def test_gru_multistep_fprop_k3_multidataset():
hidden_sizes = 50
hyperparams = {
'model': 'gru_multistep',
'k': 3,
'm': 3,
'batch_size': 16,
'SGD': "1e-2",
'hidden_sizes': hidden_sizes,
'learn_to_stop': False,
'normalize': False,
'noisy_streamlines_sigma': None,
'shuffle_streamlines': True,
'seed': 1234}
with Timer("Creating dataset", newline=True):
volume_manager = neurotools.VolumeManager()
trainset = make_dummy_dataset(volume_manager)
validset = make_dummy_dataset(volume_manager)
print("Dataset sizes:", len(trainset))
batch_scheduler = factories.batch_scheduler_factory(hyperparams, trainset, train_mode=True)
print("An epoch will be composed of {} updates.".format(batch_scheduler.nb_updates_per_epoch))
print(volume_manager.data_dimension, hidden_sizes, batch_scheduler.target_size)
with Timer("Creating model"):
model = factories.model_factory(hyperparams, input_size=volume_manager.data_dimension, output_size=batch_scheduler.target_size,
volume_manager=volume_manager)
model.initialize(factories.weigths_initializer_factory("orthogonal", seed=1234))
# Test fprop
output = model.get_output(trainset.symb_inputs)
fct = theano.function([trainset.symb_inputs], output, updates=model.graph_updates)
batch_inputs, batch_targets, batch_mask = batch_scheduler._next_batch(2)
out = fct(batch_inputs)
with Timer("Building optimizer"):
loss = factories.loss_factory(hyperparams, model, trainset)
optimizer = factories.optimizer_factory(hyperparams, loss)
fct_loss = theano.function([trainset.symb_inputs, trainset.symb_targets, trainset.symb_mask], loss.loss, updates=model.graph_updates)
loss_value = fct_loss(batch_inputs, batch_targets, batch_mask)
print("Loss:", loss_value)
fct_optim = theano.function([trainset.symb_inputs, trainset.symb_targets, trainset.symb_mask], list(optimizer.directions.values()),
updates=model.graph_updates)
dirs = fct_optim(batch_inputs, batch_targets, batch_mask)
def compute_loss_errors(streamlines, model, hyperparams):
# Create dummy dataset for these new streamlines.
tracto_data = neurotools.TractographyData(None, None, None)
tracto_data.add(streamlines, bundle_name="Generated")
tracto_data.subject_id = 0
dataset = datasets.TractographyDataset([tracto_data], "Generated", keep_on_cpu=True)
# Override K for gru_multistep
if 'k' in hyperparams:
hyperparams['k'] = 1
batch_scheduler = batch_scheduler_factory(hyperparams, dataset, train_mode=False, batch_size_override=1000, use_data_augment=False)
loss = loss_factory(hyperparams, model, dataset)
loss_view = views.LossView(loss=loss, batch_scheduler=batch_scheduler)
return loss_view.losses.view()
def compute_loss_errors(streamlines, model, hyperparams):
# Create dummy dataset for these new streamlines.
tracto_data = neurotools.TractographyData(None, None, None)
tracto_data.add(streamlines, bundle_name="Generated")
tracto_data.subject_id = 0
dataset = datasets.TractographyDataset([tracto_data],
"Generated",
keep_on_cpu=True)
# Override K for gru_multistep
if 'k' in hyperparams:
hyperparams['k'] = 1
batch_scheduler = batch_scheduler_factory(hyperparams,
dataset,
train_mode=False,
batch_size_override=1000,
use_data_augment=False)
loss = loss_factory(hyperparams, model, dataset)
loss_view = views.LossView(loss=loss, batch_scheduler=batch_scheduler)
return loss_view.losses.view()
def test_gru_multistep_fprop_k3():
hidden_sizes = 50
hyperparams = {
'model': 'gru_multistep',
'k': 3,
'm': 3,
'batch_size': 16,
'SGD': "1e-2",
'hidden_sizes': hidden_sizes,
'learn_to_stop': False,
'normalize': False,
'noisy_streamlines_sigma': None,
'shuffle_streamlines': True,
'seed': 1234
}
with Timer("Creating dataset", newline=True):
volume_manager = neurotools.VolumeManager()
trainset = make_dummy_dataset(volume_manager)
print("Dataset sizes:", len(trainset))
batch_scheduler = factories.batch_scheduler_factory(hyperparams,
trainset,
train_mode=True)
print("An epoch will be composed of {} updates.".format(
batch_scheduler.nb_updates_per_epoch))
print(volume_manager.data_dimension, hidden_sizes,
batch_scheduler.target_size)
with Timer("Creating model"):
model = factories.model_factory(
hyperparams,
input_size=volume_manager.data_dimension,
output_size=batch_scheduler.target_size,
volume_manager=volume_manager)
model.initialize(
factories.weigths_initializer_factory("orthogonal", seed=1234))
# Test fprop
output = model.get_output(trainset.symb_inputs)
fct = theano.function([trainset.symb_inputs],
output,
updates=model.graph_updates)
batch_inputs, batch_targets, batch_mask = batch_scheduler._next_batch(2)
out = fct(batch_inputs)
with Timer("Building optimizer"):
loss = factories.loss_factory(hyperparams, model, trainset)
optimizer = factories.optimizer_factory(hyperparams, loss)
fct_loss = theano.function(
[trainset.symb_inputs, trainset.symb_targets, trainset.symb_mask],
loss.loss,
updates=model.graph_updates)
loss_value = fct_loss(batch_inputs, batch_targets, batch_mask)
print("Loss:", loss_value)
fct_optim = theano.function(
[trainset.symb_inputs, trainset.symb_targets, trainset.symb_mask],
list(optimizer.directions.values()),
updates=model.graph_updates)
dirs = fct_optim(batch_inputs, batch_targets, batch_mask)
def main():
parser = build_parser()
args = parser.parse_args()
print(args)
if min(args.keep_top) < 0:
parser.error("--keep-top must be between in [0, 1].")
# Get experiment folder
experiment_path = args.name
if not os.path.isdir(experiment_path):
# If not a directory, it must be the name of the experiment.
experiment_path = pjoin(".", "experiments", args.name)
if not os.path.isdir(experiment_path):
parser.error('Cannot find experiment: {0}!'.format(args.name))
# Load experiments hyperparameters
try:
hyperparams = smartutils.load_dict_from_json_file(
pjoin(experiment_path, "hyperparams.json"))
except FileNotFoundError:
hyperparams = smartutils.load_dict_from_json_file(
pjoin(experiment_path, "..", "hyperparams.json"))
# Use this for hyperparams added in a new version, but nonexistent from older versions
retrocompatibility_defaults = {
'feed_previous_direction': False,
'predict_offset': False,
'normalize': False,
'keep_step_size': False,
'sort_streamlines': False
}
for new_hyperparams, default_value in retrocompatibility_defaults.items():
if new_hyperparams not in hyperparams:
hyperparams[new_hyperparams] = default_value
with Timer("Loading signal data and tractogram", newline=True):
volume_manager = VolumeManager()
dataset = datasets.load_tractography_dataset_from_dwi_and_tractogram(
args.signal,
args.tractogram,
volume_manager,
use_sh_coeffs=hyperparams['use_sh_coeffs'],
bvals=args.bvals,
bvecs=args.bvecs,
step_size=args.step_size)
print("Dataset size:", len(dataset))
if vizu_available and args.vizu:
vizu.check_dataset_integrity(dataset, subset=0.2)
with Timer("Loading model"):
loss_type = args.loss_type
model = None
if hyperparams['model'] == 'gru_regression':
from learn2track.models import GRU_Regression
model = GRU_Regression.create(experiment_path,
volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_mixture':
from learn2track.models import GRU_Mixture
model = GRU_Mixture.create(experiment_path,
volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_multistep':
from learn2track.models import GRU_Multistep_Gaussian
model = GRU_Multistep_Gaussian.create(
experiment_path, volume_manager=volume_manager)
model.k = 1
model.m = 1
elif hyperparams['model'] == 'ffnn_regression':
from learn2track.models import FFNN_Regression
model = FFNN_Regression.create(experiment_path,
volume_manager=volume_manager)
if loss_type in ['l2_sum', 'l2_mean']:
loss_type = "expected_value"
else:
raise NameError("Unknown model: {}".format(hyperparams['model']))
with Timer("Building evaluation function"):
# Override K for gru_multistep
if 'k' in hyperparams:
hyperparams['k'] = 1
batch_scheduler = batch_scheduler_factory(
hyperparams,
dataset,
use_data_augment=
False, # Otherwise it doubles the number of losses :-/
train_mode=False,
batch_size_override=args.batch_size)
loss = loss_factory(hyperparams, model, dataset, loss_type=loss_type)
l2_error = views.LossView(loss=loss, batch_scheduler=batch_scheduler)
with Timer("Scoring...", newline=True):
dummy_status = Status() # Forces recomputing results
losses = l2_error.losses.view(dummy_status)
if hyperparams['model'] == 'ffnn_regression':
_losses = dataset.streamlines.copy()
_losses._data = losses.copy()
_losses._lengths -= 1
_losses._offsets -= np.arange(len(dataset.streamlines))
if args.loss_type == 'l2_sum':
losses = np.asarray([l.sum() for l in _losses])
elif args.loss_type == 'l2_mean':
losses = np.asarray([l.mean() for l in _losses])
mean = float(l2_error.mean.view(dummy_status))
stderror = float(l2_error.stderror.view(dummy_status))
print("Loss: {:.4f} ± {:.4f}".format(mean, stderror))
print("Min: {:.4f}".format(losses.min()))
print("Max: {:.4f}".format(losses.max()))
print("Percentiles: {}".format(
np.percentile(losses, [0, 25, 50, 75, 100])))
with Timer("Saving streamlines"):
nii = dataset.subjects[0].signal
tractogram = nib.streamlines.Tractogram(
dataset.streamlines[batch_scheduler.indices],
affine_to_rasmm=nii.affine)
tractogram.data_per_streamline['loss'] = losses
header = {}
header[Field.VOXEL_TO_RASMM] = nii.affine.copy()
header[Field.VOXEL_SIZES] = nii.header.get_zooms()[:3]
header[Field.DIMENSIONS] = nii.shape[:3]
header[Field.VOXEL_ORDER] = "".join(aff2axcodes(nii.affine))
nib.streamlines.save(tractogram.copy(), args.out, header=header)
if len(args.keep_top) > 0:
for keep_top in args.keep_top:
with Timer("Saving top {}% streamlines".format(keep_top)):
idx = np.argsort(losses)
idx = idx[:int(keep_top * len(losses))]
print("Keeping {}/{} streamlines".format(
len(idx), len(losses)))
sub_tractogram = tractogram[idx]
out_filename = args.out[:-4] + "_top{}".format(
keep_top) + ".tck"
nib.streamlines.save(sub_tractogram, out_filename)
def prediction_tractogram(hyperparams, model, dataset, batch_size_override,
prediction_method):
loss = loss_factory(hyperparams,
model,
dataset,
loss_type=prediction_method)
batch_scheduler = batch_scheduler_factory(
hyperparams,
dataset,
train_mode=False,
batch_size_override=batch_size_override,
use_data_augment=False)
_ = loss.losses # Hack to generate update dict in loss :(
predictions = loss.samples
predict, timestep_losses, inputs, targets, masks = log_variables(
batch_scheduler, model, predictions, loss.loss_per_time_step,
dataset.symb_inputs * 1, dataset.symb_targets * 1,
dataset.symb_mask * 1)
if hyperparams['model'] == 'ffnn_regression':
# Regrouping data into streamlines will only work if the original streamlines were NOT shuffled, resampled or augmented
timesteps_prediction = ArraySequence()
timesteps_loss = ArraySequence()
timesteps_inputs = ArraySequence()
timesteps_targets = ArraySequence()
idx = 0
for length in dataset.streamlines._lengths:
start = idx
idx = end = idx + length
timesteps_prediction.extend(predict[start:end])
timesteps_loss.extend(timestep_losses[start:end])
timesteps_inputs.extend(inputs[start:end])
timesteps_targets.extend(targets[start:end])
else:
timesteps_prediction = ArraySequence(
[p[:int(m.sum())] for p, m in zip(chain(*predict), chain(*masks))])
timesteps_loss = ArraySequence([
l[:int(m.sum())]
for l, m in zip(chain(*timestep_losses), chain(*masks))
])
timesteps_inputs = ArraySequence(
[i[:int(m.sum())] for i, m in zip(chain(*inputs), chain(*masks))])
# Use np.squeeze in case gru_multistep is used to remove the empty k=1 dimension
timesteps_targets = ArraySequence([
np.squeeze(t[:int(m.sum())])
for t, m in zip(chain(*targets), chain(*masks))
])
# Debug : Print norm stats
# print("Dataset: {}; # of streamlines: {}".format(dataset.name, len(dataset)))
# all_predictions = np.array(list(chain(*timesteps_prediction)))
# prediction_norms = np.linalg.norm(all_predictions, axis=1)
# print("Prediction norm --- Mean:{}; Max:{}; Min:{}".format(np.mean(prediction_norms), np.max(prediction_norms), np.min(prediction_norms)))
# all_targets = np.array(list(chain(*timesteps_targets)))
# target_norms = np.linalg.norm(all_targets, axis=1)
# print("Target norm --- Mean:{}; Max:{}; Min:{}".format(np.mean(target_norms), np.max(target_norms), np.min(target_norms)))
# Color is based on timestep loss
cmap = cm.get_cmap('bwr')
values = np.concatenate(timesteps_loss)
vmin = np.percentile(values, 5)
vmax = np.percentile(values, 95)
scalar_map = cm.ScalarMappable(norm=mplcolors.Normalize(vmin=vmin,
vmax=vmax),
cmap=cmap)
streamlines = []
colors = []
for i, t, p, l in zip(timesteps_inputs, timesteps_targets,
timesteps_prediction, timesteps_loss):
pts = np.r_[i[:, :3], [i[-1, :3] + t[-1]]]
streamline = np.zeros(((len(pts) - 1) * 3 + 1, 3))
streamline[::3] = pts
streamline[1:-1:3] = pts[:-1] + p
streamline[2:-1:3] = pts[:-1]
streamlines.append(streamline)
# Color input streamlines in a uniform color, then color predictions based on L2 error
color = np.zeros_like(streamline)
# Base color of streamlines is minimum value (best score)
color[:] = scalar_map.to_rgba(vmin, bytes=True)[:3]
color[1:-1:3, :] = scalar_map.to_rgba(l, bytes=True)[:, :3]
colors.append(color)
tractogram = nib.streamlines.Tractogram(streamlines,
data_per_point={"colors": colors})
return tractogram
def evaluation_tractogram(hyperparams, model, dataset, batch_size_override,
metric):
loss = loss_factory(hyperparams, model, dataset, loss_type=None)
batch_scheduler = batch_scheduler_factory(
hyperparams,
dataset,
train_mode=False,
batch_size_override=batch_size_override,
use_data_augment=False)
_ = loss.losses # Hack to generate update dict in loss :(
if hyperparams['model'] == 'ffnn_regression':
timestep_losses, inputs, targets = log_variables(
batch_scheduler, model, loss.loss_per_time_step,
dataset.symb_inputs * 1, dataset.symb_targets * 1)
# Regrouping data into streamlines will only work if the original streamlines were NOT shuffled, resampled or augmented
timesteps_loss = ArraySequence()
seq_loss = []
timesteps_inputs = ArraySequence()
timesteps_targets = ArraySequence()
idx = 0
for length in dataset.streamlines._lengths:
start = idx
idx = end = idx + length
timesteps_loss.extend(timestep_losses[start:end])
seq_loss.extend(np.mean(timestep_losses[start:end]))
timesteps_inputs.extend(inputs[start:end])
timesteps_targets.extend(targets[start:end])
else:
timestep_losses, seq_losses, inputs, targets, masks = log_variables(
batch_scheduler, model, loss.loss_per_time_step, loss.loss_per_seq,
dataset.symb_inputs * 1, dataset.symb_targets * 1,
dataset.symb_mask * 1)
timesteps_loss = ArraySequence([
l[:int(m.sum())]
for l, m in zip(chain(*timestep_losses), chain(*masks))
])
seq_loss = np.array(list(chain(*seq_losses)))
timesteps_inputs = ArraySequence(
[i[:int(m.sum())] for i, m in zip(chain(*inputs), chain(*masks))])
# Use np.squeeze in case gru_multistep is used to remove the empty k=1 dimension
timesteps_targets = ArraySequence([
np.squeeze(t[:int(m.sum())])
for t, m in zip(chain(*targets), chain(*masks))
])
if metric == 'sequence':
# Color is based on sequence loss
values = seq_loss
elif metric == 'timestep' or metric == 'cumul_avg':
# Color is based on timestep loss
values = np.concatenate(timesteps_loss)
else:
raise ValueError("Unrecognized metric: {}".format(metric))
cmap = cm.get_cmap('bwr')
vmin = np.percentile(values, 5)
vmax = np.percentile(values, 95)
scalar_map = cm.ScalarMappable(norm=mplcolors.Normalize(vmin=vmin,
vmax=vmax),
cmap=cmap)
streamlines = []
colors = []
for i, t, l, seq_l in zip(timesteps_inputs, timesteps_targets,
timesteps_loss, seq_loss):
pts = np.r_[i[:, :3], [i[-1, :3] + t[-1]]]
color = np.zeros_like(pts)
if metric == 'sequence':
# Streamline color is based on sequence loss
color[:, :] = scalar_map.to_rgba(seq_l, bytes=True)[:3]
elif metric == 'timestep':
# Streamline color is based on timestep loss
# Identify first point with green
color[0, :] = [0, 255, 0]
color[1:, :] = scalar_map.to_rgba(l, bytes=True)[:, :3]
elif metric == 'cumul_avg':
# Streamline color is based on timestep loss
# Compute cumulative average
cumul_avg = np.cumsum(l) / np.arange(1, len(l) + 1)
# Identify first point with green
color[0, :] = [0, 255, 0]
color[1:, :] = scalar_map.to_rgba(cumul_avg, bytes=True)[:, :3]
else:
raise ValueError("Unrecognized metric: {}".format(metric))
streamlines.append(pts)
colors.append(color)
tractogram = nib.streamlines.Tractogram(streamlines,
data_per_point={"colors": colors})
return tractogram
def main():
parser = build_argparser()
args = parser.parse_args()
print(args)
print("Using Theano v.{}".format(theano.version.short_version))
hyperparams_to_exclude = ['max_epoch', 'force', 'name', 'view', 'shuffle_streamlines']
# Use this for hyperparams added in a new version, but nonexistent from older versions
retrocompatibility_defaults = {'feed_previous_direction': False,
'normalize': False}
experiment_path, hyperparams, resuming = utils.maybe_create_experiment_folder(args, exclude=hyperparams_to_exclude,
retrocompatibility_defaults=retrocompatibility_defaults)
# Log the command currently running.
with open(pjoin(experiment_path, 'cmd.txt'), 'a') as f:
f.write(" ".join(sys.argv) + "\n")
print("Resuming:" if resuming else "Creating:", experiment_path)
with Timer("Loading dataset", newline=True):
trainset_volume_manager = VolumeManager()
validset_volume_manager = VolumeManager()
trainset = datasets.load_tractography_dataset(args.train_subjects, trainset_volume_manager, name="trainset", use_sh_coeffs=args.use_sh_coeffs)
validset = datasets.load_tractography_dataset(args.valid_subjects, validset_volume_manager, name="validset", use_sh_coeffs=args.use_sh_coeffs)
print("Dataset sizes:", len(trainset), " |", len(validset))
if args.view:
tsne_view(trainset, trainset_volume_manager)
sys.exit(0)
batch_scheduler = batch_scheduler_factory(hyperparams, dataset=trainset, train_mode=True)
print("An epoch will be composed of {} updates.".format(batch_scheduler.nb_updates_per_epoch))
print(trainset_volume_manager.data_dimension, args.hidden_sizes, batch_scheduler.target_size)
with Timer("Creating model"):
input_size = trainset_volume_manager.data_dimension
if hyperparams['feed_previous_direction']:
input_size += 3
model = model_factory(hyperparams,
input_size=input_size,
output_size=batch_scheduler.target_size,
volume_manager=trainset_volume_manager)
model.initialize(weigths_initializer_factory(args.weights_initialization,
seed=args.initialization_seed))
with Timer("Building optimizer"):
loss = loss_factory(hyperparams, model, trainset)
if args.clip_gradient is not None:
loss.append_gradient_modifier(DirectionClipping(threshold=args.clip_gradient))
optimizer = optimizer_factory(hyperparams, loss)
with Timer("Building trainer"):
trainer = Trainer(optimizer, batch_scheduler)
# Log training error
loss_monitor = views.MonitorVariable(loss.loss)
avg_loss = tasks.AveragePerEpoch(loss_monitor)
trainer.append_task(avg_loss)
# Print average training loss.
trainer.append_task(tasks.Print("Avg. training loss: : {}", avg_loss))
# if args.learn_to_stop:
# l2err_monitor = views.MonitorVariable(T.mean(loss.mean_sqr_error))
# avg_l2err = tasks.AveragePerEpoch(l2err_monitor)
# trainer.append_task(avg_l2err)
#
# crossentropy_monitor = views.MonitorVariable(T.mean(loss.cross_entropy))
# avg_crossentropy = tasks.AveragePerEpoch(crossentropy_monitor)
# trainer.append_task(avg_crossentropy)
#
# trainer.append_task(tasks.Print("Avg. training L2 err: : {}", avg_l2err))
# trainer.append_task(tasks.Print("Avg. training stopping: : {}", avg_crossentropy))
# trainer.append_task(tasks.Print("L2 err : {0:.4f}", l2err_monitor, each_k_update=100))
# trainer.append_task(tasks.Print("stopping : {0:.4f}", crossentropy_monitor, each_k_update=100))
# Print NLL mean/stderror.
# train_loss = L2DistanceForSequences(model, trainset)
# train_batch_scheduler = StreamlinesBatchScheduler(trainset, batch_size=1000,
# noisy_streamlines_sigma=None,
# nb_updates_per_epoch=None,
# seed=1234)
# train_error = views.LossView(loss=train_loss, batch_scheduler=train_batch_scheduler)
# trainer.append_task(tasks.Print("Trainset - Error : {0:.2f} | {1:.2f}", train_error.sum, train_error.mean))
# HACK: To make sure all subjects in the volume_manager are used in a batch, we have to split the trainset/validset in 2 volume managers
model.volume_manager = validset_volume_manager
valid_loss = loss_factory(hyperparams, model, validset)
valid_batch_scheduler = batch_scheduler_factory(hyperparams,
dataset=validset,
train_mode=False)
valid_error = views.LossView(loss=valid_loss, batch_scheduler=valid_batch_scheduler)
trainer.append_task(tasks.Print("Validset - Error : {0:.2f} | {1:.2f}", valid_error.sum, valid_error.mean))
# HACK: Restore trainset volume manager
model.volume_manager = trainset_volume_manager
lookahead_loss = valid_error.sum
direction_norm = views.MonitorVariable(T.sqrt(sum(map(lambda d: T.sqr(d).sum(), loss.gradients.values()))))
# trainer.append_task(tasks.Print("||d|| : {0:.4f}", direction_norm))
# logger = tasks.Logger(train_error.mean, valid_error.mean, valid_error.sum, direction_norm)
logger = tasks.Logger(valid_error.mean, valid_error.sum, direction_norm)
trainer.append_task(logger)
if args.view:
import pylab as plt
def _plot(*args, **kwargs):
plt.figure(1)
plt.clf()
plt.show(False)
plt.subplot(121)
plt.plot(np.array(logger.get_variable_history(0)).flatten(), label="Train")
plt.plot(np.array(logger.get_variable_history(1)).flatten(), label="Valid")
plt.legend()
plt.subplot(122)
plt.plot(np.array(logger.get_variable_history(3)).flatten(), label="||d'||")
plt.draw()
trainer.append_task(tasks.Callback(_plot))
# Callback function to stop training if NaN is detected.
def detect_nan(obj, status):
if np.isnan(model.parameters[0].get_value().sum()):
print("NaN detected! Stopping training now.")
sys.exit()
trainer.append_task(tasks.Callback(detect_nan, each_k_update=1))
# Callback function to save training progression.
def save_training(obj, status):
trainer.save(experiment_path)
trainer.append_task(tasks.Callback(save_training))
# Early stopping with a callback for saving every time model improves.
def save_improvement(obj, status):
""" Save best model and training progression. """
if np.isnan(model.parameters[0].get_value().sum()):
print("NaN detected! Not saving the model. Crashing now.")
sys.exit()
print("*** Best epoch: {0} ***\n".format(obj.best_epoch))
model.save(experiment_path)
# Print time for one epoch
trainer.append_task(tasks.PrintEpochDuration())
trainer.append_task(tasks.PrintTrainingDuration())
trainer.append_task(tasks.PrintTime(each_k_update=100)) # Profiling
# Add stopping criteria
trainer.append_task(stopping_criteria.MaxEpochStopping(args.max_epoch))
early_stopping = stopping_criteria.EarlyStopping(lookahead_loss, lookahead=args.lookahead, eps=args.lookahead_eps, callback=save_improvement)
trainer.append_task(early_stopping)
with Timer("Compiling Theano graph"):
trainer.build_theano_graph()
if resuming:
if not os.path.isdir(pjoin(experiment_path, 'training')):
print("No 'training/' folder. Assuming it failed before"
" the end of the first epoch. Starting a new training.")
else:
with Timer("Loading"):
trainer.load(experiment_path)
with Timer("Training"):
trainer.train()
def main():
parser = build_argparser()
args = parser.parse_args()
print(args)
print("Using Theano v.{}".format(theano.version.short_version))
hyperparams_to_exclude = ['max_epoch', 'force', 'name', 'view', 'shuffle_streamlines']
# Use this for hyperparams added in a new version, but nonexistent from older versions
retrocompatibility_defaults = {'feed_previous_direction': False,
'predict_offset': False,
'normalize': False,
'sort_streamlines': False,
'keep_step_size': False,
'use_layer_normalization': False,
'drop_prob': 0.,
'use_zoneout': False,
'skip_connections': False}
experiment_path, hyperparams, resuming = utils.maybe_create_experiment_folder(args, exclude=hyperparams_to_exclude,
retrocompatibility_defaults=retrocompatibility_defaults)
# Log the command currently running.
with open(pjoin(experiment_path, 'cmd.txt'), 'a') as f:
f.write(" ".join(sys.argv) + "\n")
print("Resuming:" if resuming else "Creating:", experiment_path)
with Timer("Loading dataset", newline=True):
trainset_volume_manager = VolumeManager()
validset_volume_manager = VolumeManager()
trainset = datasets.load_tractography_dataset(args.train_subjects, trainset_volume_manager, name="trainset",
use_sh_coeffs=args.use_sh_coeffs)
validset = datasets.load_tractography_dataset(args.valid_subjects, validset_volume_manager, name="validset",
use_sh_coeffs=args.use_sh_coeffs)
print("Dataset sizes:", len(trainset), " |", len(validset))
batch_scheduler = batch_scheduler_factory(hyperparams, dataset=trainset, train_mode=True)
print("An epoch will be composed of {} updates.".format(batch_scheduler.nb_updates_per_epoch))
print(trainset_volume_manager.data_dimension, args.hidden_sizes, batch_scheduler.target_size)
with Timer("Creating model"):
input_size = trainset_volume_manager.data_dimension
if hyperparams['feed_previous_direction']:
input_size += 3
model = model_factory(hyperparams,
input_size=input_size,
output_size=batch_scheduler.target_size,
volume_manager=trainset_volume_manager)
model.initialize(weigths_initializer_factory(args.weights_initialization,
seed=args.initialization_seed))
with Timer("Building optimizer"):
loss = loss_factory(hyperparams, model, trainset)
if args.clip_gradient is not None:
loss.append_gradient_modifier(DirectionClipping(threshold=args.clip_gradient))
optimizer = optimizer_factory(hyperparams, loss)
with Timer("Building trainer"):
trainer = Trainer(optimizer, batch_scheduler)
# Log training error
loss_monitor = views.MonitorVariable(loss.loss)
avg_loss = tasks.AveragePerEpoch(loss_monitor)
trainer.append_task(avg_loss)
# Print average training loss.
trainer.append_task(tasks.Print("Avg. training loss: : {}", avg_loss))
# if args.learn_to_stop:
# l2err_monitor = views.MonitorVariable(T.mean(loss.mean_sqr_error))
# avg_l2err = tasks.AveragePerEpoch(l2err_monitor)
# trainer.append_task(avg_l2err)
#
# crossentropy_monitor = views.MonitorVariable(T.mean(loss.cross_entropy))
# avg_crossentropy = tasks.AveragePerEpoch(crossentropy_monitor)
# trainer.append_task(avg_crossentropy)
#
# trainer.append_task(tasks.Print("Avg. training L2 err: : {}", avg_l2err))
# trainer.append_task(tasks.Print("Avg. training stopping: : {}", avg_crossentropy))
# trainer.append_task(tasks.Print("L2 err : {0:.4f}", l2err_monitor, each_k_update=100))
# trainer.append_task(tasks.Print("stopping : {0:.4f}", crossentropy_monitor, each_k_update=100))
# Print NLL mean/stderror.
# train_loss = L2DistanceForSequences(model, trainset)
# train_batch_scheduler = StreamlinesBatchScheduler(trainset, batch_size=1000,
# noisy_streamlines_sigma=None,
# nb_updates_per_epoch=None,
# seed=1234)
# train_error = views.LossView(loss=train_loss, batch_scheduler=train_batch_scheduler)
# trainer.append_task(tasks.Print("Trainset - Error : {0:.2f} | {1:.2f}", train_error.sum, train_error.mean))
# HACK: To make sure all subjects in the volume_manager are used in a batch, we have to split the trainset/validset in 2 volume managers
model.volume_manager = validset_volume_manager
model.drop_prob = 0. # Do not use dropout/zoneout for evaluation
valid_loss = loss_factory(hyperparams, model, validset)
valid_batch_scheduler = batch_scheduler_factory(hyperparams,
dataset=validset,
train_mode=False)
valid_error = views.LossView(loss=valid_loss, batch_scheduler=valid_batch_scheduler)
trainer.append_task(tasks.Print("Validset - Error : {0:.2f} | {1:.2f}", valid_error.sum, valid_error.mean))
if hyperparams['model'] == 'ffnn_regression':
valid_batch_scheduler2 = batch_scheduler_factory(hyperparams,
dataset=validset,
train_mode=False)
valid_l2 = loss_factory(hyperparams, model, validset, loss_type="expected_value")
valid_l2_error = views.LossView(loss=valid_l2, batch_scheduler=valid_batch_scheduler2)
trainer.append_task(tasks.Print("Validset - {}".format(valid_l2.__class__.__name__) + "\t: {0:.2f} | {1:.2f}", valid_l2_error.sum, valid_l2_error.mean))
# HACK: Restore trainset volume manager
model.volume_manager = trainset_volume_manager
model.drop_prob = hyperparams['drop_prob'] # Restore dropout
lookahead_loss = valid_error.sum
direction_norm = views.MonitorVariable(T.sqrt(sum(map(lambda d: T.sqr(d).sum(), loss.gradients.values()))))
# trainer.append_task(tasks.Print("||d|| : {0:.4f}", direction_norm))
# logger = tasks.Logger(train_error.mean, valid_error.mean, valid_error.sum, direction_norm)
logger = tasks.Logger(valid_error.mean, valid_error.sum, direction_norm)
trainer.append_task(logger)
if args.view:
import pylab as plt
def _plot(*args, **kwargs):
plt.figure(1)
plt.clf()
plt.show(False)
plt.subplot(121)
plt.plot(np.array(logger.get_variable_history(0)).flatten(), label="Train")
plt.plot(np.array(logger.get_variable_history(1)).flatten(), label="Valid")
plt.legend()
plt.subplot(122)
plt.plot(np.array(logger.get_variable_history(3)).flatten(), label="||d'||")
plt.draw()
trainer.append_task(tasks.Callback(_plot))
# Callback function to stop training if NaN is detected.
def detect_nan(obj, status):
if np.isnan(model.parameters[0].get_value().sum()):
print("NaN detected! Stopping training now.")
sys.exit()
trainer.append_task(tasks.Callback(detect_nan, each_k_update=1))
# Callback function to save training progression.
def save_training(obj, status):
trainer.save(experiment_path)
trainer.append_task(tasks.Callback(save_training))
# Early stopping with a callback for saving every time model improves.
def save_improvement(obj, status):
""" Save best model and training progression. """
if np.isnan(model.parameters[0].get_value().sum()):
print("NaN detected! Not saving the model. Crashing now.")
sys.exit()
print("*** Best epoch: {0} ***\n".format(obj.best_epoch))
model.save(experiment_path)
# Print time for one epoch
trainer.append_task(tasks.PrintEpochDuration())
trainer.append_task(tasks.PrintTrainingDuration())
trainer.append_task(tasks.PrintTime(each_k_update=100)) # Profiling
# Add stopping criteria
trainer.append_task(stopping_criteria.MaxEpochStopping(args.max_epoch))
early_stopping = stopping_criteria.EarlyStopping(lookahead_loss, lookahead=args.lookahead, eps=args.lookahead_eps, callback=save_improvement)
trainer.append_task(early_stopping)
with Timer("Compiling Theano graph"):
trainer.build_theano_graph()
if resuming:
if not os.path.isdir(pjoin(experiment_path, 'training')):
print("No 'training/' folder. Assuming it failed before"
" the end of the first epoch. Starting a new training.")
else:
with Timer("Loading"):
trainer.load(experiment_path)
with Timer("Training"):
trainer.train()
def main():
parser = build_parser()
args = parser.parse_args()
print(args)
# Get experiment folder
experiment_path = args.name
if not os.path.isdir(experiment_path):
# If not a directory, it must be the name of the experiment.
experiment_path = pjoin(".", "experiments", args.name)
if not os.path.isdir(experiment_path):
parser.error('Cannot find experiment: {0}!'.format(args.name))
# Load experiments hyperparameters
try:
hyperparams = smartutils.load_dict_from_json_file(pjoin(experiment_path, "hyperparams.json"))
except FileNotFoundError:
hyperparams = smartutils.load_dict_from_json_file(pjoin(experiment_path, "..", "hyperparams.json"))
with Timer("Loading dataset", newline=True):
volume_manager = VolumeManager()
dataset = datasets.load_tractography_dataset(args.subjects, volume_manager, name="dataset", use_sh_coeffs=hyperparams['use_sh_coeffs'])
print("Dataset size:", len(dataset))
with Timer("Loading model"):
model = None
if hyperparams['model'] == 'gru_regression':
from learn2track.models import GRU_Regression
model = GRU_Regression.create(experiment_path, volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_mixture':
from learn2track.models import GRU_Mixture
model = GRU_Mixture.create(experiment_path, volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_multistep':
from learn2track.models import GRU_Multistep_Gaussian
model = GRU_Multistep_Gaussian.create(experiment_path, volume_manager=volume_manager)
model.k = 1
model.m = 1
elif hyperparams['model'] == 'ffnn_regression':
from learn2track.models import FFNN_Regression
model = FFNN_Regression.create(experiment_path, volume_manager=volume_manager)
else:
raise NameError("Unknown model: {}".format(hyperparams['model']))
with Timer("Building evaluation function"):
# Override K for gru_multistep
if 'k' in hyperparams:
hyperparams['k'] = 1
batch_scheduler = batch_scheduler_factory(hyperparams, dataset, train_mode=False, batch_size_override=args.batch_size)
loss = loss_factory(hyperparams, model, dataset, loss_type=args.loss_type)
l2_error = views.LossView(loss=loss, batch_scheduler=batch_scheduler)
with Timer("Evaluating...", newline=True):
results_file = pjoin(experiment_path, "results.json")
results = {}
if os.path.isfile(results_file) and not args.force:
print("Loading saved results... (use --force to re-run evaluation)")
results = smartutils.load_dict_from_json_file(results_file)
tag = ""
if args.loss_type == 'expected_value' or hyperparams['model'] == 'gru_regression':
tag = "_EV_L2_error"
elif args.loss_type == 'maximum_component':
tag = "_MC_L2_error"
elif hyperparams['model'] == 'gru_mixture' or hyperparams['model'] == 'gru_multistep':
tag = "_NLL"
entry = args.dataset_name + tag
if entry not in results or args.force:
with Timer("Evaluating {}".format(entry)):
dummy_status = Status() # Forces recomputing results
results[entry] = {'mean': float(l2_error.mean.view(dummy_status)), 'stderror': float(l2_error.stderror.view(dummy_status))}
smartutils.save_dict_to_json_file(results_file, results) # Update results file.
print("{}: {:.4f} ± {:.4f}".format(entry, results[entry]['mean'], results[entry]['stderror']))
def main():
parser = build_parser()
args = parser.parse_args()
print(args)
# Get experiment folder
experiment_path = args.name
if not os.path.isdir(experiment_path):
# If not a directory, it must be the name of the experiment.
experiment_path = pjoin(".", "experiments", args.name)
if not os.path.isdir(experiment_path):
parser.error('Cannot find experiment: {0}!'.format(args.name))
# Load experiments hyperparameters
try:
hyperparams = smartutils.load_dict_from_json_file(
pjoin(experiment_path, "hyperparams.json"))
except FileNotFoundError:
hyperparams = smartutils.load_dict_from_json_file(
pjoin(experiment_path, "..", "hyperparams.json"))
# Use this for hyperparams added in a new version, but nonexistent from older versions
retrocompatibility_defaults = {
'feed_previous_direction': False,
'predict_offset': False,
'normalize': False,
'keep_step_size': False,
'sort_streamlines': False,
'use_layer_normalization': False,
'drop_prob': 0.,
'use_zoneout': False
}
for new_hyperparams, default_value in retrocompatibility_defaults.items():
if new_hyperparams not in hyperparams:
hyperparams[new_hyperparams] = default_value
with Timer("Loading dataset", newline=True):
volume_manager = VolumeManager()
dataset = datasets.load_tractography_dataset(
args.subjects,
volume_manager,
name="dataset",
use_sh_coeffs=hyperparams['use_sh_coeffs'])
print("Dataset size:", len(dataset))
with Timer("Loading model"):
model = None
if hyperparams['model'] == 'gru_regression':
from learn2track.models import GRU_Regression
model = GRU_Regression.create(experiment_path,
volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_gaussian':
from learn2track.models import GRU_Gaussian
model = GRU_Gaussian.create(experiment_path,
volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_mixture':
from learn2track.models import GRU_Mixture
model = GRU_Mixture.create(experiment_path,
volume_manager=volume_manager)
elif hyperparams['model'] == 'gru_multistep':
from learn2track.models import GRU_Multistep_Gaussian
model = GRU_Multistep_Gaussian.create(
experiment_path, volume_manager=volume_manager)
model.k = 1
model.m = 1
elif hyperparams['model'] == 'ffnn_regression':
from learn2track.models import FFNN_Regression
model = FFNN_Regression.create(experiment_path,
volume_manager=volume_manager)
else:
raise NameError("Unknown model: {}".format(hyperparams['model']))
model.drop_prob = 0. # Make sure dropout/zoneout is not used when testing
with Timer("Building evaluation function"):
# Override K for gru_multistep
if 'k' in hyperparams:
hyperparams['k'] = 1
batch_scheduler = batch_scheduler_factory(
hyperparams,
dataset,
train_mode=False,
batch_size_override=args.batch_size)
loss = loss_factory(hyperparams,
model,
dataset,
loss_type=args.loss_type)
l2_error = views.LossView(loss=loss, batch_scheduler=batch_scheduler)
with Timer("Evaluating...", newline=True):
results_file = pjoin(experiment_path, "results.json")
results = {}
if os.path.isfile(results_file) and not args.force:
print(
"Loading saved results... (use --force to re-run evaluation)")
results = smartutils.load_dict_from_json_file(results_file)
tag = ""
if args.loss_type == 'expected_value' or hyperparams[
'model'] == 'gru_regression':
tag = "_EV_L2_error"
elif args.loss_type == 'maximum_component':
tag = "_MC_L2_error"
elif hyperparams['model'] in [
'gru_gaussian', 'gru_mixture', 'gru_multistep'
]:
tag = "_NLL"
entry = args.dataset_name + tag
if entry not in results or args.force:
with Timer("Evaluating {}".format(entry)):
dummy_status = Status() # Forces recomputing results
results[entry] = {
'mean': float(l2_error.mean.view(dummy_status)),
'stderror': float(l2_error.stderror.view(dummy_status))
}
smartutils.save_dict_to_json_file(
results_file, results) # Update results file.
print("{}: {:.4f} ± {:.4f}".format(entry, results[entry]['mean'],
results[entry]['stderror']))
def evaluation_tractogram(hyperparams, model, dataset, batch_size_override, metric):
loss = loss_factory(hyperparams, model, dataset, loss_type=None)
batch_scheduler = batch_scheduler_factory(hyperparams, dataset, train_mode=False, batch_size_override=batch_size_override, use_data_augment=False)
_ = loss.losses # Hack to generate update dict in loss :(
if hyperparams['model'] == 'ffnn_regression':
timestep_losses, inputs, targets = log_variables(batch_scheduler,
model,
loss.loss_per_time_step,
dataset.symb_inputs * 1,
dataset.symb_targets * 1)
# Regrouping data into streamlines will only work if the original streamlines were NOT shuffled, resampled or augmented
timesteps_loss = ArraySequence()
seq_loss = []
timesteps_inputs = ArraySequence()
timesteps_targets = ArraySequence()
idx = 0
for length in dataset.streamlines._lengths:
start = idx
idx = end = idx+length
timesteps_loss.extend(timestep_losses[start:end])
seq_loss.extend(np.mean(timestep_losses[start:end]))
timesteps_inputs.extend(inputs[start:end])
timesteps_targets.extend(targets[start:end])
else:
timestep_losses, seq_losses, inputs, targets, masks = log_variables(batch_scheduler,
model,
loss.loss_per_time_step,
loss.loss_per_seq,
dataset.symb_inputs * 1,
dataset.symb_targets * 1,
dataset.symb_mask * 1)
timesteps_loss = ArraySequence([l[:int(m.sum())] for l, m in zip(chain(*timestep_losses), chain(*masks))])
seq_loss = np.array(list(chain(*seq_losses)))
timesteps_inputs = ArraySequence([i[:int(m.sum())] for i, m in zip(chain(*inputs), chain(*masks))])
# Use np.squeeze in case gru_multistep is used to remove the empty k=1 dimension
timesteps_targets = ArraySequence([np.squeeze(t[:int(m.sum())]) for t, m in zip(chain(*targets), chain(*masks))])
if metric == 'sequence':
# Color is based on sequence loss
values = seq_loss
elif metric == 'timestep' or metric == 'cumul_avg':
# Color is based on timestep loss
values = np.concatenate(timesteps_loss)
else:
raise ValueError("Unrecognized metric: {}".format(metric))
cmap = cm.get_cmap('bwr')
vmin = np.percentile(values, 5)
vmax = np.percentile(values, 95)
scalar_map = cm.ScalarMappable(norm=mplcolors.Normalize(vmin=vmin, vmax=vmax), cmap=cmap)
streamlines = []
colors = []
for i, t, l, seq_l in zip(timesteps_inputs, timesteps_targets, timesteps_loss, seq_loss):
pts = np.r_[i[:, :3], [i[-1, :3] + t[-1]]]
color = np.zeros_like(pts)
if metric == 'sequence':
# Streamline color is based on sequence loss
color[:, :] = scalar_map.to_rgba(seq_l, bytes=True)[:3]
elif metric == 'timestep':
# Streamline color is based on timestep loss
# Identify first point with green
color[0, :] = [0, 255, 0]
color[1:, :] = scalar_map.to_rgba(l, bytes=True)[:, :3]
elif metric == 'cumul_avg':
# Streamline color is based on timestep loss
# Compute cumulative average
cumul_avg = np.cumsum(l) / np.arange(1, len(l) + 1)
# Identify first point with green
color[0, :] = [0, 255, 0]
color[1:, :] = scalar_map.to_rgba(cumul_avg, bytes=True)[:, :3]
else:
raise ValueError("Unrecognized metric: {}".format(metric))
streamlines.append(pts)
colors.append(color)
tractogram = nib.streamlines.Tractogram(streamlines, data_per_point={"colors": colors})
return tractogram
def prediction_tractogram(hyperparams, model, dataset, batch_size_override, prediction_method):
loss = loss_factory(hyperparams, model, dataset, loss_type=prediction_method)
batch_scheduler = batch_scheduler_factory(hyperparams, dataset, train_mode=False, batch_size_override=batch_size_override, use_data_augment=False)
_ = loss.losses # Hack to generate update dict in loss :(
predictions = loss.samples
predict, timestep_losses, inputs, targets, masks = log_variables(batch_scheduler,
model,
predictions,
loss.loss_per_time_step,
dataset.symb_inputs * 1,
dataset.symb_targets * 1,
dataset.symb_mask * 1)
if hyperparams['model'] == 'ffnn_regression':
# Regrouping data into streamlines will only work if the original streamlines were NOT shuffled, resampled or augmented
timesteps_prediction = ArraySequence()
timesteps_loss = ArraySequence()
timesteps_inputs = ArraySequence()
timesteps_targets = ArraySequence()
idx = 0
for length in dataset.streamlines._lengths:
start = idx
idx = end = idx+length
timesteps_prediction.extend(predict[start:end])
timesteps_loss.extend(timestep_losses[start:end])
timesteps_inputs.extend(inputs[start:end])
timesteps_targets.extend(targets[start:end])
else:
timesteps_prediction = ArraySequence([p[:int(m.sum())] for p, m in zip(chain(*predict), chain(*masks))])
timesteps_loss = ArraySequence([l[:int(m.sum())] for l, m in zip(chain(*timestep_losses), chain(*masks))])
timesteps_inputs = ArraySequence([i[:int(m.sum())] for i, m in zip(chain(*inputs), chain(*masks))])
# Use np.squeeze in case gru_multistep is used to remove the empty k=1 dimension
timesteps_targets = ArraySequence([np.squeeze(t[:int(m.sum())]) for t, m in zip(chain(*targets), chain(*masks))])
# Debug : Print norm stats
# print("Dataset: {}; # of streamlines: {}".format(dataset.name, len(dataset)))
# all_predictions = np.array(list(chain(*timesteps_prediction)))
# prediction_norms = np.linalg.norm(all_predictions, axis=1)
# print("Prediction norm --- Mean:{}; Max:{}; Min:{}".format(np.mean(prediction_norms), np.max(prediction_norms), np.min(prediction_norms)))
# all_targets = np.array(list(chain(*timesteps_targets)))
# target_norms = np.linalg.norm(all_targets, axis=1)
# print("Target norm --- Mean:{}; Max:{}; Min:{}".format(np.mean(target_norms), np.max(target_norms), np.min(target_norms)))
# Color is based on timestep loss
cmap = cm.get_cmap('bwr')
values = np.concatenate(timesteps_loss)
vmin = np.percentile(values, 5)
vmax = np.percentile(values, 95)
scalar_map = cm.ScalarMappable(norm=mplcolors.Normalize(vmin=vmin, vmax=vmax), cmap=cmap)
streamlines = []
colors = []
for i, t, p, l in zip(timesteps_inputs, timesteps_targets, timesteps_prediction, timesteps_loss):
pts = np.r_[i[:, :3], [i[-1, :3] + t[-1]]]
streamline = np.zeros(((len(pts) - 1) * 3 + 1, 3))
streamline[::3] = pts
streamline[1:-1:3] = pts[:-1] + p
streamline[2:-1:3] = pts[:-1]
streamlines.append(streamline)
# Color input streamlines in a uniform color, then color predictions based on L2 error
color = np.zeros_like(streamline)
# Base color of streamlines is minimum value (best score)
color[:] = scalar_map.to_rgba(vmin, bytes=True)[:3]
color[1:-1:3, :] = scalar_map.to_rgba(l, bytes=True)[:, :3]
colors.append(color)
tractogram = nib.streamlines.Tractogram(streamlines, data_per_point={"colors": colors})
return tractogram
def test_gru_mixture_fprop_neighborhood():
hyperparams = {
'model': 'gru_mixture',
'SGD': "1e-2",
'hidden_sizes': 50,
'batch_size': 16,
'learn_to_stop': False,
'normalize': True,
'activation': 'tanh',
'feed_previous_direction': False,
'predict_offset': False,
'use_layer_normalization': False,
'drop_prob': 0.,
'use_zoneout': False,
'skip_connections': False,
'seed': 1234,
'noisy_streamlines_sigma': None,
'keep_step_size': True,
'sort_streamlines': False,
'n_gaussians': 2,
'neighborhood_radius': 0.5
}
with Timer("Creating dataset", newline=True):
volume_manager = neurotools.VolumeManager()
trainset = make_dummy_dataset(volume_manager)
print("Dataset sizes:", len(trainset))
batch_scheduler = factories.batch_scheduler_factory(hyperparams,
dataset=trainset)
print("An epoch will be composed of {} updates.".format(
batch_scheduler.nb_updates_per_epoch))
print(volume_manager.data_dimension, hyperparams['hidden_sizes'],
batch_scheduler.target_size)
with Timer("Creating model"):
model = factories.model_factory(
hyperparams,
input_size=volume_manager.data_dimension,
output_size=batch_scheduler.target_size,
volume_manager=volume_manager)
model.initialize(
factories.weigths_initializer_factory("orthogonal", seed=1234))
print("Input size: {}".format(model.model_input_size))
# Test fprop with missing streamlines from one subject in a batch
output = model.get_output(trainset.symb_inputs)
fct = theano.function([trainset.symb_inputs],
output,
updates=model.graph_updates)
batch_inputs, batch_targets, batch_mask = batch_scheduler._next_batch(2)
out = fct(batch_inputs)
with Timer("Building optimizer"):
loss = factories.loss_factory(hyperparams, model, trainset)
optimizer = factories.optimizer_factory(hyperparams, loss)
fct_loss = theano.function(
[trainset.symb_inputs, trainset.symb_targets, trainset.symb_mask],
loss.loss,
updates=model.graph_updates)
loss_value = fct_loss(batch_inputs, batch_targets, batch_mask)
print("Loss:", loss_value)
fct_optim = theano.function(
[trainset.symb_inputs, trainset.symb_targets, trainset.symb_mask],
list(optimizer.directions.values()),
updates=model.graph_updates)
dirs = fct_optim(batch_inputs, batch_targets, batch_mask)
return True
