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' for x in (x, y))  # shapes
+                p = sum(x.numel() for x in m.parameters()) if isinstance(m, nn.Module) else 0  # parameters
+                print(f'{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}')
+                results.append([p, flops, mem, tf, tb, s_in, s_out])
+            except Exception as e:
+                print(e)
+                results.append(None)
+            torch.cuda.empty_cache()
+    return results
+
+
+def is_parallel(model):
+    # Returns True if model is of type DP or DDP
+    return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)
+
+
+def de_parallel(model):
+    # De-parallelize a model: returns single-GPU model if model is of type DP or DDP
+    return model.module if is_parallel(model) else model
+
+
+def initialize_weights(model):
+    for m in model.modules():
+        t = type(m)
+        if t is nn.Conv2d:
+            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+        elif t is nn.BatchNorm2d:
+            m.eps = 1e-3
+            m.momentum = 0.03
+        elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
+            m.inplace = True
+
+
+def find_modules(model, mclass=nn.Conv2d):
+    # Finds layer indices matching module class 'mclass'
+    return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]
+
+
+def sparsity(model):
+    # Return global model sparsity
+    a, b = 0, 0
+    for p in model.parameters():
+        a += p.numel()
+        b += (p == 0).sum()
+    return b / a
+
+
+def prune(model, amount=0.3):
+    # Prune model to requested global sparsity
+    import torch.nn.utils.prune as prune
+    for name, m in model.named_modules():
+        if isinstance(m, nn.Conv2d):
+            prune.l1_unstructured(m, name='weight', amount=amount)  # prune
+            prune.remove(m, 'weight')  # make permanent
+    LOGGER.info(f'Model pruned to {sparsity(model):.3g} global sparsity')
+
+
+def fuse_conv_and_bn(conv, bn):
+    # Fuse Conv2d() and BatchNorm2d() layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
+    fusedconv = nn.Conv2d(conv.in_channels,
+                          conv.out_channels,
+                          kernel_size=conv.kernel_size,
+                          stride=conv.stride,
+                          padding=conv.padding,
+                          dilation=conv.dilation,
+                          groups=conv.groups,
+                          bias=True).requires_grad_(False).to(conv.weight.device)
+
+    # Prepare filters
+    w_conv = conv.weight.clone().view(conv.out_channels, -1)
+    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
+    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))
+
+    # Prepare spatial bias
+    b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
+    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
+    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
+
+    return fusedconv
+
+
+def model_info(model, verbose=False, imgsz=640):
+    # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
+    n_p = sum(x.numel() for x in model.parameters())  # number parameters
+    n_g = sum(x.numel() for x in model.parameters() if x.requires_grad)  # number gradients
+    if verbose:
+        print(f"{'layer':>5} {'name':>40} {'gradient':>9} {'parameters':>12} {'shape':>20} {'mu':>10} {'sigma':>10}")
+        for i, (name, p) in enumerate(model.named_parameters()):
+            name = name.replace('module_list.', '')
+            print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
+                  (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
+
+    try:  # FLOPs
+        p = next(model.parameters())
+        stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32  # max stride
+        im = torch.empty((1, p.shape[1], stride, stride), device=p.device)  # input image in BCHW format
+        flops = thop.profile(deepcopy(model), inputs=(im,), verbose=False)[0] / 1E9 * 2  # stride GFLOPs
+        imgsz = imgsz if isinstance(imgsz, list) else [imgsz, imgsz]  # expand if int/float
+        fs = f', {flops * imgsz[0] / stride * imgsz[1] / stride:.1f} GFLOPs'  # 640x640 GFLOPs
+    except Exception:
+        fs = ''
+
+    name = Path(model.yaml_file).stem.replace('yolov5', 'YOLOv5') if hasattr(model, 'yaml_file') else 'Model'
+    LOGGER.info(f"{name} summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
+
+
+def scale_img(img, ratio=1.0, same_shape=False, gs=32):  # img(16,3,256,416)
+    # Scales img(bs,3,y,x) by ratio constrained to gs-multiple
+    if ratio == 1.0:
+        return img
+    h, w = img.shape[2:]
+    s = (int(h * ratio), int(w * ratio))  # new size
+    img = F.interpolate(img, size=s, mode='bilinear', align_corners=False)  # resize
+    if not same_shape:  # pad/crop img
+        h, w = (math.ceil(x * ratio / gs) * gs for x in (h, w))
+    return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean
+
+
+def copy_attr(a, b, include=(), exclude=()):
+    # Copy attributes from b to a, options to only include [...] and to exclude [...]
+    for k, v in b.__dict__.items():
+        if (len(include) and k not in include) or k.startswith('_') or k in exclude:
+            continue
+        else:
+            setattr(a, k, v)
+
+
+def smart_optimizer(model, name='Adam', lr=0.001, momentum=0.9, decay=1e-5):
+    # YOLOv5 3-param group optimizer: 0) weights with decay, 1) weights no decay, 2) biases no decay
+    g = [], [], []  # optimizer parameter groups
+    bn = tuple(v for k, v in nn.__dict__.items() if 'Norm' in k)  # normalization layers, i.e. BatchNorm2d()
+    for v in model.modules():
+        for p_name, p in v.named_parameters(recurse=0):
+            if p_name == 'bias':  # bias (no decay)
+                g[2].append(p)
+            elif p_name == 'weight' and isinstance(v, bn):  # weight (no decay)
+                g[1].append(p)
+            else:
+                g[0].append(p)  # weight (with decay)
+
+    if name == 'Adam':
+        optimizer = torch.optim.Adam(g[2], lr=lr, betas=(momentum, 0.999))  # adjust beta1 to momentum
+    elif name == 'AdamW':
+        optimizer = torch.optim.AdamW(g[2], lr=lr, betas=(momentum, 0.999), weight_decay=0.0)
+    elif name == 'RMSProp':
+        optimizer = torch.optim.RMSprop(g[2], lr=lr, momentum=momentum)
+    elif name == 'SGD':
+        optimizer = torch.optim.SGD(g[2], lr=lr, momentum=momentum, nesterov=True)
+    else:
+        raise NotImplementedError(f'Optimizer {name} not implemented.')
+
+    optimizer.add_param_group({'params': g[0], 'weight_decay': decay})  # add g0 with weight_decay
+    optimizer.add_param_group({'params': g[1], 'weight_decay': 0.0})  # add g1 (BatchNorm2d weights)
+    LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__}(lr={lr}) with parameter groups "
+                f"{len(g[1])} weight(decay=0.0), {len(g[0])} weight(decay={decay}), {len(g[2])} bias")
+    return optimizer
+
+
+def smart_hub_load(repo='ultralytics/yolov5', model='yolov5s', **kwargs):
+    # YOLOv5 torch.hub.load() wrapper with smart error/issue handling
+    if check_version(torch.__version__, '1.9.1'):
+        kwargs['skip_validation'] = True  # validation causes GitHub API rate limit errors
+    if check_version(torch.__version__, '1.12.0'):
+        kwargs['trust_repo'] = True  # argument required starting in torch 0.12
+    try:
+        return torch.hub.load(repo, model, **kwargs)
+    except Exception:
+        return torch.hub.load(repo, model, force_reload=True, **kwargs)
+
+
+def smart_resume(ckpt, optimizer, ema=None, weights='yolov5s.pt', epochs=300, resume=True):
+    # Resume training from a partially trained checkpoint
+    best_fitness = 0.0
+    start_epoch = ckpt['epoch'] + 1
+    if ckpt['optimizer'] is not None:
+        optimizer.load_state_dict(ckpt['optimizer'])  # optimizer
+        best_fitness = ckpt['best_fitness']
+    if ema and ckpt.get('ema'):
+        ema.ema.load_state_dict(ckpt['ema'].float().state_dict())  # EMA
+        ema.updates = ckpt['updates']
+    if resume:
+        assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.\n' \
+                                f"Start a new training without --resume, i.e. 'python train.py --weights {weights}'"
+        LOGGER.info(f'Resuming training from {weights} from epoch {start_epoch} to {epochs} total epochs')
+    if epochs < start_epoch:
+        LOGGER.info(f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs.")
+        epochs += ckpt['epoch']  # finetune additional epochs
+    return best_fitness, start_epoch, epochs
+
+
+class EarlyStopping:
+    # YOLOv5 simple early stopper
+    def __init__(self, patience=30):
+        self.best_fitness = 0.0  # i.e. mAP
+        self.best_epoch = 0
+        self.patience = patience or float('inf')  # epochs to wait after fitness stops improving to stop
+        self.possible_stop = False  # possible stop may occur next epoch
+
+    def __call__(self, epoch, fitness):
+        if fitness >= self.best_fitness:  # >= 0 to allow for early zero-fitness stage of training
+            self.best_epoch = epoch
+            self.best_fitness = fitness
+        delta = epoch - self.best_epoch  # epochs without improvement
+        self.possible_stop = delta >= (self.patience - 1)  # possible stop may occur next epoch
+        stop = delta >= self.patience  # stop training if patience exceeded
+        if stop:
+            LOGGER.info(f'Stopping training early as no improvement observed in last {self.patience} epochs. '
+                        f'Best results observed at epoch {self.best_epoch}, best model saved as best.pt.\n'
+                        f'To update EarlyStopping(patience={self.patience}) pass a new patience value, '
+                        f'i.e. `python train.py --patience 300` or use `--patience 0` to disable EarlyStopping.')
+        return stop
+
+
+class ModelEMA:
+    """ Updated Exponential Moving Average (EMA) from https://github.com/rwightman/pytorch-image-models
+    Keeps a moving average of everything in the model state_dict (parameters and buffers)
+    For EMA details see https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
+    """
+
+    def __init__(self, model, decay=0.9999, tau=2000, updates=0):
+        # Create EMA
+        self.ema = deepcopy(de_parallel(model)).eval()  # FP32 EMA
+        self.updates = updates  # number of EMA updates
+        self.decay = lambda x: decay * (1 - math.exp(-x / tau))  # decay exponential ramp (to help early epochs)
+        for p in self.ema.parameters():
+            p.requires_grad_(False)
+
+    def update(self, model):
+        # Update EMA parameters
+        self.updates += 1
+        d = self.decay(self.updates)
+
+        msd = de_parallel(model).state_dict()  # model state_dict
+        for k, v in self.ema.state_dict().items():
+            if v.dtype.is_floating_point:  # true for FP16 and FP32
+                v *= d
+                v += (1 - d) * msd[k].detach()
+        # assert v.dtype == msd[k].dtype == torch.float32, f'{k}: EMA {v.dtype} and model {msd[k].dtype} must be FP32'
+
+    def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):
+        # Update EMA attributes
+        copy_attr(self.ema, model, include, exclude)

utils/triton.py

@@ -0,0 +1,85 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+""" Utils to interact with the Triton Inference Server
+"""
+
+import typing
+from urllib.parse import urlparse
+
+import torch
+
+
+class TritonRemoteModel:
+    """ A wrapper over a model served by the Triton Inference Server. It can
+    be configured to communicate over GRPC or HTTP. It accepts Torch Tensors
+    as input and returns them as outputs.
+    """
+
+    def __init__(self, url: str):
+        """
+        Keyword arguments:
+        url: Fully qualified address of the Triton server - for e.g. grpc://localhost:8000
+        """
+
+        parsed_url = urlparse(url)
+        if parsed_url.scheme == "grpc":
+            from tritonclient.grpc import InferenceServerClient, InferInput
+
+            self.client = InferenceServerClient(parsed_url.netloc)  # Triton GRPC client
+            model_repository = self.client.get_model_repository_index()
+            self.model_name = model_repository.models[0].name
+            self.metadata = self.client.get_model_metadata(self.model_name, as_json=True)
+
+            def create_input_placeholders() -> typing.List[InferInput]:
+                return [
+                    InferInput(i['name'], [int(s) for s in i["shape"]], i['datatype']) for i in self.metadata['inputs']]
+
+        else:
+            from tritonclient.http import InferenceServerClient, InferInput
+
+            self.client = InferenceServerClient(parsed_url.netloc)  # Triton HTTP client
+            model_repository = self.client.get_model_repository_index()
+            self.model_name = model_repository[0]['name']
+            self.metadata = self.client.get_model_metadata(self.model_name)
+
+            def create_input_placeholders() -> typing.List[InferInput]:
+                return [
+                    InferInput(i['name'], [int(s) for s in i["shape"]], i['datatype']) for i in self.metadata['inputs']]
+
+        self._create_input_placeholders_fn = create_input_placeholders
+
+    @property
+    def runtime(self):
+        """Returns the model runtime"""
+        return self.metadata.get("backend", self.metadata.get("platform"))
+
+    def __call__(self, *args, **kwargs) -> typing.Union[torch.Tensor, typing.Tuple[torch.Tensor, ...]]:
+        """ Invokes the model. Parameters can be provided via args or kwargs.
+        args, if provided, are assumed to match the order of inputs of the model.
+        kwargs are matched with the model input names.
+        """
+        inputs = self._create_inputs(*args, **kwargs)
+        response = self.client.infer(model_name=self.model_name, inputs=inputs)
+        result = []
+        for output in self.metadata['outputs']:
+            tensor = torch.as_tensor(response.as_numpy(output['name']))
+            result.append(tensor)
+        return result[0] if len(result) == 1 else result
+
+    def _create_inputs(self, *args, **kwargs):
+        args_len, kwargs_len = len(args), len(kwargs)
+        if not args_len and not kwargs_len:
+            raise RuntimeError("No inputs provided.")
+        if args_len and kwargs_len:
+            raise RuntimeError("Cannot specify args and kwargs at the same time")
+
+        placeholders = self._create_input_placeholders_fn()
+        if args_len:
+            if args_len != len(placeholders):
+                raise RuntimeError(f"Expected {len(placeholders)} inputs, got {args_len}.")
+            for input, value in zip(placeholders, args):
+                input.set_data_from_numpy(value.cpu().numpy())
+        else:
+            for input in placeholders:
+                value = kwargs[input.name]
+                input.set_data_from_numpy(value.cpu().numpy())
+        return placeholders