def fine_tune_model():
print(f'\n\nfine tuning model...')
data_manager = DataManager()
plotter = Plotter()
train_manager = get_train_manager(data_manager)
n_filters_list = []
train_losses, test_losses = [], []
for i in range(1, 11):
n_filters = 4 * i
net = Net(conv1_out_channels=n_filters)
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
train_manager.init_model(net, optimizer)
train_and_save_model(train_manager)
(train_loss,
train_accuracy), (test_loss,
test_accuracy) = train_manager.get_losses()
n_filters_list.append(n_filters)
train_losses.append(train_loss)
test_losses.append(test_loss)
plotter.plot_filters_losses(n_filters_list, train_losses, test_losses)
return train_losses, test_losses, n_filters_list
def test_no_spatial_structure():
print(f'testing no spatial structure...')
fixed_shuffle_data_manager = DataManager(shuffle_type='fixed')
fresh_shuffle_data_manager = DataManager(shuffle_type='fresh')
for name, data_manager in [('fixed_shuffle', fixed_shuffle_data_manager),
('fresh_shuffle', fresh_shuffle_data_manager)]:
print(f'\ntesting {name}...')
train_manager = get_train_manager(data_manager)
net = Net()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
train_manager.init_model(net, optimizer)
train_manager.train()
print(f'\ntesting no spatial...')
train_res, test_res = train_manager.get_losses()
print(f'train loss: {train_res[0]}, train accuracy: {train_res[1]}\n'
f'test loss: {test_res[0]}, test accuracy: {test_res[1]}')
def test_locality_of_receptive_field():
print(f'testing locality of receptive field...')
noshuffle_data_manager = DataManager(shuffle_type='none')
shuffled_data_manager = DataManager(shuffle_type='fixed')
for name, data_manager in [('no_shuffle', noshuffle_data_manager),
('shuffle', shuffled_data_manager)]:
print(f'\ntesting {name}...')
train_manager = get_train_manager(data_manager)
net = Net()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
train_manager.init_model(net, optimizer)
train_manager.train()
print(f'\ntesting local field...')
train_res, test_res = train_manager.get_losses()
print(f'train loss: {train_res[0]}, train accuracy: {train_res[1]}\n'
f'test loss: {test_res[0]}, test accuracy: {test_res[1]}')
# # print(trainset[i])
# print(trainset[i]["sensitive"])
# print("Testset:")
# for i in range(len(testset)):
# print("testset " + str(i))
# print(testset[i]["sensitive"])
#
writer = SummaryWriter("runs/noConv1D_ascad_desync_50_3")
#TODO: Change the model for the one of the paper
net = Net()
#TODO: propose NLLloss (Categorical Cross Entropy), Adam optimizer and Cyclic Learning Rate
criterion = nn.NLLLoss()
optimizer = optim.Adam(net.parameters(), lr=float(config.train.lr))
scheduler = optim.lr_scheduler.OneCycleLR(optimizer,
max_lr=float(config.train.lr),
epochs=50,
steps_per_epoch=len(trainloader))
# TODO: plot in tensorboard the curves loss and accuracy for train and val
for epoch in range(
config.train.epochs): # loop over the dataset multiple times
print('Epoch {}/{}'.format(epoch + 1, config.train.epochs))
print('-' * 10)
for phase in ["train", "val"]:
if phase == "train":
net.train()
else:
net.eval()
noise_type=noise_type,
noise_level=noise_level,
init_val=init_val).to(device, dtype)
save_model(net,
experiment_name,
0,
noise_type,
noise_level,
model_dir=model_dir)
else:
print("starting from epoch {}".format(start_epoch))
net = recreate_model(model_to_load, dataset=dataset, act=act)
# optimiser parameters
optimiser = get_optimiser(net.parameters(), op, learning_rate,
momentum)
# training criterion
criterion = torch.nn.CrossEntropyLoss()
# train network
train(net,
train_loader,
test_loader,
criterion,
optimiser,
epochs,
noise_type,
noise_level,
save=True,