Hello,
I use two different custom functions to load the network parameters. I load the network into two different variables ‘network1’ and ‘network2’.
When I compare them as network1 == network2, the result is False even though state-dictionaries seem to coincide, although I havent checked every single entry. Is this a supposed behaviour?
These are the function for saving and loading the network state dictionaries (one pair of function is save_network() and load_network(), and another pair of functions is save_network_params() and load_network_params()).
def save_network(self, model, network_parameters, input_depth, conv_layer_activations, linear_layer_activations, comment):
checkpoint = dict(network_parameters) # or orig.copy()
extra = {"input_depth": input_depth,
"conv_activation_functions": conv_layer_activations,
"fc_activation_functions": linear_layer_activations,
"state_dict": model.state_dict()}
checkpoint.update(extra)
print()
print()
print(checkpoint["state_dict"])
torch.save(checkpoint, os.path.join(self.path_params,"_r" + str(self.run_count) + "_e" + str(self.epoch_count) + ".pth"))
return self.subfolder
def load_network(self,date_,run,epoch):
print("Available subfolders:")
print(np.transpose(next(os.walk(self.params_root))[1]))
print()
run_epoch,filenames = self.retrieve_run_epoch_from_filenames(date_)
print("Files in " + "[" + date_ + "]")
print(filenames)
# find the index of the corresponding filename in filenames-list
for row, sublist in enumerate(run_epoch):
if (sublist[0] == run) & (sublist[1] == epoch):
idx_ = row
filename = filenames[idx_]
folder = os.path.join(self.params_root,date_)
checkpoint = torch.load(os.path.join(folder,filename))
model = Network3(checkpoint["input_depth"],
checkpoint["conv_layer_out_channels"],
checkpoint["conv_layer_kernel_sizes"],
checkpoint["conv_layer_strides"],
checkpoint["conv_layer_paddings"],
checkpoint["max_pool_layer_numbers"],
checkpoint["max_pool_kernel_sizes"],
checkpoint["max_pool_strides"],
checkpoint["max_pool_paddings"],
checkpoint["batch_norm_2d_layer_numbers"],
checkpoint["dropout2d_layer_numbers"],
checkpoint["dropout2d_probabilities"],
checkpoint["linear_layer_out_features"],
checkpoint["batch_norm_1d_layer_numbers"],
checkpoint["conv_activation_functions"],
checkpoint["fc_activation_functions"],
checkpoint["dropout1d_layer_numbers"],
checkpoint["dropout1d_probabilities"]
)
model.load_state_dict(checkpoint['state_dict'])
print("\nLoaded: ")
print(filename)
# model.eval()
return model
def save_network_params(self,model,comment):
print()
print()
print(model.state_dict())
torch.save(model.state_dict(), os.path.join(self.path_params,"_r" + str(self.run_count) + "_e" + str(self.epoch_count) + ".pt"))
# torch.save(model.state_dict(), os.path.join(self.path_params,comment + "_r" + str(self.run_count) + "_e" + str(self.epoch_count) + ".pt"))
return self.subfolder
def load_network_params(self,date_,model,run,epoch):
print("Available subfolders:")
print(np.transpose(next(os.walk(self.params_root))[1]))
print()
run_epoch,filenames = self.retrieve_run_epoch_from_filenames(date_)
print("Files in " + "[" + date_ + "]")
print(filenames)
# find the index of the corresponding filename in filenames-list
for row, sublist in enumerate(run_epoch):
if (sublist[0] == run) & (sublist[1] == epoch):
idx_ = row
filename = filenames[idx_]
folder = os.path.join(self.params_root,date_)
model.load_state_dict(torch.load(os.path.join(folder,filename)))
print("\nLoaded: ")
print(filename)
model.eval()
Then, I use them as follows:
network3 = Network3(input_depth, conv_layer_out_channels,
conv_layer_kernel_sizes,
conv_layer_strides,
conv_layer_paddings,
max_pool_layer_numbers,
max_pool_kernel_sizes,
max_pool_strides,
max_pool_paddings,
batch_norm_2d_layer_numbers,
dropout2d_layer_numbers,
dropout2d_probabilities,
linear_layer_out_features,
batch_norm_1d_layer_numbers,
conv_layer_activations,
linear_layer_activations,
dropout1d_layer_numbers,
dropout1d_probabilities
)
network3.eval()
network_parameters = {"conv_layer_out_channels": conv_layer_out_channels,
"conv_layer_kernel_sizes": conv_layer_kernel_sizes,
"conv_layer_strides": conv_layer_strides,
"conv_layer_paddings": conv_layer_paddings,
"max_pool_layer_numbers": max_pool_layer_numbers,
"max_pool_kernel_sizes": max_pool_kernel_sizes,
"max_pool_strides": max_pool_strides,
"max_pool_paddings": max_pool_paddings,
"batch_norm_2d_layer_numbers": batch_norm_2d_layer_numbers,
"dropout2d_layer_numbers": dropout2d_layer_numbers,
"dropout2d_probabilities": dropout2d_probabilities,
"linear_layer_out_features": linear_layer_out_features,
"batch_norm_1d_layer_numbers": batch_norm_1d_layer_numbers,
"dropout1d_layer_numbers": dropout1d_layer_numbers,
"dropout1d_probabilities": dropout1d_probabilities
}
mmm1 = RunManager('test', 'test')
dat = mmm1.save_network_params(network3, 'comment')
network1 = network3
mmm1.load_network_params_2( date_=dat, model=network1, run=0, epoch=0)
mmm2 = RunManager('test', 'test')
dat = mmm2.save_network(network3, network_parameters, input_depth, conv_layer_activations, linear_layer_activations, 'test_save')
network2 = mmm2.load_network(dat, 0, 0)
network2.eval()
Then I print comparisons:
print(network3 == network_loaded2)
print(network1 == network_loaded2)
print(network3 == network_loaded1)
And the results are
False
False
True
Why do I have the first two False?
@ptrblck Would you have any ideas/suggestions? Can it be the case the the two networks are located at different addresses and this makes them not equal or something like that?