encoder = Encoder()
# encoder = Encoder()
encoder.to(device)
encoder.eval()
for i, orient in enumerate(orients):
print("Processing {} features".format(orient))
data_path = os.path.join(experiment_dir, 'tractseg_data', subj_id,
'training_slices', orient)
features_save_path = os.path.join(experiment_dir, 'tractseg_data', subj_id,
'learned_features')
dataset = Datasets.OrientationDataset(data_path,
scale=True,
normalize=False,
bg_zero=True,
noise_prob=noise_prob,
alpha=1)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=10)
encoder_path = "{}saved_models/{}_encoder_epoch_{}".format(
experiment_dir, model_name, epoch)
encoder.load_state_dict(torch.load(encoder_path))
print("Loaded pretrained weights starting from epoch {} for {}".format(
epoch, model_name))
with torch.no_grad():
if bool(noise_prob):
noise_prob = None
num_latent = 128
model = ConvVAE(num_latent, device)
model.to(device)
model.eval()
for i, orient in enumerate(orients):
print("Processing {} features".format(orient))
data_path = os.path.join(experiment_dir, 'tractseg_data', subj_id,
'training_slices', orient)
features_save_path = os.path.join(experiment_dir, 'tractseg_data', subj_id,
'learned_features')
dataset = Datasets.OrientationDataset(data_path,
scale=True,
normalize=False)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=10)
model_path = "{}saved_models/{}_epoch_{}".format(experiment_dir,
model_name, epoch)
model.load_state_dict(torch.load(model_path))
print("Loaded pretrained weights starting from epoch {} for {}".format(
epoch, model_name))
with torch.no_grad():
orient_features = torch.zeros(feature_shapes[i])
model_name = 'MODEL1_DENOISING_P50'
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") # device
deterministic = True # reproducibility
seed = 0 # random seed for reproducibility
if deterministic:
torch.manual_seed(seed)
torch.backends.cudnn.benchmark = (not deterministic) # set False whenever input size varies
torch.backends.cudnn.deterministic = deterministic
batch_size = 2 ** 15
start_epoch = 10
channels = 10
noise_prob = None
trainset = Datasets.VoxelDataset(data_path,
file_name='data.nii.gz',
normalize=False,
scale=True)
total_examples = len(trainset)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=False, num_workers=6)
print("Total training examples: {}, Batch size: {}, Iters per epoch: {}".format(total_examples,
batch_size,
len(trainloader)))
# ------------------------------------------Model-----------------------------------------------------------------------
# model settings
encoder = Encoder(h=10)
decoder = Decoder(h=10)
# encoder = Encoder(288, 10, 10, device)
# decoder = Decoder(288, 10, 10)
encoder.to(device)
decoder.to(device)
batch_size = 1
start_epoch = 0 # for loading pretrained weights
num_epochs = 10 # number of epochs to trains
checkpoint = 10 # save model every checkpoint epoch
masked_loss = False
denoising = False
# ------------------------------------------Data------------------------------------------------------------------------
# trainset = Datasets.VoxelDataset(data_path,
# file_name='data.nii.gz',
# normalize=False,
# scale=True)
# trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=10)
trainset = Datasets.OrientationDataset(data_path, scale=True, normalize=False, bg_zero=True)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=10)
total_examples = len(trainset)
print("Total training examples: {}, Batch size: {}, Iters per epoch: {}".format(total_examples,
batch_size,
len(trainloader)))
# ------------------------------------------Model-----------------------------------------------------------------------
for h in range(1, 101):
model_name = "ConvModel1_prelu_h{}".format(h)
# model settings
# encoder = Encoder(h=h)
# decoder = Decoder(h=h)
encoder = Encoder(input_size=(145, 145), h=h)
decoder = Decoder(h)
encoder.to(device)