def test_init_pdnet():
run_params = {
'n_primal': 5,
'n_dual': 5,
'n_iter': 10,
'n_filters': 32,
}
model = pdnet(lr=1e-3, **run_params)
def fine_tune_pdnet(original_run_id, af, contrast, cuda_visible_devices):
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(cuda_visible_devices)
af = int(af)
# generators
train_set = train_masked_kspace_dataset_from_indexable(
train_path,
AF=af,
contrast=contrast,
inner_slices=8,
rand=True,
scale_factor=1e6,
)
val_set = train_masked_kspace_dataset_from_indexable(
val_path,
AF=af,
contrast=contrast,
scale_factor=1e6,
)
run_params = {
'n_primal': 5,
'n_dual': 5,
'n_iter': 10,
'n_filters': 32,
}
n_epochs = 50
new_run_id = f'pdnet_af{af}_{contrast}_{int(time.time())}'
chkpt_path = f'{CHECKPOINTS_DIR}checkpoints/{new_run_id}' + '-{epoch:02d}.hdf5'
chkpt_cback = ModelCheckpoint(chkpt_path,
period=n_epochs,
save_weights_only=True)
log_dir = op.join(f'{LOGS_DIR}logs', new_run_id)
tboard_cback = TensorBoard(
profile_batch=0,
log_dir=log_dir,
histogram_freq=0,
write_graph=False,
write_images=False,
)
model = pdnet(lr=1e-6, **run_params)
model.load_weights(
f'{CHECKPOINTS_DIR}checkpoints/{original_run_id}-300.hdf5')
print(model.summary(line_length=150))
model.fit(
train_set,
steps_per_epoch=n_volumes_train // 2,
epochs=n_epochs,
validation_data=val_set,
validation_steps=5,
verbose=0,
callbacks=[
tboard_cback,
chkpt_cback,
],
)
run_id = f'pdnet_af{AF}_oasis_{int(time.time())}'
chkpt_path = f'checkpoints/{run_id}' + '-{epoch:02d}.hdf5'
chkpt_cback = ModelCheckpoint(chkpt_path, period=100, save_weights_only=True)
log_dir = op.join('logs', run_id)
tboard_cback = TensorBoard(
profile_batch=0,
log_dir=log_dir,
histogram_freq=0,
write_graph=True,
write_images=False,
)
tqdm_cb = TQDMCallback(metric_format="{name}: {value:e}")
tqdm_cb.on_train_batch_begin = tqdm_cb.on_batch_begin
tqdm_cb.on_train_batch_end = tqdm_cb.on_batch_end
model = pdnet(input_size=(None, None, 1), fastmri=False, lr=1e-3, **run_params)
print(model.summary(line_length=150))
model.fit_generator(
train_gen,
steps_per_epoch=n_train,
epochs=n_epochs,
validation_data=val_gen,
validation_steps=1,
verbose=0,
callbacks=[tqdm_cb, tboard_cback, chkpt_cback,],
# max_queue_size=35,
use_multiprocessing=True,
workers=35,
shuffle=False,
def train_pdnet(subclassed, af, contrast, cuda_visible_devices, n_samples, n_epochs, n_iter):
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(cuda_visible_devices)
af = int(af)
# generators
train_set = train_masked_kspace_dataset_from_indexable(
train_path,
AF=af,
contrast=contrast,
inner_slices=8,
rand=True,
scale_factor=1e6,
n_samples=n_samples,
)
val_set = train_masked_kspace_dataset_from_indexable(
val_path,
AF=af,
contrast=contrast,
scale_factor=1e6,
)
run_params = {
'n_primal': 5,
'n_dual': 5,
'n_iter': n_iter,
'n_filters': 32,
}
additional_info = f'af{af}'
if contrast is not None:
additional_info += f'_{contrast}'
if n_samples is not None:
additional_info += f'_{n_samples}'
if n_iter != 10:
additional_info += f'_i{n_iter}'
run_id = f'pdnet_{additional_info}_{int(time.time())}'
chkpt_path = f'{CHECKPOINTS_DIR}checkpoints/{run_id}' + '-{epoch:02d}.hdf5'
chkpt_cback = ModelCheckpoint(chkpt_path, period=n_epochs, save_weights_only=True)
log_dir = op.join(f'{LOGS_DIR}logs', run_id)
tboard_cback = TensorBoard(
profile_batch=0,
log_dir=log_dir,
histogram_freq=0,
write_graph=False,
write_images=False,
)
if subclassed:
model = PDNet(**run_params)
default_model_compile(model, lr=1e-3)
else:
def adapt_dataset_tensors_to_functional_model(model_input, model_output):
mask = model_input[1]
new_mask = tf.tile(mask, [1, 640, 1])
return (model_input[0], new_mask), model_output
train_set = train_set.map(adapt_dataset_tensors_to_functional_model, num_parallel_calls=tf.data.experimental.AUTOTUNE)
val_set = val_set.map(adapt_dataset_tensors_to_functional_model, num_parallel_calls=tf.data.experimental.AUTOTUNE)
model = pdnet(lr=1e-3, **run_params)
print(run_id)
model.fit(
train_set,
steps_per_epoch=n_volumes_train,
epochs=n_epochs,
validation_data=val_set,
validation_steps=5,
verbose=0,
callbacks=[tboard_cback, chkpt_cback,],
)
def train_pdnet(af, contrast, cuda_visible_devices, n_samples, n_epochs):
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(cuda_visible_devices)
af = int(af)
# generators
train_set = train_masked_kspace_dataset_from_indexable(
train_path,
AF=af,
contrast=contrast,
inner_slices=8,
rand=True,
scale_factor=1e6,
n_samples=n_samples,
)
val_set = train_masked_kspace_dataset_from_indexable(
val_path,
AF=af,
contrast=contrast,
inner_slices=8,
rand=True,
scale_factor=1e6,
)
run_params = {
'n_primal': 5,
'n_dual': 5,
'n_iter': 10,
'n_filters': 32,
'fastmri': False,
}
additional_info = f'af{af}'
if contrast is not None:
additional_info += f'_{contrast}'
if n_samples is not None:
additional_info += f'_{n_samples}'
run_id = f'pdnet_multicoil_{additional_info}_{int(time.time())}'
chkpt_path = f'{CHECKPOINTS_DIR}checkpoints/{run_id}' + '-{epoch:02d}.hdf5'
chkpt_cback = ModelCheckpoint(chkpt_path,
period=n_epochs,
save_weights_only=True)
log_dir = op.join(f'{LOGS_DIR}logs', run_id)
tboard_cback = TensorBoard(
profile_batch=0,
log_dir=log_dir,
histogram_freq=0,
write_graph=False,
write_images=False,
)
model = pdnet(lr=1e-3, **run_params)
print(run_id)
print(model.summary(line_length=150))
model.fit(
train_set,
steps_per_epoch=n_volumes_train,
epochs=n_epochs,
validation_data=val_set,
validation_steps=5,
verbose=0,
callbacks=[
tboard_cback,
chkpt_cback,
],
)