def test_serialization(self):
from inferno.trainers.basic import Trainer
import os
# Make model
net = self._make_test_model()
# Make trainer
trainer = Trainer(model=net) \
.build_optimizer('Adam') \
.build_criterion('CrossEntropyLoss') \
.build_metric('CategoricalError') \
.validate_every((1, 'epochs')) \
.save_every((1, 'epochs'), to_directory=os.path.join(self.ROOT_DIR, 'saves')) \
.save_at_best_validation_score() \
.set_max_num_epochs(2)
# Try to serialize
trainer.save()
# Try to unserialize
trainer = Trainer(net).save_to_directory(
os.path.join(self.ROOT_DIR, 'saves')).load()
# Make sure everything survived (especially the logger)
self.assertEqual(trainer._logger.__class__.__name__,
'BasicTensorboardLogger')
def set_up_training(project_directory, config, data_config,
load_pretrained_model):
# Get model
if load_pretrained_model:
model = Trainer().load(from_directory=project_directory,
filename='Weights/checkpoint.pytorch').model
else:
model_name = config.get('model_name')
model = getattr(models, model_name)(**config.get('model_kwargs'))
criterion = SorensenDiceLoss()
loss_train = LossWrapper(criterion=criterion,
transforms=Compose(ApplyAndRemoveMask(),
InvertTarget()))
loss_val = LossWrapper(criterion=criterion,
transforms=Compose(RemoveSegmentationFromTarget(),
ApplyAndRemoveMask(),
InvertTarget()))
# Build trainer and validation metric
logger.info("Building trainer.")
smoothness = 0.95
offsets = data_config['volume_config']['segmentation']['affinity_config'][
'offsets']
metric = ArandErrorFromMulticut(average_slices=False,
use_2d_ws=True,
n_threads=8,
weight_edges=True,
offsets=offsets)
trainer = Trainer(model)\
.save_every((1000, 'iterations'),
to_directory=os.path.join(project_directory, 'Weights'))\
.build_criterion(loss_train)\
.build_validation_criterion(loss_val)\
.build_optimizer(**config.get('training_optimizer_kwargs'))\
.evaluate_metric_every('never')\
.validate_every((100, 'iterations'), for_num_iterations=1)\
.register_callback(SaveAtBestValidationScore(smoothness=smoothness, verbose=True))\
.build_metric(metric)\
.register_callback(AutoLR(factor=0.98,
patience='100 iterations',
monitor_while='validating',
monitor_momentum=smoothness,
consider_improvement_with_respect_to='previous'))\
.register_callback(GarbageCollection())
logger.info("Building logger.")
# Build logger
tensorboard = TensorboardLogger(
log_scalars_every=(1, 'iteration'),
log_images_every=(100, 'iterations'),
log_histograms_every='never').observe_states(
['validation_input', 'validation_prediction, validation_target'],
observe_while='validating')
trainer.build_logger(tensorboard,
log_directory=os.path.join(project_directory, 'Logs'))
return trainer
def test_save(self):
from inferno.trainers.basic import Trainer
trainer = Trainer().save_to_directory(
to_directory=self.ROOT_DIR, checkpoint_filename='dummy.pytorch')
trainer.save()
# Instantiate new trainer and load
trainer = Trainer().load(from_directory=self.ROOT_DIR,
filename='dummy.pytorch')
def set_up_training(project_directory, config, data_config, criterion, balance,
load_pretrained_model):
# Get model
if load_pretrained_model:
model = Trainer().load(from_directory=project_directory,
filename='Weights/checkpoint.pytorch').model
else:
model_name = config.get('model_name')
model = getattr(models, model_name)(**config.get('model_kwargs'))
# TODO
logger.info("Using criterion: %s" % criterion)
# TODO this should go somewhere more prominent
affinity_offsets = data_config['volume_config']['segmentation'][
'affinity_offsets']
# TODO implement affinities on gpu again ?!
criterion = CRITERIA[criterion]
loss = LossWrapper(
criterion=criterion(),
transforms=Compose(MaskTransitionToIgnoreLabel(affinity_offsets),
RemoveSegmentationFromTarget(), InvertTarget()),
weight_function=BalanceAffinities(
ignore_label=0, offsets=affinity_offsets) if balance else None)
# Build trainer and validation metric
logger.info("Building trainer.")
smoothness = 0.95
# use multicut pipeline for validation
metric = ArandErrorFromSegmentationPipeline(
local_affinity_multicut_from_wsdt2d(n_threads=10, time_limit=120))
trainer = Trainer(model)\
.save_every((1000, 'iterations'), to_directory=os.path.join(project_directory, 'Weights'))\
.build_criterion(loss)\
.build_optimizer(**config.get('training_optimizer_kwargs'))\
.evaluate_metric_every('never')\
.validate_every((100, 'iterations'), for_num_iterations=1)\
.register_callback(SaveAtBestValidationScore(smoothness=smoothness, verbose=True))\
.build_metric(metric)\
.register_callback(AutoLR(factor=0.98,
patience='100 iterations',
monitor_while='validating',
monitor_momentum=smoothness,
consider_improvement_with_respect_to='previous'))
logger.info("Building logger.")
# Build logger
tensorboard = TensorboardLogger(
log_scalars_every=(1, 'iteration'),
log_images_every=(100, 'iterations')).observe_states(
['validation_input', 'validation_prediction, validation_target'],
observe_while='validating')
trainer.build_logger(tensorboard,
log_directory=os.path.join(project_directory, 'Logs'))
return trainer
def set_up_training(project_directory, config, data_config,
load_pretrained_model):
# Get model
if load_pretrained_model:
model = Trainer().load(from_directory=project_directory,
filename='Weights/checkpoint.pytorch').model
else:
model_name = config.get('model_name')
model = getattr(models, model_name)(**config.get('model_kwargs'))
affinity_offsets = data_config['volume_config']['segmentation'][
'affinity_offsets']
loss = MultiOutputLossWrapper(
criterion=SorensenDiceLoss(),
transforms=Compose(MaskTransitionToIgnoreLabel(affinity_offsets),
RemoveSegmentationFromTarget(), InvertTarget()))
# Build trainer and validation metric
logger.info("Building trainer.")
smoothness = 0.95
# use multicut pipeline for validation
# metric = ArandErrorFromSegmentationPipeline(local_affinity_multicut_from_wsdt2d(n_threads=10,
# time_limit=120))
# use damws for validation
stride = [2, 10, 10]
metric = ArandErrorFromSegmentationPipeline(
DamWatershed(affinity_offsets, stride, randomize_bounds=False))
trainer = Trainer(model)\
.save_every((1000, 'iterations'), to_directory=os.path.join(project_directory, 'Weights'))\
.build_criterion(loss)\
.build_optimizer(**config.get('training_optimizer_kwargs'))\
.evaluate_metric_every('never')\
.validate_every((100, 'iterations'), for_num_iterations=1)\
.register_callback(SaveAtBestValidationScore(smoothness=smoothness, verbose=True))\
.build_metric(metric)\
.register_callback(AutoLR(factor=0.98,
patience='100 iterations',
monitor_while='validating',
monitor_momentum=smoothness,
consider_improvement_with_respect_to='previous'))
# FIXME some issues with conda tf for torch0.3 env
# logger.info("Building logger.")
# # Build logger
# tensorboard = TensorboardLogger(log_scalars_every=(1, 'iteration'),
# log_images_every=(100, 'iterations')).observe_states(
# ['validation_input', 'validation_prediction, validation_target'],
# observe_while='validating'
# )
# trainer.build_logger(tensorboard, log_directory=os.path.join(project_directory, 'Logs'))
return trainer
def test_serialization(self):
trainer = Trainer()\
.build_logger(logger=DummyLogger())\
.save_to_directory(join(self.ROOT, 'saves'))
trainer.save()
# Unserialize
trainer = Trainer().load(from_directory=join(self.ROOT, 'saves'))
# Check if the loggers are consistent
logger_from_trainer = trainer._logger
logger_from_callback_engine = \
next(iter(trainer.callbacks._callback_registry['end_of_training_iteration']))
self.assertIs(logger_from_trainer, logger_from_callback_engine)
self.assertIs(logger_from_callback_engine.trainer, trainer)
def set_up_training(project_directory, config, data_config,
load_pretrained_model, max_iters):
# Get model
if load_pretrained_model:
model = Trainer().load(from_directory=project_directory,
filename='Weights/checkpoint.pytorch').model
else:
model_name = config.get('model_name')
model = getattr(models, model_name)(**config.get('model_kwargs'))
loss = LossWrapper(criterion=SorensenDiceLoss(),
transforms=Compose(MaskIgnoreLabel(),
RemoveSegmentationFromTarget()))
# TODO loss transforms:
# - Invert Target ???
# Build trainer and validation metric
logger.info("Building trainer.")
# smoothness = 0.95
# TODO set up validation ?!
trainer = Trainer(model)\
.save_every((1000, 'iterations'), to_directory=os.path.join(project_directory, 'Weights'))\
.build_criterion(loss)\
.build_optimizer(**config.get('training_optimizer_kwargs'))\
.evaluate_metric_every('never')\
.register_callback(ManualLR(decay_specs=[((k * 100, 'iterations'), 0.99)
for k in range(1, max_iters // 100)]))
# .validate_every((100, 'iterations'), for_num_iterations=1)\
# .register_callback(SaveAtBestValidationScore(smoothness=smoothness, verbose=True))\
# .build_metric(metric)\
# .register_callback(AutoLR(factor=0.98,
# patience='100 iterations',
# monitor_while='validating',
# monitor_momentum=smoothness,
# consider_improvement_with_respect_to='previous'))
logger.info("Building logger.")
# Build logger
tensorboard = TensorboardLogger(
log_scalars_every=(1, 'iteration'),
log_images_every=(100, 'iterations')) # .observe_states(
# ['validation_input', 'validation_prediction, validation_target'],
# observe_while='validating'
# )
trainer.build_logger(tensorboard,
log_directory=os.path.join(project_directory, 'Logs'))
return trainer
def train_model(args):
"""
Performs the training
"""
if os.path.exists(
os.path.join(args.save_directory,
Trainer()._checkpoint_filename)):
# Skip training if checkpoint exists
return
model = WsjModel(args)
kwargs = {
'num_workers': args.num_workers,
'pin_memory': True
} if args.cuda else {}
train_loader = DataLoader(WsjDataset('train', args),
shuffle=True,
batch_size=args.batch_size,
collate_fn=wsj_collate_fn,
**kwargs)
validate_loader = DataLoader(WsjDataset('dev', args),
shuffle=True,
batch_size=args.batch_size,
collate_fn=wsj_collate_fn,
**kwargs)
# Build trainer
trainer = Trainer(model) \
.build_criterion(ctc_loss()) \
.build_metric(ctc_loss()) \
.build_optimizer('Adam') \
.validate_every((1, 'epochs')) \
.save_every((1, 'epochs')) \
.save_to_directory(args.save_directory) \
.set_max_num_epochs(args.epochs) \
#.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
#log_images_every='never'),
#log_directory=args.save_directory)
# Bind loaders
trainer.bind_loader('train', train_loader, num_inputs=3)
trainer.bind_loader('validate', validate_loader, num_inputs=3)
trainer.register_callback(EpochTimer())
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
trainer.save()
def set_up_training(project_directory, config, data_config,
load_pretrained_model):
# Get model
if load_pretrained_model:
model = Trainer().load(from_directory=project_directory,
filename='Weights/checkpoint.pytorch').model
else:
model_name = config.get('model_name')
model = getattr(models, model_name)(**config.get('model_kwargs'))
loss = dice_loss()
loss_val = dice_loss(is_val=True)
metric = mws_metric()
# metric = loss_val
# Build trainer and validation metric
logger.info("Building trainer.")
smoothness = 0.9
trainer = Trainer(model)\
.save_every((1000, 'iterations'),
to_directory=os.path.join(project_directory, 'Weights'))\
.build_criterion(loss)\
.build_validation_criterion(loss_val)\
.build_optimizer(**config.get('training_optimizer_kwargs'))\
.evaluate_metric_every('never')\
.validate_every((100, 'iterations'), for_num_iterations=5)\
.register_callback(SaveAtBestValidationScore(smoothness=smoothness, verbose=True))\
.build_metric(metric)\
.register_callback(AutoLR(factor=0.98,
patience='100 iterations',
monitor_while='validating',
monitor_momentum=smoothness,
consider_improvement_with_respect_to='previous'))\
.register_callback(GarbageCollection())\
logger.info("Building logger.")
# Build logger
tensorboard = TensorboardLogger(
log_scalars_every=(1, 'iteration'),
log_images_every=(100, 'iterations'),
log_histograms_every='never').observe_states(
['validation_input', 'validation_prediction, validation_target'],
observe_while='validating')
trainer.build_logger(tensorboard,
log_directory=os.path.join(project_directory, 'Logs'))
return trainer
def test_multi_gpu(self):
import torch
if not torch.cuda.is_available():
return
from inferno.trainers.basic import Trainer
from inferno.io.box.cifar10 import get_cifar10_loaders
import os
# Make model
net = self._make_test_model()
# Make trainer
trainer = Trainer(model=net) \
.build_optimizer('Adam') \
.build_criterion('CrossEntropyLoss') \
.build_metric('CategoricalError') \
.validate_every((1, 'epochs')) \
.save_every((1, 'epochs'), to_directory=os.path.join(self.ROOT_DIR, 'saves')) \
.save_at_best_validation_score() \
.set_max_num_epochs(2)\
.cuda(devices=[0, 1, 2, 3])
train_loader, validate_loader = get_cifar10_loaders(
root_directory=self.ROOT_DIR, download=True)
trainer.bind_loader('train', train_loader)
trainer.bind_loader('validate', validate_loader)
trainer.fit()
def train_model(model, train_dataset, valid_dataset, args):
kw = {'num_workers': 2, 'pin_memory': True} if args.cuda else {}
train_loader = DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True,
**kw)
validate_loader = DataLoader(dataset=valid_dataset,
batch_size=args.batch_size,
**kw)
trainer = Trainer(model) \
.build_criterion(CrossEntropyLoss3D) \
.build_metric(CategoricalError3D) \
.build_optimizer('Adam', weight_decay=1e-6) \
.save_every((1, 'epochs')) \
.validate_every((100, 'iteration')) \
.save_to_directory(args.save_directory) \
.set_max_num_epochs(args.epochs) \
.build_logger(
TensorboardLogger(log_scalars_every=(1, 'iteration'), log_images_every='never'),
log_directory='tb_log/'
) \
.bind_loader('train', train_loader) \
.bind_loader('validate', validate_loader)
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
def train_model(args):
model = MNISTCNNModel()
train_loader, validate_loader = mnist_data_loaders(args)
# Build trainer
trainer = Trainer(model) \
.build_criterion('CrossEntropyLoss') \
.build_metric('CategoricalError') \
.build_optimizer('Adam') \
.validate_every((2, 'epochs')) \
.save_every((5, 'epochs')) \
.save_to_directory(args.save_directory) \
.set_max_num_epochs(args.epochs) \
.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
log_images_every='never'),
log_directory=args.save_directory)
# Bind loaders
trainer \
.bind_loader('train', train_loader) \
.bind_loader('validate', validate_loader)
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
def test_cifar(self):
from inferno.trainers.basic import Trainer
from inferno.io.box.cifar10 import get_cifar10_loaders
# Build cifar10 loaders
trainloader, testloader = get_cifar10_loaders(
root_directory=join(self.ROOT_DIR, 'data'),
download=self.DOWNLOAD_CIFAR)
# Make model
net = self._make_test_model()
tic = time.time()
# Make trainer
trainer = Trainer(model=net)\
.build_optimizer('Adam')\
.build_criterion('CrossEntropyLoss')\
.build_metric('CategoricalError')\
.validate_every((1, 'epochs'))\
.save_every((1, 'epochs'), to_directory=join(self.ROOT_DIR, 'saves'))\
.save_at_best_validation_score()\
.set_max_num_epochs(2)
# Bind trainer to datasets
trainer.bind_loader('train',
trainloader).bind_loader('validate', testloader)
# Check device and fit
if self.CUDA:
if self.HALF_PRECISION:
trainer.cuda().set_precision('half').fit()
else:
trainer.cuda().fit()
else:
trainer.fit()
toc = time.time()
print("[*] Elapsed time: {} seconds.".format(toc - tic))
def setUp(self):
# Build model
net = self._make_test_model()
# Build trainer
self.trainer = Trainer(net)\
.build_logger(TensorboardLogger(send_image_at_batch_indices=0,
send_image_at_channel_indices='all',
log_images_every=(20, 'iterations')),
log_directory=os.path.join(self.ROOT_DIR, 'logs'))\
.build_criterion('CrossEntropyLoss')\
.build_metric('CategoricalError')\
.build_optimizer('Adam')\
.validate_every((1, 'epochs'))\
.save_every((2, 'epochs'), to_directory=os.path.join(self.ROOT_DIR, 'saves'))\
.save_at_best_validation_score()\
.set_max_num_epochs(2)\
.cuda().set_precision(self.PRECISION)
# Load CIFAR10 data
train_loader, test_loader = \
get_cifar10_loaders(root_directory=os.path.join(self.ROOT_DIR, 'data'),
download=self.DOWNLOAD_CIFAR)
# Bind loaders
self.trainer.bind_loader('train', train_loader).bind_loader('validate', test_loader)
def __init__(self,
model_path,
halo,
gpu=0,
use_best=True,
prep_model=None,
**augmentation_kwargs):
# load the model and prep it if specified
assert os.path.exists(model_path), model_path
trainer = Trainer().load(from_directory=model_path, best=use_best)
self.model = trainer.model.cuda(gpu)
self.model.eval()
if prep_model is not None:
self.model = prep_model(self.model)
self.gpu = gpu
self.halo = halo
self.lock = threading.Lock()
# build the test-time-augmenter if we have augmentation kwargs
if augmentation_kwargs:
assert TestTimeAugmenter is not None, "Need neurofire for test-time-augmentation"
self.offsets = augmentation_kwargs.pop('offsets', None)
print(augmentation_kwargs)
self.augmenter = self.build_augmenter(**augmentation_kwargs)
else:
self.augmenter = None
def run(args):
dataset = AlderleyWrapper()
save_args(args) # save command line to a file for reference
train_loader = alderley_cgan_data_loader(args, dataset=dataset) # get the data
model = CGANModel(
args,
discriminators=[PixelDiscriminator(), patchCDiscriminatorNetwork(args)],
generator=UnetUpsample())
# Build trainer
trainer = Trainer(model)
trainer.build_criterion(CWGANDiscriminatorLoss(penalty_weight=args.penalty_weight, model=model))
trainer.build_optimizer('Adam', [parameter for discriminator in model.discriminators for parameter in discriminator.parameters()], lr=args.discriminator_lr)
trainer.save_every((1, 'epochs'))
trainer.save_to_directory(args.save_directory)
trainer.set_max_num_epochs(args.epochs)
trainer.register_callback(CGenerateDataCallback(args, dataset=dataset))
trainer.register_callback(CGeneratorTrainingCallback(
args,
parameters=model.generator.parameters(),
criterion=WGANGeneratorLoss(), dataset=dataset))
trainer.bind_loader('train', train_loader, num_inputs=2)
# Custom logging configuration so it knows to log our images
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
def train_model(args):
model = model_fn()
train_loader, validate_loader = mnist_data_loaders(args)
# Build trainer
trainer = Trainer(model) \
.build_criterion('RegularizedCrossEntropyLoss') \
.build_metric('CategoricalError') \
.build_optimizer('Adam') \
.validate_every((1, 'epochs')) \
.save_every((1, 'epochs')) \
.save_to_directory(args.save_directory) \
.set_max_num_epochs(args.epochs) \
.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
log_images_every='never'),
log_directory=args.save_directory)
# Record regularization losses
trainer.logger.observe_training_and_validation_states([
'main_loss', 'total_regularization_loss', 'activity_regularization',
'l1_weight_regularization'
])
# Bind loaders
trainer \
.bind_loader('train', train_loader) \
.bind_loader('validate', validate_loader)
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
def predict(project_folder, sample, only_nn_channels=True):
gpu = 0
checkpoint = os.path.join(project_folder, 'Weights')
data_config_file = './template_config/prediction_configs/sample%s.yml' % sample
print(
"[*] Loading CREMI with Configuration at: {}".format(data_config_file))
# Load CREMI sample
cremi = RawVolumeWithDefectAugmentation.from_config(data_config_file)
# Load model
trainer = Trainer().load(from_directory=checkpoint, best=True).cuda(gpu)
model = trainer.model
inference_config_file = './template_config/prediction_configs/inference_config.yml'
inference_engine = SimpleInferenceEngine.from_config(
inference_config_file, model)
output = inference_engine.infer(cremi)
print("[*] Output has shape {}".format(str(output.shape)))
save_folder = os.path.join(project_folder, 'Predictions')
if not os.path.exists(save_folder):
os.mkdir(save_folder)
save_path = os.path.join(save_folder, 'prediction_sample%s.h5' % sample)
if only_nn_channels:
output = output[:3]
save_path = save_path[:-3] + '_nnaffinities.h5'
toh5(output.astype('float32'), save_path, compression='lzf')
def train(net, dataset, criterion, num_epochs, batch_size, learn_rate,
dir_name):
dir_name = os.path.join('net/', dir_name)
trainer = Trainer(net[0])
if (os.path.exists(os.path.join(dir_name, 'model.pytorch'))):
net_temp = trainer.load_model(dir_name).model
net[0].load_state_dict(net_temp.state_dict())
print("Loaded checkpoint directly")
else:
if (not os.path.exists(dir_name)):
os.makedirs(dir_name)
data_loader = torch.utils.data.DataLoader(dataset,
shuffle=True,
batch_size=batch_size)
net[0].train()
trainer \
.build_criterion(LossPrinter(criterion)) \
.bind_loader('train', data_loader) \
.build_optimizer('Adam', lr=learn_rate) \
.set_max_num_epochs(num_epochs)
if torch.cuda.is_available():
trainer.cuda()
trainer.fit()
trainer.save_model(dir_name)
net[0].cpu()
net[0].eval()
def train_net_with_inferno(config_dict, net, criterion, optimizer, trainloader,
valloader):
"""
Trains the NeuralNet with inferno
:param config_dict: dict with configs
:param net: NeuralNet
:param criterion: criterion for NN
:param optimizer: optimizer for NN
:param trainloader: dataloader with traind ata
:param valloader: dataloader with validation data
"""
print("Start training with inferno!")
from inferno.trainers.basic import Trainer
from inferno.trainers.callbacks.essentials import SaveAtBestValidationScore
model_folder = os.path.join(config_dict["project_folder"], "model/")
if not os.path.exists(model_folder):
os.mkdir(model_folder)
trainer = Trainer(net) \
.save_every((1, 'epochs'), to_directory=model_folder) \
.build_criterion(criterion) \
.build_optimizer(optimizer) \
.build_metric(sorensen_dice_metric) \
.evaluate_metric_every('never') \
.validate_every((1, 'epochs'), for_num_iterations=50) \
.register_callback(SaveAtBestValidationScore(smoothness=.5))
trainer.set_max_num_epochs(config_dict['max_train_epochs'])
trainer.bind_loader('train',
trainloader).bind_loader('validate', valloader)
trainer.cuda()
trainer.fit()
def test_auto_registry(self):
callback_engine = CallbackEngine().bind_trainer(Trainer())
callback_engine.register_callback(DummyCallback())
self.assertIsInstance(next(iter(callback_engine
._callback_registry
.get('end_of_training_iteration'))),
DummyCallback)
with self.assertRaises(AssertionError):
callback_engine.register_callback(WrongDummyCallback())
def run(args):
model = SpeechModel(args)
trainer = Trainer()
if os.path.exists(
os.path.join(args.save_directory, trainer._checkpoint_filename)):
trainer.load(from_directory=args.save_directory)
model.load_state_dict(trainer.model.state_dict())
if args.cuda:
model = model.cuda()
else:
train_loader = make_loader('train', args, batch_size=args.batch_size)
dev_loader = make_loader('dev', args, batch_size=args.batch_size)
# Build trainer
trainer = Trainer(model) \
.build_criterion(CTCCriterion, size_average=True) \
.build_optimizer('Adam', weight_decay=1e-7, lr=5e-5) \
.validate_every((1, 'epochs')) \
.save_every((1, 'epochs')) \
.save_to_directory(args.save_directory) \
.set_max_num_epochs(args.epochs) \
.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
log_images_every='never'),
log_directory=args.save_directory)
# Bind loaders
trainer.bind_loader('train', train_loader, num_inputs=3, num_targets=1)
trainer.bind_loader('validate',
dev_loader,
num_inputs=3,
num_targets=1)
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
trainer.save()
test_loader = make_loader('test',
args=args,
shuffle=False,
batch_size=1,
test=True)
run_logits(loader=test_loader, model=model, args=args)
def train_model(args):
"""
Perform training then call prediction
"""
model = all_cnn_module()
model.apply(initializer)
train_loader, validate_loader, test_loader = make_loaders(args)
# Build trainer
savepath = os.path.join(args.save_directory,
Trainer()._checkpoint_filename)
if os.path.exists(savepath):
trainer = Trainer().load(from_directory=args.save_directory)
if args.cuda:
trainer.cuda()
else:
trainer = Trainer(model) \
.build_criterion('CrossEntropyLoss') \
.build_metric('CategoricalError') \
.save_every((1, 'epochs')) \
.validate_every((1, 'epochs')) \
.save_to_directory(args.save_directory) \
.set_max_num_epochs(args.epochs) \
.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
log_images_every='never'),
log_directory=args.save_directory)
# These are the params from the paper
trainer.build_optimizer('SGD',
lr=0.01,
momentum=0.9,
weight_decay=0.001)
# Also works with Adam and default settings
# trainer.build_optimizer('Adam')
trainer.bind_loader('train', train_loader)
trainer.bind_loader('validate', validate_loader)
if args.cuda:
trainer.cuda()
# Go!
trainer.fit()
trainer.save()
write_predictions(args, trainer.model, test_loader)
def __init__(self, model_path, halo, gpu=0, use_best=True, prep_model=None):
assert os.path.exists(model_path), model_path
trainer = Trainer().load(from_directory=model_path, best=use_best)
self.model = trainer.model.cuda(gpu)
self.model.eval()
if prep_model is not None:
self.model = prep_model(self.model)
self.gpu = gpu
self.halo = halo
self.lock = threading.Lock()
def test_serialization(self):
trainer = self.get_trainer(3)
# Serialize
trainer.save()
# Unserialize
trainer = Trainer().load(os.path.join(self.ROOT_DIR, 'saves'))
train_loader, test_loader = self.get_random_dataloaders(
input_channels=3)
trainer.bind_loader('train',
train_loader).bind_loader('validate', test_loader)
trainer.fit()
def load_model(args):
trainer = Trainer()
trainer.load(from_directory=args.load_directory, best=False)
# trainer.load(from_directory=args.load_directory, best=False)
trainer.set_max_num_epochs(args.epochs + trainer.epoch_count)
model = trainer.model
trainer.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
log_images_every='never'),
log_directory=args.save_directory)
trainer.save_to_directory(args.save_directory)
return (model, trainer)
def test_multi_gpu_setup(self):
from torch.nn import CrossEntropyLoss
from inferno.trainers.basic import Trainer
# Test base_device = 'cpu'
# Build model
net = self._make_test_model()
# Make dummy criterion
criterion = CrossEntropyLoss(weight=torch.rand(10))
# Make trainer
trainer = Trainer(net).build_criterion(criterion).cuda(
[0, 1], base_device='cpu')
self.assertIsInstance(trainer.criterion.weight, torch.FloatTensor)
# Test base_device = 'cpu'
# Build model
net = self._make_test_model()
criterion = CrossEntropyLoss(weight=torch.rand(10))
# Make trainer
trainer = Trainer(net).build_criterion(criterion).cuda(
[0, 1], base_device='cuda')
self.assertIsInstance(trainer.criterion.weight, torch.cuda.FloatTensor)
def set_up_training(project_directory, config, data_config):
# Get model
model_name = config.get('model_name')
model = getattr(models, model_name)(**config.get('model_kwargs'))
criterion = SorensenDiceLoss()
loss_train = LossWrapper(criterion=criterion, transforms=InvertTarget())
loss_val = LossWrapper(criterion=criterion,
transforms=Compose(RemoveSegmentationFromTarget(),
InvertTarget()))
# Build trainer and validation metric
logger.info("Building trainer.")
smoothness = 0.75
offsets = data_config['volume_config']['segmentation']['affinity_config'][
'offsets']
strides = [1, 10, 10]
metric = ArandErrorFromMWS(average_slices=False,
offsets=offsets,
strides=strides,
randomize_strides=False)
trainer = Trainer(model)\
.save_every((1000, 'iterations'),
to_directory=os.path.join(project_directory, 'Weights'))\
.build_criterion(loss_train)\
.build_validation_criterion(loss_val)\
.build_optimizer(**config.get('training_optimizer_kwargs'))\
.evaluate_metric_every('never')\
.validate_every((100, 'iterations'), for_num_iterations=1)\
.register_callback(SaveAtBestValidationScore(smoothness=smoothness,
verbose=True))\
.build_metric(metric)\
.register_callback(AutoLR(factor=0.99,
patience='100 iterations',
monitor_while='validating',
monitor_momentum=smoothness,
consider_improvement_with_respect_to='previous'))\
logger.info("Building logger.")
# Build logger
tensorboard = TensorboardLogger(
log_scalars_every=(1, 'iteration'),
log_images_every=(100, 'iterations'),
log_histograms_every='never').observe_states(
['validation_input', 'validation_prediction, validation_target'],
observe_while='validating')
trainer.build_logger(tensorboard,
log_directory=os.path.join(project_directory, 'Logs'))
return trainer
def __init__(self, model_path, crop=None, gpu=0, use_best=True):
from inferno.trainers.basic import Trainer
assert os.path.exists(model_path), model_path
trainer = Trainer().load(from_directory=model_path, best=use_best)
self.model = trainer.model.cuda(gpu)
self.gpu = gpu
# validate cropping
if crop is not None:
assert isinstance(crop, (list, tuple))
assert len(crop) == 3
self.crop = crop
self.lock = threading.Lock()
def main(args):
img_size = 224
train_trans = transforms.Compose([
transforms.RandomResizedCrop(img_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
test_trans = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(img_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
train_ds = IBMDataset(args.data_dir, transform=train_trans, name='train')
train_loader = data.DataLoader(train_ds,
batch_size = args.batch_size,
shuffle = True)
val_ds = IBMDataset(args.data_dir, transform=test_trans, name='val')
print(len(val_ds), len(train_ds))
val_loader = data.DataLoader(val_ds,
batch_size = args.batch_size,
shuffle = True)
LOG_DIRECTORY = '../logs/'
SAVE_DIRECTORY = '../models/resnet50/'
model = ResNet50()
trainer = Trainer(model) \
.build_criterion('CrossEntropyLoss') \
.build_metric('CategoricalError') \
.build_optimizer('Adam') \
.validate_every((1, 'epochs')) \
.save_every((1, 'epochs')) \
.save_to_directory(SAVE_DIRECTORY) \
.set_max_num_epochs(args.num_epochs) \
.build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),
log_images_every='never'),
log_directory=LOG_DIRECTORY)
# Bind loaders
trainer \
.bind_loader('train', train_loader) \
.bind_loader('validate', val_loader)
if torch.cuda.is_available():
trainer.cuda()
trainer.fit()
