def _retrain(
self,
experiment: Experiment,
train_generator: DatasetContainer,
valid_generator: DatasetContainer,
epochs: int,
seed: int,
callbacks: List,
disable_tensorboard: bool,
capturing_context: bool,
) -> List[Dict]:
# pylint: disable=too-many-arguments
# If Poutyne 1.7 and before, we capture poutyne print since it print some exception.
# Otherwise, we use a null context manager.
with Capturing() if capturing_context else contextlib.nullcontext():
train_res = experiment.train(
train_generator,
valid_generator=valid_generator,
epochs=epochs,
seed=seed,
callbacks=callbacks,
verbose=self.verbose,
disable_tensorboard=disable_tensorboard,
)
return train_res
def main(cfg: DictConfig) -> None:
log.info("Init of the training")
seed = cfg.setting.seed
set_seeds(seed)
# cfg.MlFlow.params["root_path"] = hydra.utils.get_original_cwd()
writer_callback = instantiate(cfg.MlFlow)
writer_callback.log_config_params(cfg)
log.info("Loading of the dataset and embedding model")
transform = transforms.Compose([
transforms.ToTensor()
])
train = DataLoader(data.CIFAR10(root="..\..\data", train=True, download=True, transform=transform), batch_size=cfg.data_loader.batch_size)
val = DataLoader(data.CIFAR10(root="..\..\data", train=False, download=True, transform=transform), batch_size=cfg.data_loader.batch_size)
network = instantiate(cfg.model)
optimizer = instantiate(cfg.optimizer, network.parameters())
loss = nn.CrossEntropyLoss()
saving_directory = os.path.join(hydra.utils.get_original_cwd(), cfg.poutyne.root_logging_directory,
writer_callback.experiment_id,
writer_callback.run_id)
monitored_metric = "val_acc"
monitor_mode = "max"
experiment = Experiment(directory=saving_directory, network=network, device=cfg.device, optimizer=optimizer,
loss_function=loss, batch_metrics=["acc"], epoch_metrics=[F1()],
logging=cfg.poutyne.logging, monitor_metric=monitored_metric, monitor_mode=monitor_mode)
log.info("Start of the training")
experiment.train(train_generator=train, valid_generator=val, epochs=cfg.trainer.num_epochs,
seed=seed, callbacks=[writer_callback])
log.info("Start of the testing of the trained model")
test_result = experiment.test(test_generator=val, seed=seed)
writer_callback.on_test_end(test_result)
def retrain(self,
dataset_container: DatasetContainer,
train_ratio: float,
batch_size: int,
epochs: int,
num_workers: int = 1,
learning_rate: float = 0.01,
callbacks: Union[List, None] = None,
seed: int = 42,
logging_path: str = "./chekpoints") -> List[Dict]:
# pylint: disable=too-many-arguments, line-too-long, too-many-locals
"""
Method to retrain the address parser model using a dataset with the same tags. We train using
`experiment <https://poutyne.org/experiment.html>`_ from `poutyne <https://poutyne.org/index.html>`_
framework. The experiment module allow us to save checkpoints ``ckpt`` (pickle format) and a log.tsv where
the best epochs can be found (the best epoch is used in test).
Args:
dataset_container (~deepparse.deepparse.dataset_container.dataset_container.DatasetContainer): The
dataset container of the data to use.
train_ratio (float): The ratio to use of the dataset for the training. The rest of the data is used for the validation
(e.g. a train ratio of 0.8 mean a 80-20 train-valid split).
batch_size (int): The size of the batch.
epochs (int): number of training epochs.
num_workers (int): Number of workers to use for the data loader (default is 1 worker).
learning_rate (float): The learning rate (LR) to use for training (default 0.01). To reduce the LR during
training, use `Poutyne learning rate scheduler callback
<https://github.com/GRAAL-Research/poutyne/blob/master/poutyne/framework/callbacks/lr_scheduler.py>`_.
callbacks (Union[List, None]): List of callbacks to use during training.
See Poutyne `callback <https://poutyne.org/callbacks.html#callback-class>`_ for more information. By default
we set no callback.
seed (int): Seed to use (by default 42).
logging_path (str): The logging path for the checkpoints. By default the path is ``./chekpoints``.
Return:
A list of dictionary with the best epoch stats (see `Experiment class
<https://poutyne.org/experiment.html#poutyne.Experiment.train>`_ for details).
Note:
We use SGD optimizer, NLL loss and accuracy as a metric, the data is shuffled and we use teacher forcing during
training (with a prob of 0.5) as in the `article <https://arxiv.org/abs/2006.16152>`_.
Note:
Due to pymagnitude, we could not train using the Magnitude embeddings, meaning it's not possible to
train using the fasttext-light model. But, since we don't update the embeddings weights, one can retrain
using the fasttext model and later on use the weights with the fasttext-light.
Example:
.. code-block:: python
address_parser = AddressParser(device=0) #on gpu device 0
data_path = 'path_to_a_pickle_dataset.p'
container = PickleDatasetContainer(data_path)
address_parser.retrain(container, 0.8, epochs=1, batch_size=128)
Using learning rate scheduler callback.
.. code-block:: python
import poutyne
address_parser = AddressParser(device=0)
data_path = 'path_to_a_pickle_dataset.p'
container = PickleDatasetContainer(data_path)
lr_scheduler = poutyne.StepLR(step_size=1, gamma=0.1) # reduce LR by a factor of 10 each epoch
address_parser.retrain(container, 0.8, epochs=5, batch_size=128, callbacks=[lr_scheduler])
See `this <https://github.com/GRAAL-Research/deepparse/blob/master/examples/fine_tuning.py>`_ for a fine
tuning example.
"""
if self.model_type == "fasttext-light":
raise ValueError(
"It's not possible to retrain a fasttext-light due to pymagnitude problem."
)
callbacks = [] if callbacks is None else callbacks
train_generator, valid_generator = self._create_training_data_generator(
dataset_container, train_ratio, batch_size, num_workers, seed=seed)
optimizer = SGD(self.model.parameters(), learning_rate)
exp = Experiment(logging_path,
self.model,
device=self.device,
optimizer=optimizer,
loss_function=nll_loss,
batch_metrics=[accuracy])
train_res = exp.train(train_generator,
valid_generator=valid_generator,
epochs=epochs,
seed=seed,
callbacks=callbacks)
return train_res
def launch(dataset, experiment_name, network, hidden_size, hidden_layers, sample_size, weight_decay, prior,\
learning_rate, lr_patience, optim_algo, epochs, batch_size, valid_size, pre_epochs, stop_early,\
gpu_device, random_seed, logging):
# Setting random seed for reproducibility
random_state = check_random_state(random_seed)
torch.manual_seed(random_seed)
# Pac-Bayes Bound parameters
delta = 0.05
C_range = torch.Tensor(np.arange(0.1, 20.0, 0.01))
# Setting GPU device
device = None
if torch.cuda.is_available() and gpu_device != -1:
torch.cuda.set_device(gpu_device)
device = torch.device('cuda:%d' % gpu_device)
print("Running on GPU %d" % gpu_device)
else:
print("Running on CPU")
# Logging
experiment_setting = dict([('experiment_name', experiment_name),
('dataset', dataset), ('network', network),
('hidden_size', hidden_size),
('hidden_layers', hidden_layers),
('sample_size', sample_size),
('epochs', epochs),
('weight_decay', weight_decay),
('prior', prior),
('learning_rate', learning_rate),
('lr_patience', lr_patience),
('optim_algo', optim_algo),
('batch_size', batch_size),
('valid_size', valid_size),
('pre_epochs', pre_epochs),
('stop_early', stop_early),
('random_seed', random_seed)])
directory_name = get_logging_dir_name(experiment_setting)
logging_path = join(RESULTS_PATH, experiment_name, dataset, directory_name)
if logging:
if not exists(logging_path): makedirs(logging_path)
with open(join(logging_path, "setting.json"), 'w') as out_file:
json.dump(experiment_setting, out_file, sort_keys=True, indent=4)
# Loading dataset
dataset_loader = DatasetLoader(random_state=random_state)
X_train, X_test, y_train, y_test = dataset_loader.load(dataset)
X_train, X_valid, y_train, y_valid = train_test_split(
X_train, y_train, test_size=valid_size, random_state=random_state)
# Experiment
batch_metrics = [accuracy]
epoch_metrics = []
save_every_epoch = False
cost_function = linear_loss
monitor_metric = 'val_loss'
valid_set_use = 'val'
callbacks = []
if network in ['pbgnet', 'pbcombinet']:
print("### Using Pac-Bayes Binary Gradient Network ###")
if prior in ['zero', 'init']:
valid_set_use = 'train'
X_train = np.vstack([X_train, X_valid])
y_train = np.vstack([y_train, y_valid])
elif prior == 'pretrain':
valid_set_use = 'pretrain'
if network == 'pbgnet':
net = PBGNet(X_train.shape[1], hidden_layers * [hidden_size],
X_train.shape[0], sample_size, delta)
else:
net = PBCombiNet(X_train.shape[1], hidden_layers * [hidden_size],
X_train.shape[0], delta)
monitor_metric = 'bound'
cost_function = net.bound
epoch_metrics.append(
MasterMetricLogger(network=net,
loss_function=linear_loss,
delta=delta,
n_examples=X_train.shape[0]))
elif network in ['pbgnet_ll', 'pbcombinet_ll']:
print(
"### Using PAC-Bayes Gradient Network Architecture and Optimizing Linear Loss ###"
)
if network == 'pbgnet_ll':
net = PBGNet(X_train.shape[1], hidden_layers * [hidden_size],
X_train.shape[0], sample_size, delta)
else:
net = PBCombiNet(X_train.shape[1], hidden_layers * [hidden_size],
X_train.shape[0], delta)
epoch_metrics.append(
MasterMetricLogger(network=net,
loss_function=linear_loss,
delta=delta,
n_examples=X_train.shape[0],
C_range=C_range.to(device)))
callbacks.append(
ModelCheckpoint(join(logging_path, 'bound_checkpoint_epoch.ckpt'),
temporary_filename=join(
logging_path,
'bound_checkpoint_epoch.tmp.ckpt'),
monitor='bound',
mode='min',
save_best_only=True))
elif network == "baseline":
print("### Running the Baseline Network with Tanh activations ###")
net = BaselineNet(X_train.shape[1], hidden_layers * [hidden_size],
torch.nn.Tanh)
if network.startswith('pb'):
epoch_metrics.append(MetricLogger(network=net, key='bound'))
epoch_metrics.append(MetricLogger(network=net, key='kl'))
epoch_metrics.append(MetricLogger(network=net, key='C'))
# Parameters initialization
if prior in ['zero', 'init']:
net.init_weights()
elif prior == 'pretrain':
print("### Pre-training network ###")
if network == 'pbgnet':
pre_net = PBGNet(X_valid.shape[1], hidden_layers * [hidden_size],
X_valid.shape[0], sample_size, delta)
else:
pre_net = PBCombiNet(X_valid.shape[1],
hidden_layers * [hidden_size],
X_valid.shape[0], delta)
pre_net.init_weights()
pre_optimizer = torch.optim.Adam(pre_net.parameters(),
lr=learning_rate,
weight_decay=0.0)
pre_logging_path = join(logging_path, 'pretrain')
if not exists(pre_logging_path): makedirs(pre_logging_path)
pretrain = Experiment(directory=pre_logging_path,
network=pre_net,
optimizer=pre_optimizer,
loss_function=linear_loss,
monitor_metric='loss',
device=device,
logging=logging,
batch_metrics=[accuracy])
pretrain_loader = DataLoader(TensorDataset(torch.Tensor(X_valid),
torch.Tensor(y_valid)),
batch_size,
shuffle=True)
pretrain.train(train_generator=pretrain_loader,
valid_generator=None,
epochs=pre_epochs,
save_every_epoch=False,
disable_tensorboard=True,
seed=random_seed)
history = pd.read_csv(pretrain.log_filename, sep='\t')
best_epoch_index = history['loss'].idxmin()
best_epoch_stats = history.iloc[best_epoch_index:best_epoch_index + 1]
best_epoch = best_epoch_stats['epoch'].item()
ckpt_filename = pretrain.best_checkpoint_filename.format(
epoch=best_epoch)
weights = torch.load(ckpt_filename, map_location='cpu')
net.load_state_dict(weights, strict=False)
print("### Training ###")
# Setting prior
if network.startswith('pb') and prior in ['init', 'pretrain']:
net.set_priors(net.state_dict())
# Adding early stopping and lr scheduler
reduce_lr = ReduceLROnPlateau(monitor=monitor_metric, mode='min', patience=lr_patience, factor=0.5, \
threshold_mode='abs', threshold=1e-4, verbose=True)
lr_schedulers = [reduce_lr]
early_stopping = EarlyStopping(monitor=monitor_metric,
mode='min',
min_delta=1e-4,
patience=stop_early,
verbose=True)
if stop_early > 0:
callbacks.append(early_stopping)
# Initializing optimizer
if optim_algo == "sgd":
optimizer = torch.optim.SGD(net.parameters(),
lr=learning_rate,
momentum=0.9,
weight_decay=weight_decay)
elif optim_algo == "adam":
optimizer = torch.optim.Adam(net.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
# Creating Poutyne experiment
expt = Experiment(directory=logging_path,
network=net,
optimizer=optimizer,
loss_function=cost_function,
monitor_metric=monitor_metric,
device=device,
logging=logging,
batch_metrics=batch_metrics,
epoch_metrics=epoch_metrics)
# Initializing data loaders
train_loader = DataLoader(TensorDataset(torch.Tensor(X_train),
torch.Tensor(y_train)),
batch_size,
shuffle=True)
valid_loader = None
if valid_set_use == 'val':
valid_loader = DataLoader(
TensorDataset(torch.Tensor(X_valid), torch.Tensor(y_valid)),
batch_size)
# Launching training
expt.train(train_generator=train_loader,
valid_generator=valid_loader,
epochs=epochs,
callbacks=callbacks,
lr_schedulers=lr_schedulers,
save_every_epoch=save_every_epoch,
disable_tensorboard=True,
seed=random_seed)
print("### Testing ###")
sign_act_fct = lambda: Lambda(lambda x: torch.sign(x))
test_loader = DataLoader(
TensorDataset(torch.Tensor(X_test), torch.Tensor(y_test)), batch_size)
if network == 'baseline':
expt.test(test_generator=test_loader,
checkpoint='best',
seed=random_seed)
# Binary network testing (sign activation)
best_epoch = expt.get_best_epoch_stats()['epoch'].item()
ckpt_filename = expt.best_checkpoint_filename.format(epoch=best_epoch)
binary_net = BaselineNet(X_test.shape[1],
hidden_layers * [hidden_size], sign_act_fct)
weights = torch.load(ckpt_filename, map_location='cpu')
binary_net.load_state_dict(weights, strict=False)
binary_model = Model(binary_net,
'sgd',
linear_loss,
batch_metrics=[accuracy])
test_loss, test_accuracy = binary_model.evaluate_generator(test_loader,
steps=None)
test_stats = pd.read_csv(expt.test_log_filename.format(name='test'),
sep='\t')
test_stats['bin_test_linear_loss'] = test_loss
test_stats['bin_test_accuracy'] = test_accuracy
test_stats['linear_loss'] = test_stats['loss']
test_stats['val_linear_loss'] = test_stats['val_loss']
test_stats['test_linear_loss'] = test_stats['test_loss']
test_stats.to_csv(expt.test_log_filename.format(name='test'),
sep='\t',
index=False)
def pbgnet_testing(target_metric, irrelevant_columns, n_repetitions=20):
print(f"Restoring best model according to {target_metric}")
# Cleaning logs
history = pd.read_csv(expt.log_filename,
sep='\t').drop(irrelevant_columns,
axis=1,
errors='ignore')
history.to_csv(expt.log_filename, sep='\t', index=False)
# Loading best weights
best_epoch_index = history[target_metric].idxmin()
best_epoch_stats = history.iloc[best_epoch_index:best_epoch_index +
1].reset_index(drop=True)
best_epoch = best_epoch_stats['epoch'].item()
print(f"Found best checkpoint at epoch: {best_epoch}")
ckpt_filename = expt.best_checkpoint_filename.format(epoch=best_epoch)
if network in ['pbgnet_ll', 'pbcombinet_ll'
] and target_metric == 'bound':
ckpt_filename = join(logging_path, 'bound_checkpoint_epoch.ckpt')
weights = torch.load(ckpt_filename, map_location='cpu')
# Binary network testing (sign activation)
binary_net = BaselineNet(X_test.shape[1],
hidden_layers * [hidden_size], sign_act_fct)
updated_weights = {}
for name, weight in weights.items():
if name.startswith('layers'):
name = name.split('.', 2)
name[1] = str(2 * int(name[1]))
name = '.'.join(name)
updated_weights[name] = weight
binary_net.load_state_dict(updated_weights, strict=False)
binary_model = Model(binary_net,
'sgd',
linear_loss,
batch_metrics=[accuracy])
test_loss, test_accuracy = binary_model.evaluate_generator(test_loader,
steps=None)
best_epoch_stats['bin_test_linear_loss'] = test_loss
best_epoch_stats['bin_test_accuracy'] = test_accuracy
model = expt.model
model.load_weights(ckpt_filename)
def repeat_inference(loader, prefix='', drop_keys=[], n_times=20):
metrics_names = [prefix + 'loss'] + [
prefix + metric_name for metric_name in model.metrics_names
]
metrics_list = []
for _ in range(n_times):
loss, metrics = model.evaluate_generator(loader, steps=None)
if not isinstance(metrics, np.ndarray):
metrics = np.array([metrics])
metrics_list.append(np.concatenate(([loss], metrics)))
metrics_list = [list(e) for e in zip(*metrics_list)]
metrics_stats = pd.DataFrame(
{col: val
for col, val in zip(metrics_names, metrics_list)})
return metrics_stats.drop(drop_keys, axis=1, errors='ignore')
metrics_stats = repeat_inference(train_loader, n_times=n_repetitions)
metrics_stats = metrics_stats.join(
repeat_inference(test_loader,
prefix='test_',
drop_keys=['test_bound', 'test_kl', 'test_C'],
n_times=n_repetitions))
best_epoch_stats = best_epoch_stats.drop(metrics_stats.keys().tolist(),
axis=1,
errors='ignore')
metrics_stats = metrics_stats.join(
pd.concat([best_epoch_stats] * n_repetitions, ignore_index=True))
log_filename = expt.test_log_filename.format(name='test')
if network in ['pbgnet_ll', 'pbcombinet_ll'
] and target_metric == 'bound':
log_filename = join(logging_path, 'bound_test_log.tsv')
metrics_stats.to_csv(log_filename, sep='\t', index=False)
default_irrelevant_columns = ['val_bound', 'val_kl', 'val_C']
if network == 'pbgnet_ll':
pbgnet_testing(target_metric='val_loss',
irrelevant_columns=default_irrelevant_columns,
n_repetitions=20)
pbgnet_testing(target_metric='bound',
irrelevant_columns=default_irrelevant_columns,
n_repetitions=20)
elif network == 'pbgnet':
pbgnet_testing(
target_metric='bound',
irrelevant_columns=['val_loss', 'val_accuracy', 'val_linear_loss'
] + default_irrelevant_columns,
n_repetitions=20)
elif network == 'pbcombinet_ll':
pbgnet_testing(target_metric='val_loss',
irrelevant_columns=default_irrelevant_columns,
n_repetitions=1)
pbgnet_testing(target_metric='bound',
irrelevant_columns=default_irrelevant_columns,
n_repetitions=1)
elif network == 'pbcombinet':
pbgnet_testing(
target_metric='bound',
irrelevant_columns=['val_loss', 'val_accuracy', 'val_linear_loss'
] + default_irrelevant_columns,
n_repetitions=1)
if logging:
with open(join(logging_path, 'done.txt'), 'w') as done_file:
done_file.write("done")
print("### DONE ###")
def main():
vec_model = KeyedVectors.load_word2vec_format(config.pretrained_embs[0],
limit=500000)
print("Breakpoint 1")
x = load_anto_syn_graph(config.synonyms_graph[0],
config.antonyms_graph[0],
vec_model,
neg_sample=config.nb_false)
weight = compute_weight(x)
print("Breakpoint 2")
train_generator, valid_generator, test_generator = prepare_generator_graph(
x)
print("Breakpoint 3")
device = torch.device(
'cuda:%d' % config.device if torch.cuda.is_available() else 'cpu')
network = Retrofit(vec_model, weight)
embeddings_param_set = set(network.embedding.parameters())
other_params_list = [
p for p in network.parameters() if p not in embeddings_param_set
]
optimizer = optim.SGD([{
'params': other_params_list,
**config.optimizer_other_params
}, {
'params': network.embedding.parameters(),
**config.optimizer_embeddings_params
}])
#scheduler = LambdaLR(lr_lambda=[lambda_lr_other, lambda_lr_embedding])
#scheduler = StepLR(step_size=8, gamma=0.1)
scheduler = ReduceLROnPlateau(monitor='val_loss',
mode='min',
patience=2,
verbose=True)
callbacks = [scheduler]
exp = Experiment(config.dir_experiment,
network,
device=device,
optimizer=optimizer,
loss_function=None,
batch_metrics=['acc'])
exp.train(train_generator,
valid_generator,
epochs=config.epoch,
lr_schedulers=callbacks)
exp.test(test_generator)
steps = len(test_generator)
test_loss, test_metrics, pred_y, true_y = exp.model.evaluate_generator(
test_generator,
return_pred=True,
return_ground_truth=True,
steps=steps)
pred_y = np.argmax(np.concatenate(pred_y), 1)
true_y = np.concatenate(true_y)
true_syn, false_syn, false_anto, true_anto = confusion_matrix(
true_y, pred_y).ravel()
print(true_syn, false_syn, false_anto, true_anto)
learning_visualizer = LearningVisualizer(exp, config.epoch)
learning_visualizer.visualize_learning()
exp._load_best_checkpoint()
for file in config.evaluations_file:
print(evaluation(file, vec_model.vocab, exp.model.model.embedding))
vec_model_initial = KeyedVectors.load_word2vec_format(
config.pretrained_embs[0], limit=500000)
original_weights = torch.FloatTensor(vec_model_initial.vectors)
original_weights.to("cuda")
original_embs = nn.Embedding.from_pretrained(original_weights)
original_embs.cuda()
original_embs.weight.requires_grad = False
for file in config.evaluations_file:
print(evaluation(file, vec_model.vocab, original_embs))
class IterateLossFunctionOnEpoch(Callback):
def __init__(self, epoch_number):
self.epoch_number = epoch_number
super().__init__()
def on_epoch_end(self, epoch_number, logs):
if epoch_number % self.epoch_number == 0:
logs["loss"] = self.model.network.loss
self.model.network.switch_loss_type()
self.model.loss_function = self.model.network.loss_function
experiment = Experiment("model_weights/hc_{}_{}_{}".format(aggregation, dataset, test_name), lstm, optimizer=optimizer, device=0,loss_function=loss_functon, monitor_metric="val_fscore_macro", monitor_mode="max", epoch_metrics=[TopLevelAccuracy(), FBetaTopLevel(average='macro')])
# monitor_metric="top_level_accuracy", monitor_mode="max"
experiment.train(train_loader, valid_loader, lr_schedulers=[ReduceLROnPlateau(patience=patience, cooldown=cooldown)],epochs=epoch)
# BestModelRestoreOnEpoch(epoch=50,monitor="val_fscore_macro", mode="max" )
import tqdm
import json
import spacy
nlp = spacy.load('en_core_web_sm')
# for index, paragraph in enumerate(test_pickled):
# current_test_ids = ["{}-{}-{}-{}".format(doc_id, par_id, sentence_id) for doc_id, par_id, sentence_id, _, _ in paragraph]
# current_test_context = [text for _, _, _, text, _ in test_pickled]
# current_preprocessed_test = [datapoint for _, _, _, _, datapoint in test_pickled]
def run(args):
# Logging
if args.verbose:
logging.basicConfig(level=logging.DEBUG)
# Initialization
random_seed = 42
np.random.seed(random_seed)
torch.manual_seed(random_seed)
# Fix bug in PyTorch where memory is still allocated on GPU0 when
# asked to allocate memory on GPU1.
if torch.cuda.is_available():
torch.cuda.set_device(args.device)
device = torch.device('cuda:%d' % args.device if torch.cuda.is_available() else 'cpu')
# Building dataset
dataset = TextDatasetBuilder(name=args.dataset,
word_vectors=args.words,
vector_size=args.vector,
random_state=random_seed)
logging.debug("Dataset built.")
dataset.pre_process(min_freq=args.freq, max_len=args.len)
embeddings = dataset.build_embeddings()
logging.debug("Vocab size {}".format(len(dataset.vocab)))
pos_enc_len = None
if args.pos:
pos_enc_len = args.len
traind, validd, testd = dataset.get_train_valid_test()
logging.debug("Split: train = {}, valid = {} and test = {}".format(len(traind), len(validd), len(testd)))
# Creating Data Loaders
train_loader = DataLoader(traind,
batch_size=args.batch,
shuffle=True,
collate_fn=collate_padding)
valid_loader = DataLoader(validd,
batch_size=args.batch,
shuffle=False,
collate_fn=collate_padding)
test_loader = DataLoader(testd,
batch_size=args.batch,
shuffle=False,
collate_fn=collate_padding)
model = SANet(input_size=args.vector,
hidden_size=args.hidden,
n_classes=len(dataset.classes),
embeddings=embeddings,
n_blocks=args.blocks,
pos_enc_len=pos_enc_len)
init_model(model)
params = [p for n, p in model.named_parameters() if n != 'word_embedding.weight']
optimizer = optim.SGD([{'params': model.word_embedding.parameters(), 'lr': args.lr * 0.1},
{'params': params, 'lr':args.lr, 'momentum':args.momentum}])
# Preparing results output
expt_path = join(RESULTS_PATH, args.dataset, args.exp)
expt = Experiment(expt_path, model, device=device, logging=True, optimizer=optimizer, task='classifier')
reduce_lr = ReduceLROnPlateau(monitor='loss',
mode='min',
patience=2,
factor=0.5,
threshold_mode='abs',
threshold=1e-3,
verbose=True)
expt.train(train_loader, valid_loader,
epochs=args.epochs,
lr_schedulers=[reduce_lr])
expt.test(test_loader)
print("### DONE ###")
def main(args):
raw_dataset = RegressionDatasetFolder(os.path.join(
args.root_dir, 'Images/1024_with_jedi'),
input_only_transform=None,
transform=Compose([ToTensor()]))
mean, std = compute_mean_std(raw_dataset)
print(mean)
print(std)
pos_weights = compute_pos_weight(raw_dataset)
print(pos_weights)
test_dataset = RegressionDatasetFolder(
os.path.join(args.root_dir, 'Images/1024_with_jedi'),
input_only_transform=Compose([Normalize(mean, std)]),
transform=Compose(
[Lambda(lambda img: pad_resize(img, 1024, 1024)),
ToTensor()]),
in_memory=True)
valid_dataset = RegressionDatasetFolder(
os.path.join(args.root_dir, 'Images/1024_with_jedi'),
input_only_transform=Compose([Normalize(mean, std)]),
transform=Compose([ToTensor()]),
include_fname=True)
train_split, valid_split, test_split, train_weights = get_splits(
valid_dataset)
valid_loader = DataLoader(Subset(test_dataset, valid_split),
batch_size=8,
num_workers=8,
pin_memory=False)
# module = deeplabv3_efficientnet(n=5)
module = fcn_resnet50(dropout=0.8)
# module = deeplabv3_resnet50()
optim = torch.optim.Adam(module.parameters(), lr=5e-4, weight_decay=2e-3)
exp = Experiment(directory=os.path.join(args.root_dir, 'moar'),
module=module,
device=torch.device(args.device),
optimizer=optim,
loss_function=LovaszSoftmax(),
metrics=[miou, PixelWiseF1(None)],
monitor_metric='val_miou',
monitor_mode='max')
lr_schedulers = [
ReduceLROnPlateau(monitor='val_miou',
mode='max',
factor=0.2,
patience=3,
threshold=1e-1,
threshold_mode='abs')
]
callbacks = [
EarlyStopping(monitor='val_miou',
min_delta=1e-1,
patience=8,
verbose=True,
mode='max')
]
for i, (crop_size, batch_size) in enumerate(zip([512], [5])):
train_loader = get_loader_for_crop_batch(crop_size, batch_size,
train_split, mean, std,
train_weights, args.root_dir)
exp.train(train_loader=train_loader,
valid_loader=valid_loader,
epochs=(1 + i) * 30,
lr_schedulers=lr_schedulers,
callbacks=callbacks)
raw_dataset.print_filenames()
pure_dataset = RegressionDatasetFolder(os.path.join(
args.root_dir, 'Images/1024_with_jedi'),
transform=Compose([ToTensor()]),
include_fname=True)
test_loader = DataLoader(Subset(test_dataset, test_split),
batch_size=8,
num_workers=8,
pin_memory=False)
valid_loader = DataLoader(valid_dataset,
batch_size=1,
num_workers=8,
pin_memory=False)
pure_loader = DataLoader(pure_dataset,
batch_size=1,
num_workers=8,
pin_memory=False)
exp.test(test_loader)
# for checkpoint in [11, 15, 16, 17, 21]:
# print("Testing checkpoint {}".format(checkpoint))
# exp.load_checkpoint(checkpoint)
# test_model_on_checkpoint(exp.model, test_loader)
exp.load_checkpoint(11)
module = exp.model.model
module.eval()
generate_output_folders(args.root_dir)
splits = [(train_split, 'train'), (valid_split, 'valid'),
(test_split, 'test')]
results_csv = [[
'Name', 'Type', 'Split', 'iou_nothing', 'iou_bark', 'iou_node',
'iou_mean', 'f1_nothing', 'f1_bark', 'f1_node', 'f1_mean',
'Output Bark %', 'Output Node %', 'Target Bark %', 'Target Node %'
]]
with torch.no_grad():
for image_number, (batch, pure_batch) in enumerate(
zip(valid_loader, pure_loader)):
input = pure_batch[0]
target = pure_batch[1]
fname = pure_batch[2][0]
wood_type = pure_batch[3][0]
del pure_batch
outputs = module(batch[0].to(torch.device(args.device)))
outputs = remove_small_zones(outputs)
del batch
names = ['Input', 'Target', 'Generated image']
try:
class_accs = iou(outputs, target.to(torch.device(args.device)))
f1s = PixelWiseF1('all')(outputs, target) * 100
acc = class_accs.mean()
f1 = f1s.mean()
except ValueError as e:
print('Error on file {}'.format(fname))
print(outputs.shape)
print(target.shape)
raise e
outputs = torch.argmax(outputs, dim=1)
imgs = [input, target, outputs]
imgs = [img.detach().cpu().squeeze().numpy() for img in imgs]
fig, axs = plt.subplots(1, 3)
class_names = ['Nothing', 'Bark', 'Node']
for i, ax in enumerate(axs.flatten()):
img = imgs[i]
raw = (len(img.shape) == 3)
if raw: # Raw input
img = img.transpose(1, 2, 0)
values = np.unique(img.ravel())
plotted_img = ax.imshow(img, vmax=2)
ax.set_title(names[i])
ax.axis('off')
if not raw: # Predicted image
patches = [
mpatches.Patch(
color=plotted_img.cmap(plotted_img.norm(value)),
label='{} zone'.format(class_names[value]))
for value in values
]
suptitle = 'Mean iou : {:.3f}\n'.format(acc)
for split_idxs, split_name in splits:
if image_number in split_idxs:
split = split_name
running_csv_stats = [fname, wood_type, split]
class_names = ['Nothing', 'Bark', 'Node']
for c, c_acc in zip(class_names, class_accs):
suptitle += '{} : {:.3f}; '.format('iou_' + c, c_acc)
running_csv_stats.append('{:.3f}'.format(c_acc))
running_csv_stats.append('{:.3f}'.format(acc))
suptitle += '\nMean f1 : {:.3f}\n'.format(f1)
for c, c_f1 in zip(class_names, f1s):
suptitle += '{} : {:.3f}; '.format('f1_' + c, c_f1)
running_csv_stats.append('{:.3f}'.format(c_f1))
running_csv_stats.append('{:.3f}'.format(f1))
for class_idx in [1, 2]:
class_percent = (outputs == class_idx).float().mean().cpu()
running_csv_stats.append('{:.5f}'.format(class_percent * 100))
for class_idx in [1, 2]:
class_percent = (target == class_idx).float().mean().cpu()
running_csv_stats.append('{:.5f}'.format(class_percent * 100))
fig.legend(handles=patches,
title='Classes',
bbox_to_anchor=(0.4, -0.2, 0.5, 0.5))
plt.suptitle(suptitle)
plt.tight_layout()
# plt.show()
plt.savefig(os.path.join(
args.root_dir,
'Images/results/moar/combined_images/{}/{}/{}').format(
wood_type, split, fname),
format='png',
dpi=900)
plt.close()
outputs = outputs.squeeze().cpu().numpy()
dual_outputs = np.zeros((outputs.shape[0], outputs.shape[1]),
dtype=np.uint8)
dual_outputs[outputs == 1] = 127
dual_outputs[outputs == 2] = 255
dual = Image.fromarray(dual_outputs, mode='L')
dual.save(
os.path.join(args.root_dir,
'Images/results/moar/outputs/{}/{}/{}').format(
wood_type, split, fname))
results_csv.append(running_csv_stats)
csv_file = os.path.join(args.root_dir, 'Images', 'results', 'moar',
'final_stats.csv')
with open(csv_file, 'w') as f:
csv_writer = csv.writer(f, delimiter='\t')
csv_writer.writerows(results_csv)
from poutyne.framework import Experiment
# Instead of `task`, you can provide your own loss function and metrics.
expt = Experiment('my_directory', network, task='classifier', optimizer='sgd')
expt.train(train_loader,
valid_loader,
epochs=epochs,
callbacks=callbacks,
seed=42)
expt.test(test_loader)