158 lines
4 KiB
Python
158 lines
4 KiB
Python
import torch
|
|
import matplotlib.pyplot as plt
|
|
import matplotlib.patches as patch
|
|
import matplotlib as mpl
|
|
from functions.analyse_network import analyse_network
|
|
|
|
mpl.rcParams["text.usetex"] = True
|
|
mpl.rcParams["font.family"] = "serif"
|
|
|
|
|
|
# define parameters
|
|
num_iterations: int = 100000
|
|
learning_rate: float = 0.1
|
|
apply_input_mask: bool = True
|
|
mark_region_in_plot: bool = False
|
|
sheduler_patience: int = 500
|
|
sheduler_factor: float = 0.9
|
|
sheduler_eps = 1e-08
|
|
target_image_active: float = 1e4
|
|
|
|
# path to NN
|
|
nn = "network_0_seed0_Coignless_83Epoch_2807-1455"
|
|
PATH = f"./trained_models/{nn}.pt"
|
|
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
|
|
|
|
|
# load and eval model
|
|
model = torch.load(PATH).to(device)
|
|
model.eval()
|
|
print("Full network:")
|
|
print(model)
|
|
print("")
|
|
|
|
# enter index to plot:
|
|
idx = int(input("Please select layer: "))
|
|
print(f"Selected layer {idx}:")
|
|
assert idx < len(model)
|
|
model = model[: idx + 1]
|
|
|
|
# random input
|
|
input_img = torch.randn(1, 200, 200).to(device)
|
|
input_img = input_img.unsqueeze(0)
|
|
input_img.requires_grad_(True) # type: ignore
|
|
|
|
input_shape = input_img.shape
|
|
assert input_shape[-2] == input_shape[-1]
|
|
coordinate_list, layer_type_list, pixel_used = analyse_network(
|
|
model=model, input_shape=int(input_shape[-1])
|
|
)
|
|
|
|
output_shape = model(input_img).shape
|
|
target_image = torch.zeros(
|
|
(*output_shape,), dtype=input_img.dtype, device=input_img.device
|
|
)
|
|
|
|
input_parameter = torch.nn.Parameter(input_img)
|
|
|
|
|
|
print(
|
|
(
|
|
f"Available max positions: f:{target_image.shape[1]} "
|
|
f"x:{target_image.shape[2]} y:{target_image.shape[3]}"
|
|
)
|
|
)
|
|
|
|
# select neuron and plot for all feature maps (?)
|
|
neuron_f = int(input("Please select neuron_f: "))
|
|
neuron_x = target_image.shape[2] // 2
|
|
neuron_y = target_image.shape[3] // 2
|
|
print(f"Selected neuron {neuron_f}, {neuron_x}, {neuron_y}")
|
|
|
|
# Input mask ->
|
|
active_input_x = coordinate_list[-1][:, neuron_x].clone()
|
|
active_input_y = coordinate_list[-1][:, neuron_y].clone()
|
|
|
|
input_mask: torch.Tensor = torch.zeros_like(input_img)
|
|
|
|
input_mask[
|
|
:,
|
|
:,
|
|
active_input_x.type(torch.int64).unsqueeze(-1),
|
|
active_input_y.type(torch.int64).unsqueeze(0),
|
|
] = 1
|
|
|
|
rect_x = [int(active_input_x.min()), int(active_input_x.max())]
|
|
rect_y = [int(active_input_y.min()), int(active_input_y.max())]
|
|
# <- Input mask
|
|
|
|
if apply_input_mask:
|
|
with torch.no_grad():
|
|
input_img *= input_mask
|
|
|
|
|
|
optimizer = torch.optim.Adam([{"params": input_parameter}], lr=learning_rate)
|
|
|
|
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
|
|
optimizer,
|
|
patience=sheduler_patience,
|
|
factor=sheduler_factor,
|
|
eps=sheduler_eps * 0.1,
|
|
)
|
|
|
|
target_image[0, neuron_f, neuron_x, neuron_y] = target_image_active
|
|
|
|
counter: int = 0
|
|
while (optimizer.param_groups[0]["lr"] > sheduler_eps) and (counter < num_iterations):
|
|
optimizer.zero_grad()
|
|
|
|
output = model(input_parameter)
|
|
|
|
loss = torch.nn.functional.mse_loss(output, target_image)
|
|
loss.backward()
|
|
|
|
if counter % 1000 == 0:
|
|
print(
|
|
f"{counter} : loss={float(loss):.3e} lr={optimizer.param_groups[0]['lr']:.3e}"
|
|
)
|
|
|
|
optimizer.step()
|
|
|
|
if apply_input_mask:
|
|
with torch.no_grad():
|
|
input_parameter.data[torch.where(input_mask == 0)] = 0.0
|
|
|
|
with torch.no_grad():
|
|
max_data = torch.abs(input_parameter.data).max()
|
|
if max_data > 1.0:
|
|
input_parameter.data /= max_data
|
|
|
|
if (
|
|
torch.isfinite(input_parameter.data).sum().cpu()
|
|
!= torch.tensor(input_parameter.data.size()).prod()
|
|
):
|
|
print(f"Found NaN in step: {counter}, use a smaller initial lr")
|
|
exit()
|
|
|
|
scheduler.step(float(loss))
|
|
counter += 1
|
|
|
|
# plot image:
|
|
_, ax = plt.subplots()
|
|
|
|
ax.imshow(input_img.squeeze().detach().cpu().numpy(), cmap="gray")
|
|
|
|
if mark_region_in_plot:
|
|
edgecolor = "sienna"
|
|
kernel = patch.Rectangle(
|
|
(rect_y[0], rect_x[0]),
|
|
int(rect_y[1] - rect_y[0]),
|
|
int(rect_x[1] - rect_x[0]),
|
|
linewidth=1.2,
|
|
edgecolor=edgecolor,
|
|
facecolor="none",
|
|
)
|
|
ax.add_patch(kernel)
|
|
|
|
|
|
plt.show(block=True)
|