def load_step_sfcn_preloaded(run,
task,
bin_min,
bin_max,
reinit_with_scaling=None):
model = give_pretrained_sfcn(run, task)
model.module.change_output_dim(bin_max - bin_min)
if reinit_with_scaling is not None:
model.module.reinit_final_layers_pres_scale(reinit_with_scaling)
return model
def load_final_sfcn_preloaded(run,
task,
bin_min,
bin_max,
reinit_with_scaling=None):
model = give_pretrained_sfcn(run, task)
new_output_dim = bin_max - bin_min
old_output_dim = model.module.output_dim
if new_output_dim != old_output_dim:
print("Output dimension is changed. Reshape final layer.")
model.module.change_output_dim(new_output_dim)
else:
print("Output dimensions match. No reshaping of final layer.")
if reinit_with_scaling is not None:
model.module.reinit_final_layers_pres_scale(reinit_with_scaling)
return model
def test_sfcn_preloaded(test_loader,hps):
model,bin_min,bin_max=give_pretrained_sfcn(hps['run'],hps['task'],give_range=True)
model.module.train_nothing()
info_start="Full model is being tested with weights loaded from pre-trained model."
info_end="Full model was being tested with weights loaded from pre-trained model."
return sfcn_test(model,
test_loader=test_loader,
bin_min=bin_min,
bin_max=bin_max,
space=hps['space'],
loss_met=hps['loss_met'],
eval_met=hps['eval_met'],
bin_step=hps['bin_step'],
sigma=hps['sigma'],
n_epochs=hps['n_epochs'],
print_corr=hps['print_corr'],
info_start=info_start,
info_end=info_end)
x = Variable(x, requires_grad=True)
model.module.train_nothing()
model.eval()
loss_list = []
for it in range(n_epochs):
loss = compute_activation(model, x, filter_index)
loss.backward()
x.data += lr * x.grad.data - alpha * x.data
x.grad.zero_()
if it % print_every == 0:
print("Epoch: %3d || Activ: %.6f" % (it, loss))
loss_list.append(loss)
return (x, loss_list)
model = give_pretrained_sfcn("0", "age")
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Number of GPUs: ", torch.cuda.device_count())
model = model.to(device)
_, train_loader = give_oasis_data('train',
batch_size=1,
num_workers=4,
shuffle=True,
debug='debug',
preprocessing='min',
task='age',
share=1.)
n_it = 100
lr_list = [0.1, 1., 10., 100., 0.1]
num_workers=4,
shuffle=True,
debug=False,
preprocessing='min',
task=label,
share=1.0,
balance=False)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Number of GPUs: ", torch.cuda.device_count())
for method in methods:
for run in runs:
for pretrained_task in pretrained_tasks:
model = give_pretrained_sfcn(run, pretrained_task)
model.eval()
model.module.train_nothing()
model = model.to(device)
if ARGS['DEBUG'] == 'debug':
n_samples = 10
elif ARGS['DEBUG'] == 'full':
n_samples = min(train_dataset._len, ARGS['N_SAMPLES'])
else:
sys.exit("Unknown debug flag. Either 'debug' or 'full'.")
#Get random subset of indices of the dataset:
size_dataset = train_dataset._len
inds = torch.randperm(size_dataset)[:n_samples].tolist()
n_final_channels = model.module.channel_number[-1]