def get_patch_dimensions(sample_data, patch_size=16, overlap=2):
"""
Gets the dimensions of the patched inputs. This is used for the initialization
of GreedyInfoMaxBlocks.
"""
x, n_patches_x, n_patches_y = patchify_inputs(sample_data, patch_size,
overlap)
return n_patches_x, n_patches_y
def forward(self, x):
# Patchify inputs
x, n_patches_x, n_patches_y = data_utils.patchify_inputs(
x, self.patch_size, self.overlap)
x = self.encoder[0](x)
# Save positive/contrastive samples for each encoder block
log_f_module_list, true_f_module_list, z = self.encoder[1](x,
n_patches_x,
n_patches_y)
# Lists of lists: each list has num_modules internal lists, with each
# internal list containing k_predictions elements
return log_f_module_list, true_f_module_list
def encode(self, x):
# Patchify inputs
x, n_patches_x, n_patches_y = data_utils.patchify_inputs(
x, self.patch_size, self.overlap)
x = self.encoder[0](x)
# Compute encoded patch-level representation for each encoder block
for module in self.encoder[1:]:
# no need to detach between modules as .encode() will only be called
# under a torch.no_grad() scope
x, out = module.encode(x, n_patches_x, n_patches_y)
# Return patch-level representation from the last block
return out
def forward(self, x):
# Patchify inputs
x, n_patches_x, n_patches_y = data_utils.patchify_inputs(
x, self.patch_size, self.overlap)
x = self.encoder[0](x)
# Save positive/contrastive samples for each encoder block
log_f_module_list = []
for module in self.encoder[1:]:
# log_f_list and true_f_list each have k_predictions elements
log_f_list, z = module(x, n_patches_x, n_patches_y)
log_f_module_list.append(log_f_list)
# Detach x to make sure no gradients are flowing in between modules
x = z.detach()
# Lists of lists: each list has num_modules internal lists, with each
# internal list containing k_predictions elements
return log_f_module_list
def forward(self, x):
x, n_patches_x, n_patches_y = data_utils.patchify_inputs(
x, self.patch_size, self.overlap)
return x