import torch from network.CPP.PySpikeGeneration2DManyIP import SpikeGeneration2DManyIP from network.CPP.PyHDynamicCNNManyIP import HDynamicCNNManyIP from network.calculate_output_size import calculate_output_size class SbS(torch.nn.Module): _epsilon_xy: torch.Tensor | None = None _epsilon_0: float _epsilon_t: torch.Tensor | None = None _weights: torch.nn.parameter.Parameter _weights_exists: bool = False _kernel_size: list[int] _stride: list[int] _dilation: list[int] _padding: list[int] _output_size: torch.Tensor _number_of_spikes: int _number_of_cpu_processes: int _number_of_neurons: int _number_of_input_neurons: int _epsilon_xy_intitial: float _h_initial: torch.Tensor | None = None _w_trainable: bool # _last_grad_scale: torch.nn.parameter.Parameter # _keep_last_grad_scale: bool # _disable_scale_grade: bool _forgetting_offset: torch.Tensor | None = None _weight_noise_range: list[float] _skip_gradient_calculation: bool _is_pooling_layer: bool _input_size: list[int] _output_layer: bool = False _local_learning: bool = False device: torch.device default_dtype: torch.dtype _gpu_tuning_factor: int _max_grad_weights: torch.Tensor | None = None _number_of_grad_weight_contributions: float = 0.0 def __init__( self, number_of_input_neurons: int, number_of_neurons: int, input_size: list[int], forward_kernel_size: list[int], number_of_spikes: int, epsilon_t: torch.Tensor, epsilon_xy_intitial: float = 0.1, epsilon_0: float = 1.0, weight_noise_range: list[float] = [0.0, 1.0], is_pooling_layer: bool = False, strides: list[int] = [1, 1], dilation: list[int] = [0, 0], padding: list[int] = [0, 0], number_of_cpu_processes: int = 1, w_trainable: bool = False, # keep_last_grad_scale: bool = False, # disable_scale_grade: bool = True, forgetting_offset: float = -1.0, skip_gradient_calculation: bool = False, device: torch.device | None = None, default_dtype: torch.dtype | None = None, gpu_tuning_factor: int = 5, ) -> None: super().__init__() assert device is not None assert default_dtype is not None self.device = device self.default_dtype = default_dtype self._w_trainable = bool(w_trainable) # self._keep_last_grad_scale = bool(keep_last_grad_scale) self._skip_gradient_calculation = bool(skip_gradient_calculation) # self._disable_scale_grade = bool(disable_scale_grade) self._epsilon_xy_intitial = float(epsilon_xy_intitial) self._stride = strides self._dilation = dilation self._padding = padding self._kernel_size = forward_kernel_size self._number_of_input_neurons = int(number_of_input_neurons) self._number_of_neurons = int(number_of_neurons) self._epsilon_0 = float(epsilon_0) self._number_of_cpu_processes = int(number_of_cpu_processes) self._number_of_spikes = int(number_of_spikes) self._weight_noise_range = weight_noise_range self._is_pooling_layer = bool(is_pooling_layer) assert len(input_size) == 2 self._input_size = input_size # The GPU hates me... # Too many SbS threads == bad # Thus I need to limit them... # (Reminder: We cannot access the mini-batch size here, # which is part of the GPU thread size calculation...) if (self._input_size[0] * self._input_size[1]) > gpu_tuning_factor: self._gpu_tuning_factor = gpu_tuning_factor else: self._gpu_tuning_factor = 0 # self._last_grad_scale = torch.nn.parameter.Parameter( # torch.tensor(-1.0, dtype=self.default_dtype), # requires_grad=True, # ) self._forgetting_offset = torch.tensor( forgetting_offset, dtype=self.default_dtype, device=self.device ) self.epsilon_t = epsilon_t.type(dtype=self.default_dtype).to(device=self.device) self._output_size = calculate_output_size( value=input_size, kernel_size=self._kernel_size, stride=self._stride, dilation=self._dilation, padding=self._padding, ) self.set_h_init_to_uniform() self.functional_sbs = FunctionalSbS.apply # ############################################################### # Initialize the weights # ############################################################### if self._is_pooling_layer is True: self.weights = self._make_pooling_weights() else: assert len(self._weight_noise_range) == 2 weights = torch.empty( ( int(self._kernel_size[0]) * int(self._kernel_size[1]) * int(self._number_of_input_neurons), int(self._number_of_neurons), ), dtype=self.default_dtype, device=self.device, ) torch.nn.init.uniform_( weights, a=float(self._weight_noise_range[0]), b=float(self._weight_noise_range[1]), ) self.weights = weights #################################################################### # Variables in and out # #################################################################### @property def epsilon_t(self) -> torch.Tensor | None: return self._epsilon_t @epsilon_t.setter def epsilon_t(self, value: torch.Tensor): assert value is not None assert torch.is_tensor(value) is True assert value.dim() == 1 assert value.dtype == self.default_dtype self._epsilon_t = ( value.detach() .clone(memory_format=torch.contiguous_format) .type(dtype=self.default_dtype) .to(device=self.device) .requires_grad_(False) ) @property def weights(self) -> torch.Tensor | None: if self._weights_exists is False: return None else: return self._weights @weights.setter def weights(self, value: torch.Tensor): assert value is not None assert torch.is_tensor(value) is True assert value.dim() == 2 temp: torch.Tensor = ( value.detach() .clone(memory_format=torch.contiguous_format) .type(dtype=self.default_dtype) .to(device=self.device) ) temp /= temp.sum(dim=0, keepdim=True, dtype=self.default_dtype) if self._weights_exists is False: self._weights = torch.nn.parameter.Parameter(temp, requires_grad=True) self._weights_exists = True else: self._weights.data = temp @property def h_initial(self) -> torch.Tensor | None: return self._h_initial @h_initial.setter def h_initial(self, value: torch.Tensor): assert value is not None assert torch.is_tensor(value) is True assert value.dim() == 1 assert value.dtype == self.default_dtype self._h_initial = ( value.detach() .clone(memory_format=torch.contiguous_format) .type(dtype=self.default_dtype) .to(device=self.device) .requires_grad_(False) ) def update_pre_care(self): if self._weights.grad is not None: assert self._number_of_grad_weight_contributions > 0 self._weights.grad /= self._number_of_grad_weight_contributions self._number_of_grad_weight_contributions = 0.0 def update_after_care(self, threshold_weight: float): if self._w_trainable is True: self.norm_weights() self.threshold_weights(threshold_weight) self.norm_weights() # def after_batch(self, new_state: bool = False): # if self._keep_last_grad_scale is True: # self._last_grad_scale.data = self._last_grad_scale.grad # self._keep_last_grad_scale = new_state # self._last_grad_scale.grad = torch.zeros_like(self._last_grad_scale.grad) #################################################################### # Helper functions # #################################################################### def _make_pooling_weights(self) -> torch.Tensor: """For generating the pooling weights.""" assert self._number_of_neurons is not None assert self._kernel_size is not None weights: torch.Tensor = torch.zeros( ( int(self._kernel_size[0]), int(self._kernel_size[1]), int(self._number_of_neurons), int(self._number_of_neurons), ), dtype=self.default_dtype, device=self.device, ) for i in range(0, int(self._number_of_neurons)): weights[:, :, i, i] = 1.0 weights = weights.moveaxis(-1, 0).moveaxis(-1, 1) weights = torch.nn.functional.unfold( input=weights, kernel_size=(int(self._kernel_size[0]), int(self._kernel_size[1])), dilation=(1, 1), padding=(0, 0), stride=(1, 1), ).squeeze() weights = torch.moveaxis(weights, 0, 1) return weights def set_h_init_to_uniform(self) -> None: assert self._number_of_neurons > 2 self.h_initial: torch.Tensor = torch.full( (self._number_of_neurons,), (1.0 / float(self._number_of_neurons)), dtype=self.default_dtype, device=self.device, ) def norm_weights(self) -> None: assert self._weights_exists is True temp: torch.Tensor = ( self._weights.data.detach() .clone(memory_format=torch.contiguous_format) .type(dtype=self.default_dtype) .to(device=self.device) ) temp /= temp.sum(dim=0, keepdim=True, dtype=self.default_dtype) self._weights.data = temp def threshold_weights(self, threshold: float) -> None: assert self._weights_exists is True assert threshold >= 0 torch.clamp( self._weights.data, min=float(threshold), max=None, out=self._weights.data, ) #################################################################### # Forward # #################################################################### def forward( self, input: torch.Tensor, labels: torch.Tensor | None = None ) -> torch.Tensor: # Are we happy with the input? assert input is not None assert torch.is_tensor(input) is True assert input.dim() == 4 assert input.dtype == self.default_dtype assert input.shape[1] == self._number_of_input_neurons assert input.shape[2] == self._input_size[0] assert input.shape[3] == self._input_size[1] # Are we happy with the rest of the network? assert self._epsilon_0 is not None assert self._epsilon_t is not None assert self._h_initial is not None assert self._forgetting_offset is not None assert self._weights_exists is True assert self._weights is not None input_convolved = torch.nn.functional.fold( torch.nn.functional.unfold( input.requires_grad_(True), kernel_size=(int(self._kernel_size[0]), int(self._kernel_size[1])), dilation=(int(self._dilation[0]), int(self._dilation[1])), padding=(int(self._padding[0]), int(self._padding[1])), stride=(int(self._stride[0]), int(self._stride[1])), ), output_size=tuple(self._output_size.tolist()), kernel_size=(1, 1), dilation=(1, 1), padding=(0, 0), stride=(1, 1), ) epsilon_t_0: torch.Tensor = ( (self._epsilon_t * self._epsilon_0).type(input.dtype).to(input.device) ) parameter_list = torch.tensor( [ int(self._w_trainable), # 0 int(0), # int(self._disable_scale_grade), # 1 int(0), # int(self._keep_last_grad_scale), # 2 int(self._skip_gradient_calculation), # 3 int(self._number_of_spikes), # 4 int(self._number_of_cpu_processes), # 5 int(self._output_size[0]), # 6 int(self._output_size[1]), # 7 int(self._gpu_tuning_factor), # 8 int(self._output_layer), # 9 int(self._local_learning), # 10 ], dtype=torch.int64, ) if self._epsilon_xy is None: self._epsilon_xy = torch.full( ( input_convolved.shape[1], input_convolved.shape[2], input_convolved.shape[3], ), float(self._epsilon_xy_intitial), dtype=self.default_dtype, device=self.device, ) assert self._epsilon_xy is not None # In the case somebody tried to replace the matrix with wrong dimensions assert self._epsilon_xy.shape[0] == input_convolved.shape[1] assert self._epsilon_xy.shape[1] == input_convolved.shape[2] assert self._epsilon_xy.shape[2] == input_convolved.shape[3] # SbS forward functional output = self.functional_sbs( input_convolved, self._epsilon_xy, epsilon_t_0, self._weights, self._h_initial, parameter_list, # self._last_grad_scale, self._forgetting_offset, ) self._number_of_grad_weight_contributions += ( output.shape[0] * output.shape[-2] * output.shape[-1] ) return output class FunctionalSbS(torch.autograd.Function): @staticmethod def forward( # type: ignore ctx, input: torch.Tensor, epsilon_xy: torch.Tensor, epsilon_t_0: torch.Tensor, weights: torch.Tensor, h_initial: torch.Tensor, parameter_list: torch.Tensor, # grad_output_scale: torch.Tensor, forgetting_offset: torch.Tensor, ) -> torch.Tensor: assert input.dim() == 4 number_of_spikes: int = int(parameter_list[4]) if input.device == torch.device("cpu"): spike_number_of_cpu_processes: int = int(parameter_list[5]) else: spike_number_of_cpu_processes = -1 if input.device == torch.device("cpu"): hdyn_number_of_cpu_processes: int = int(parameter_list[5]) else: hdyn_number_of_cpu_processes = -1 output_size_0: int = int(parameter_list[6]) output_size_1: int = int(parameter_list[7]) gpu_tuning_factor: int = int(parameter_list[8]) # ########################################################### # Spike generation # ########################################################### # ############################################ # Normalized cumsum # (beware of the pytorch bug! Thus .clone()!) # ############################################ input_cumsum: torch.Tensor = torch.cumsum(input, dim=1, dtype=input.dtype) input_cumsum_last: torch.Tensor = input_cumsum[:, -1, :, :].unsqueeze(1).clone() input_cumsum /= input_cumsum_last # ############################################ # Get the required random numbers # ############################################ random_values = torch.rand( size=[ input_cumsum.shape[0], number_of_spikes, input_cumsum.shape[2], input_cumsum.shape[3], ], dtype=input.dtype, device=input.device, ) # ############################################ # Make space for the results # ############################################ spikes = torch.empty_like(random_values, dtype=torch.int64, device=input.device) assert input_cumsum.is_contiguous() is True assert random_values.is_contiguous() is True assert spikes.is_contiguous() is True # time_start: float = time.perf_counter() spike_generation: SpikeGeneration2DManyIP = SpikeGeneration2DManyIP() spike_generation.spike_generation( input_cumsum.data_ptr(), int(input_cumsum.shape[0]), int(input_cumsum.shape[1]), int(input_cumsum.shape[2]), int(input_cumsum.shape[3]), random_values.data_ptr(), int(random_values.shape[0]), int(random_values.shape[1]), int(random_values.shape[2]), int(random_values.shape[3]), spikes.data_ptr(), int(spikes.shape[0]), int(spikes.shape[1]), int(spikes.shape[2]), int(spikes.shape[3]), int(spike_number_of_cpu_processes), ) del random_values del input_cumsum # ########################################################### # H dynamic # ########################################################### assert epsilon_t_0.ndim == 1 assert epsilon_t_0.shape[0] >= number_of_spikes # ############################################ # Make space for the results # ############################################ output = torch.empty( ( int(input.shape[0]), int(weights.shape[1]), output_size_0, output_size_1, ), dtype=input.dtype, device=input.device, ) assert output.is_contiguous() is True assert epsilon_xy.is_contiguous() is True assert epsilon_t_0.is_contiguous() is True assert weights.is_contiguous() is True assert spikes.is_contiguous() is True assert h_initial.is_contiguous() is True assert epsilon_xy.ndim == 3 assert weights.ndim == 2 assert h_initial.ndim == 1 h_dynamic: HDynamicCNNManyIP = HDynamicCNNManyIP() h_dynamic.update( output.data_ptr(), int(output.shape[0]), int(output.shape[1]), int(output.shape[2]), int(output.shape[3]), epsilon_xy.data_ptr(), int(epsilon_xy.shape[0]), int(epsilon_xy.shape[1]), int(epsilon_xy.shape[2]), epsilon_t_0.data_ptr(), int(epsilon_t_0.shape[0]), weights.data_ptr(), int(weights.shape[0]), int(weights.shape[1]), spikes.data_ptr(), int(spikes.shape[0]), int(spikes.shape[1]), int(spikes.shape[2]), int(spikes.shape[3]), h_initial.data_ptr(), int(h_initial.shape[0]), hdyn_number_of_cpu_processes, float(forgetting_offset.item()), int(gpu_tuning_factor), ) del spikes # ########################################################### # Save the necessary data for the backward pass # ########################################################### ctx.save_for_backward( input, weights, output, parameter_list, # grad_output_scale, ) return output @staticmethod def backward(ctx, grad_output): # ############################################## # Get the variables back # ############################################## ( input, weights, output, parameter_list, # last_grad_scale, ) = ctx.saved_tensors # ############################################## # Default output # ############################################## grad_input = None grad_eps_xy = None grad_epsilon_t_0 = None grad_weights = None grad_h_initial = None grad_parameter_list = None grad_forgetting_offset = None # ############################################## # Parameters # ############################################## parameter_w_trainable: bool = bool(parameter_list[0]) # parameter_disable_scale_grade: bool = bool(parameter_list[1]) # parameter_keep_last_grad_scale: bool = bool(parameter_list[2]) parameter_skip_gradient_calculation: bool = bool(parameter_list[3]) parameter_output_layer: bool = bool(parameter_list[9]) parameter_local_learning: bool = bool(parameter_list[10]) # ############################################## # Dealing with overall scale of the gradient # ############################################## # if parameter_disable_scale_grade is False: # if parameter_keep_last_grad_scale is True: # last_grad_scale = torch.tensor( # [torch.abs(grad_output).max(), last_grad_scale] # ).max() # grad_output /= last_grad_scale # grad_output_scale = last_grad_scale.clone() input /= input.sum(dim=1, keepdim=True, dtype=weights.dtype) # ################################################# # User doesn't want us to calculate the gradients # ################################################# if parameter_skip_gradient_calculation is True: return ( grad_input, grad_eps_xy, grad_epsilon_t_0, grad_weights, grad_h_initial, grad_parameter_list, # grad_output_scale, grad_forgetting_offset, ) # ################################################# # Calculate backprop error (grad_input) # ################################################# backprop_r: torch.Tensor = weights.unsqueeze(0).unsqueeze(-1).unsqueeze( -1 ) * output.unsqueeze(1) backprop_bigr: torch.Tensor = backprop_r.sum(dim=2) backprop_z: torch.Tensor = backprop_r * ( 1.0 / (backprop_bigr + 1e-20) ).unsqueeze(2) grad_input: torch.Tensor = (backprop_z * grad_output.unsqueeze(1)).sum(2) del backprop_z # ################################################# # Calculate weight gradient (grad_weights) # ################################################# if parameter_w_trainable is False: # ################################################# # We don't train this weight # ################################################# grad_weights = None elif (parameter_output_layer is False) and (parameter_local_learning is True): # ################################################# # Local learning # ################################################# grad_weights = ( (-2 * (input - backprop_bigr).unsqueeze(2) * output.unsqueeze(1)) .sum(0) .sum(-1) .sum(-1) ) else: # ################################################# # Backprop # ################################################# backprop_f: torch.Tensor = output.unsqueeze(1) * ( input / (backprop_bigr**2 + 1e-20) ).unsqueeze(2) result_omega: torch.Tensor = backprop_bigr.unsqueeze( 2 ) * grad_output.unsqueeze(1) result_omega -= (backprop_r * grad_output.unsqueeze(1)).sum(2).unsqueeze(2) result_omega *= backprop_f del backprop_f grad_weights = result_omega.sum(0).sum(-1).sum(-1) del result_omega del backprop_bigr del backprop_r return ( grad_input, grad_eps_xy, grad_epsilon_t_0, grad_weights, grad_h_initial, grad_parameter_list, # grad_output_scale, grad_forgetting_offset, )