class Learning(object):
def __init__(self,
model,
criterion,
optimizer,
scheduler,
metric_ftns,
device,
num_epoch,
grad_clipping,
grad_accumulation_steps,
early_stopping,
validation_frequency,
tensorboard,
checkpoint_dir,
resume_path):
self.device, device_ids = self._prepare_device(device)
# self.model = model.to(self.device)
self.start_epoch = 1
if resume_path is not None:
self._resume_checkpoint(resume_path)
if len(device_ids) > 1:
# self.model = torch.nn.DataParallel(model, device_ids=device_ids)
self.model = torch.nn.DataParallel(model)
# cudnn.benchmark = True
self.model = model.cuda()
self.criterion = criterion
self.metric_ftns = metric_ftns
self.optimizer = optimizer
self.num_epoch = num_epoch
self.scheduler = scheduler
self.grad_clipping = grad_clipping
self.grad_accumulation_steps = grad_accumulation_steps
self.early_stopping = early_stopping
self.validation_frequency =validation_frequency
self.checkpoint_dir = checkpoint_dir
self.best_epoch = 1
self.best_score = 0
self.writer = TensorboardWriter(os.path.join(checkpoint_dir, 'tensorboard'), tensorboard)
self.train_metrics = MetricTracker('loss', writer = self.writer)
self.valid_metrics = MetricTracker('loss', *[m.__name__ for m in self.metric_ftns], writer = self.writer)
def train(self, train_dataloader):
score = 0
for epoch in range(self.start_epoch, self.num_epoch+1):
print("{} epoch: \t start training....".format(epoch))
start = time.time()
train_result = self._train_epoch(epoch, train_dataloader)
train_result.update({'time': time.time()-start})
for key, value in train_result.items():
print(' {:15s}: {}'.format(str(key), value))
# if (epoch+1) % self.validation_frequency!=0:
# print("skip validation....")
# continue
# print('{} epoch: \t start validation....'.format(epoch))
# start = time.time()
# valid_result = self._valid_epoch(epoch, valid_dataloader)
# valid_result.update({'time': time.time() - start})
# for key, value in valid_result.items():
# if 'score' in key:
# score = value
# print(' {:15s}: {}'.format(str(key), value))
score+=0.001
self.post_processing(score, epoch)
if epoch - self.best_epoch > self.early_stopping:
print('WARNING: EARLY STOPPING')
break
def _train_epoch(self, epoch, data_loader):
self.model.train()
self.optimizer.zero_grad()
self.train_metrics.reset()
for idx, (data, target) in enumerate(data_loader):
data = Variable(data.cuda())
target = [ann.to(self.device) for ann in target]
output = self.model(data)
loss = self.criterion(output, target)
loss.backward()
self.writer.set_step((epoch - 1) * len(data_loader) + idx)
self.train_metrics.update('loss', loss.item())
if (idx+1) % self.grad_accumulation_steps == 0:
torch.nn.utils.clip_grad_norm_(self.model.parameters(), self.grad_clipping)
self.optimizer.step()
self.optimizer.zero_grad()
if (idx+1) % int(np.sqrt(len(data_loader))) == 0:
self.writer.add_image('input', make_grid(data.cpu(), nrow=8, normalize=True))
return self.train_metrics.result()
def _valid_epoch(self, epoch, data_loader):
self.valid_metrics.reset()
self.model.eval()
with torch.no_grad():
for idx, (data, target) in enumerate(data_loader):
data, target = data.to(self.device), target.to(self.device)
output = self.model(data)
loss = self.criterion(output, target)
self.writer.set_step((epoch - 1) * len(data_loader) + idx, 'valid')
self.valid_metrics.update('loss', loss.item())
for met in self.metric_ftns:
self.valid_metrics.update(met.__name__, met(output, target))
self.writer.add_image('input', make_grid(data.cpu(), nrow=8, normalize=True))
for name, p in self.model.named_parameters():
self.writer.add_histogram(name, p, bins='auto')
return self.valid_metrics.result()
def post_processing(self, score, epoch):
best = False
if score > self.best_score:
self.best_score = score
self.best_epoch = epoch
best = True
print("best model: {} epoch - {:.5}".format(epoch, score))
self._save_checkpoint(epoch = epoch, save_best = best)
if self.scheduler.__class__.__name__ == 'ReduceLROnPlateau':
self.scheduler.step(score)
else:
self.scheduler.step()
def _save_checkpoint(self, epoch, save_best=False):
"""
Saving checkpoints
:param epoch: current epoch number
:param save_best: if True, rename the saved checkpoint to 'model_best.pth'
"""
arch = type(self.model).__name__
state = {
'arch': arch,
'epoch': epoch,
'state_dict': self.get_state_dict(self.model),
'best_score': self.best_score
}
filename = os.path.join(self.checkpoint_dir, 'checkpoint_epoch{}.pth'.format(epoch))
torch.save(state, filename)
print("Saving checkpoint: {} ...".format(filename))
if save_best:
best_path = os.path.join(self.checkpoint_dir, 'model_best.pth')
torch.save(state, best_path)
print("Saving current best: model_best.pth ...")
@staticmethod
def get_state_dict(model):
if type(model) == torch.nn.DataParallel:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
return state_dict
def _resume_checkpoint(self, resume_path):
resume_path = str(resume_path)
print("Loading checkpoint: {} ...".format(resume_path))
checkpoint = torch.load(resume_path, map_location=lambda storage, loc: storage)
self.start_epoch = checkpoint['epoch'] + 1
self.best_epoch = checkpoint['epoch']
self.best_score = checkpoint['best_score']
self.model.load_state_dict(checkpoint['state_dict'])
print("Checkpoint loaded. Resume training from epoch {}".format(self.start_epoch))
@staticmethod
def _prepare_device(device):
n_gpu_use = len(device)
n_gpu = torch.cuda.device_count()
if n_gpu_use > 0 and n_gpu == 0:
print("Warning: There\'s no GPU available on this machine, training will be performed on CPU.")
n_gpu_use = 0
if n_gpu_use > n_gpu:
print("Warning: The number of GPU\'s configured to use is {}, but only {} are available on this machine.".format(n_gpu_use, n_gpu))
n_gpu_use = n_gpu
list_ids = device
device = torch.device('cuda:{}'.format(device[0]) if n_gpu_use > 0 else 'cpu')
return device, list_ids
class Trainer:
"""
Training pipeline
Parameters
----------
num_epochs : int
We should train the model for __ epochs
start_epoch : int
We should start training the model from __th epoch
train_loader : DataLoader
DataLoader for training data
model : nn.Module
Model
model_name : str
Name of the model
loss_function : nn.Module
Loss function (cross entropy)
optimizer : optim.Optimizer
Optimizer (Adam)
lr_decay : float
A factor in interval (0, 1) to multiply the learning rate with
dataset_name : str
Name of the dataset
word_map : Dict[str, int]
Word2id map
grad_clip : float, optional
Gradient threshold in clip gradients
print_freq : int
Print training status every __ batches
checkpoint_path : str, optional
Path to the folder to save checkpoints
checkpoint_basename : str, optional, default='checkpoint'
Basename of the checkpoint
tensorboard : bool, optional, default=False
Enable tensorboard or not?
log_dir : str, optional
Path to the folder to save logs for tensorboard
"""
def __init__(self,
num_epochs: int,
start_epoch: int,
train_loader: DataLoader,
model: nn.Module,
model_name: str,
loss_function: nn.Module,
optimizer,
lr_decay: float,
dataset_name: str,
word_map: Dict[str, int],
grad_clip=Optional[None],
print_freq: int = 100,
checkpoint_path: Optional[str] = None,
checkpoint_basename: str = 'checkpoint',
tensorboard: bool = False,
log_dir: Optional[str] = None) -> None:
self.num_epochs = num_epochs
self.start_epoch = start_epoch
self.train_loader = train_loader
self.model = model
self.model_name = model_name
self.loss_function = loss_function
self.optimizer = optimizer
self.lr_decay = lr_decay
self.dataset_name = dataset_name
self.word_map = word_map
self.print_freq = print_freq
self.grad_clip = grad_clip
self.checkpoint_path = checkpoint_path
self.checkpoint_basename = checkpoint_basename
# setup visualization writer instance
self.writer = TensorboardWriter(log_dir, tensorboard)
self.len_epoch = len(self.train_loader)
def train(self, epoch: int) -> None:
"""
Train an epoch
Parameters
----------
epoch : int
Current number of epoch
"""
self.model.train() # training mode enables dropout
batch_time = AverageMeter(
) # forward prop. + back prop. time per batch
data_time = AverageMeter() # data loading time per batch
losses = AverageMeter(tag='loss',
writer=self.writer) # cross entropy loss
accs = AverageMeter(tag='acc', writer=self.writer) # accuracies
start = time.time()
# batches
for i, batch in enumerate(self.train_loader):
data_time.update(time.time() - start)
if self.model_name in ['han']:
documents, sentences_per_document, words_per_sentence, labels = batch
documents = documents.to(
device) # (batch_size, sentence_limit, word_limit)
sentences_per_document = sentences_per_document.squeeze(1).to(
device) # (batch_size)
words_per_sentence = words_per_sentence.to(
device) # (batch_size, sentence_limit)
labels = labels.squeeze(1).to(device) # (batch_size)
# forward
scores, _, _ = self.model(
documents, sentences_per_document, words_per_sentence
) # (n_documents, n_classes), (n_documents, max_doc_len_in_batch, max_sent_len_in_batch), (n_documents, max_doc_len_in_batch)
else:
sentences, words_per_sentence, labels = batch
sentences = sentences.to(device) # (batch_size, word_limit)
words_per_sentence = words_per_sentence.squeeze(1).to(
device) # (batch_size)
labels = labels.squeeze(1).to(device) # (batch_size)
# for torchtext
# sentences = batch.text[0].to(device) # (batch_size, word_limit)
# words_per_sentence = batch.text[1].to(device) # (batch_size)
# labels = batch.label.to(device) # (batch_size)
scores = self.model(
sentences, words_per_sentence) # (batch_size, n_classes)
# calc loss
loss = self.loss_function(scores, labels) # scalar
# backward
self.optimizer.zero_grad()
loss.backward()
# clip gradients
if self.grad_clip is not None:
clip_gradient(self.optimizer, grad_clip)
# update weights
self.optimizer.step()
# find accuracy
_, predictions = scores.max(dim=1) # (n_documents)
correct_predictions = torch.eq(predictions, labels).sum().item()
accuracy = correct_predictions / labels.size(0)
# set step for tensorboard
step = (epoch - 1) * self.len_epoch + i
self.writer.set_step(step=step, mode='train')
# keep track of metrics
batch_time.update(time.time() - start)
losses.update(loss.item(), labels.size(0))
accs.update(accuracy, labels.size(0))
start = time.time()
# print training status
if i % self.print_freq == 0:
print(
'Epoch: [{0}][{1}/{2}]\t'
'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data Load Time {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Accuracy {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch,
i,
len(self.train_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accs))
def run_train(self):
start = time.time()
# epochs
for epoch in range(self.start_epoch, self.num_epochs):
# trian an epoch
self.train(epoch=epoch)
# time per epoch
epoch_time = time.time() - start
print('Epoch: [{0}] finished, time consumed: {epoch_time:.3f}'.
format(epoch, epoch_time=epoch_time))
# decay learning rate every epoch
adjust_learning_rate(self.optimizer, self.lr_decay)
# save checkpoint
if self.checkpoint_path is not None:
save_checkpoint(epoch=epoch,
model=self.model,
model_name=self.model_name,
optimizer=self.optimizer,
dataset_name=self.dataset_name,
word_map=self.word_map,
checkpoint_path=self.checkpoint_path,
checkpoint_basename=self.checkpoint_basename)
start = time.time()
class Trainer:
"""
Encoder-decoder pipeline. Tearcher Forcing is used during training and validation.
Parameters
----------
caption_model : str
Type of the caption model
epochs : int
We should train the model for __ epochs
device : torch.device
Use GPU or not
word_map : Dict[str, int]
Word2id map
rev_word_map : Dict[int, str]
Id2word map
start_epoch : int
We should start training the model from __th epoch
epochs_since_improvement : int
Number of epochs since last improvement in BLEU-4 score
best_bleu4 : float
Best BLEU-4 score until now
train_loader : DataLoader
DataLoader for training data
val_loader : DataLoader
DataLoader for validation data
encoder : nn.Module
Encoder (based on CNN)
decoder : nn.Module
Decoder (based on LSTM)
encoder_optimizer : optim.Optimizer
Optimizer for encoder (Adam) (if fine-tune)
decoder_optimizer : optim.Optimizer
Optimizer for decoder (Adam)
loss_function : nn.Module
Loss function (cross entropy)
grad_clip : float
Gradient threshold in clip gradients
tau : float
Penalty term τ for doubly stochastic attention in paper: show, attend and tell
fine_tune_encoder : bool
Fine-tune encoder or not
tensorboard : bool, optional, default=False
Enable tensorboard or not?
log_dir : str, optional
Path to the folder to save logs for tensorboard
"""
def __init__(
self,
caption_model: str,
epochs: int,
device: torch.device,
word_map: Dict[str, int],
rev_word_map: Dict[int, str],
start_epoch: int,
epochs_since_improvement: int,
best_bleu4: float,
train_loader: DataLoader,
val_loader: DataLoader,
encoder: nn.Module,
decoder: nn.Module,
encoder_optimizer: optim.Optimizer,
decoder_optimizer: optim.Optimizer,
loss_function: nn.Module,
grad_clip: float,
tau: float,
fine_tune_encoder: bool,
tensorboard: bool = False,
log_dir: Optional[str] = None
) -> None:
self.device = device # GPU / CPU
self.caption_model = caption_model
self.epochs = epochs
self.word_map = word_map
self.rev_word_map = rev_word_map
self.start_epoch = start_epoch
self.epochs_since_improvement = epochs_since_improvement
self.best_bleu4 = best_bleu4
self.train_loader = train_loader
self.val_loader = val_loader
self.encoder = encoder
self.decoder = decoder
self.encoder_optimizer = encoder_optimizer
self.decoder_optimizer = decoder_optimizer
self.loss_function = loss_function
self.tau = tau
self.grad_clip = grad_clip
self.fine_tune_encoder = fine_tune_encoder
self.print_freq = 100 # print training/validation stats every __ batches
# setup visualization writer instance
self.writer = TensorboardWriter(log_dir, tensorboard)
self.len_epoch = len(self.train_loader)
def train(self, epoch: int) -> None:
"""
Train an epoch
Parameters
----------
epoch : int
Current number of epoch
"""
self.decoder.train() # train mode (dropout and batchnorm is used)
self.encoder.train()
batch_time = AverageMeter() # forward prop. + back prop. time
data_time = AverageMeter() # data loading time
losses = AverageMeter(tag='loss', writer=self.writer) # loss (per word decoded)
top5accs = AverageMeter(tag='top5acc', writer=self.writer) # top5 accuracy
start = time.time()
# batches
for i, (imgs, caps, caplens) in enumerate(self.train_loader):
data_time.update(time.time() - start)
# Move to GPU, if available
imgs = imgs.to(self.device)
caps = caps.to(self.device)
caplens = caplens.to(self.device)
# forward encoder
imgs = self.encoder(imgs)
# forward decoder
if self.caption_model == 'att2all':
scores, caps_sorted, decode_lengths, alphas, sort_ind = self.decoder(imgs, caps, caplens)
else:
scores, caps_sorted, decode_lengths, sort_ind = self.decoder(imgs, caps, caplens)
# since we decoded starting with <start>, the targets are all words after <start>, up to <end>
targets = caps_sorted[:, 1:]
# remove timesteps that we didn't decode at, or are pads
# pack_padded_sequence is an easy trick to do this
scores = pack_padded_sequence(scores, decode_lengths, batch_first=True)[0]
targets = pack_padded_sequence(targets, decode_lengths, batch_first=True)[0]
# calc loss
loss = self.loss_function(scores, targets)
# doubly stochastic attention regularization (in paper: show, attend and tell)
if self.caption_model == 'att2all':
loss += self.tau * ((1. - alphas.sum(dim = 1)) ** 2).mean()
# clear gradient of last batch
self.decoder_optimizer.zero_grad()
if self.encoder_optimizer is not None:
self.encoder_optimizer.zero_grad()
# backward
loss.backward()
# clip gradients
if self.grad_clip is not None:
clip_gradient(self.decoder_optimizer, self.grad_clip)
if self.encoder_optimizer is not None:
clip_gradient(self.encoder_optimizer, self.grad_clip)
# update weights
self.decoder_optimizer.step()
if self.encoder_optimizer is not None:
self.encoder_optimizer.step()
# set step for tensorboard
step = (epoch - 1) * self.len_epoch + i
self.writer.set_step(step=step, mode='train')
# keep track of metrics
top5 = accuracy(scores, targets, 5)
losses.update(loss.item(), sum(decode_lengths))
top5accs.update(top5, sum(decode_lengths))
batch_time.update(time.time() - start)
start = time.time()
# print status
if i % self.print_freq == 0:
print(
'Epoch: [{0}][{1}/{2}]\t'
'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data Load Time {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch, i, len(self.train_loader),
batch_time = batch_time,
data_time = data_time,
loss = losses,
top5 = top5accs
)
)
def validate(self) -> float:
"""
Validate an epoch.
Returns
-------
bleu4 : float
BLEU-4 score
"""
self.decoder.eval() # eval mode (no dropout or batchnorm)
if self.encoder is not None:
self.encoder.eval()
batch_time = AverageMeter()
losses = AverageMeter()
top5accs = AverageMeter()
start = time.time()
ground_truth = list() # ground_truth (true captions) for calculating BLEU-4 score
prediction = list() # prediction (predicted captions)
# explicitly disable gradient calculation to avoid CUDA memory error
# solves the issue #57
with torch.no_grad():
# Batches
for i, (imgs, caps, caplens, allcaps) in enumerate(self.val_loader):
# move to device, if available
imgs = imgs.to(self.device)
caps = caps.to(self.device)
caplens = caplens.to(self.device)
# forward encoder
if self.encoder is not None:
imgs = self.encoder(imgs)
# forward decoder
if self.caption_model == 'att2all':
scores, caps_sorted, decode_lengths, alphas, sort_ind = self.decoder(imgs, caps, caplens)
else:
scores, caps_sorted, decode_lengths, sort_ind = self.decoder(imgs, caps, caplens)
# since we decoded starting with <start>, the targets are all words after <start>, up to <end>
targets = caps_sorted[:, 1:]
# remove timesteps that we didn't decode at, or are pads
# pack_padded_sequence is an easy trick to do this
scores_copy = scores.clone()
scores = pack_padded_sequence(scores, decode_lengths, batch_first = True)[0]
targets = pack_padded_sequence(targets, decode_lengths, batch_first = True)[0]
# calc loss
loss = self.loss_function(scores, targets)
# doubly stochastic attention regularization (in paper: show, attend and tell)
if self.caption_model == 'att2all':
loss += self.tau * ((1. - alphas.sum(dim = 1)) ** 2).mean()
# keep track of metrics
losses.update(loss.item(), sum(decode_lengths))
top5 = accuracy(scores, targets, 5)
top5accs.update(top5, sum(decode_lengths))
batch_time.update(time.time() - start)
start = time.time()
if i % self.print_freq == 0:
print('Validation: [{0}/{1}]\t'
'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})\t'.format(i, len(self.val_loader),
batch_time = batch_time,
loss = losses,
top5 = top5accs)
)
# store ground truth captions and predicted captions of each image
# for n images, each of them has one prediction and multiple ground truths (a, b, c...):
# prediction = [ [hyp1], [hyp2], ..., [hypn] ]
# ground_truth = [ [ [ref1a], [ref1b], [ref1c] ], ..., [ [refna], [refnb] ] ]
# ground truth
allcaps = allcaps[sort_ind] # because images were sorted in the decoder
for j in range(allcaps.shape[0]):
img_caps = allcaps[j].tolist()
img_captions = list(
map(
lambda c: [w for w in c if w not in {self.word_map['<start>'], self.word_map['<pad>']}],
img_caps
)
) # remove <start> and pads
ground_truth.append(img_captions)
# prediction
_, preds = torch.max(scores_copy, dim = 2)
preds = preds.tolist()
temp_preds = list()
for j, p in enumerate(preds):
temp_preds.append(preds[j][:decode_lengths[j]]) # remove pads
preds = temp_preds
prediction.extend(preds)
assert len(ground_truth) == len(prediction)
# calc BLEU-4 and CIDEr score
metrics = Metrics(ground_truth, prediction, self.rev_word_map)
bleu4 = metrics.belu[3] # BLEU-4
cider = metrics.cider # CIDEr
print(
'\n * LOSS - {loss.avg:.3f}, TOP-5 ACCURACY - {top5.avg:.3f}, BLEU-4 - {bleu}, CIDEr - {cider}\n'.format(
loss = losses,
top5 = top5accs,
bleu = bleu4,
cider = cider
)
)
return bleu4
def run_train(self) -> None:
# epochs
for epoch in range(self.start_epoch, self.epochs):
# decay learning rate if there is no improvement for 8 consecutive epochs
# terminate training if there is no improvement for 20 consecutive epochs
if self.epochs_since_improvement == 20:
break
if self.epochs_since_improvement > 0 and self.epochs_since_improvement % 8 == 0:
adjust_learning_rate(self.decoder_optimizer, 0.8)
if self.fine_tune_encoder:
adjust_learning_rate(self.encoder_optimizer, 0.8)
# train an epoch
self.train(epoch = epoch)
# validate an epoch
recent_bleu4 = self.validate()
# epochs num since last improvement
is_best = recent_bleu4 > self.best_bleu4
self.best_bleu4 = max(recent_bleu4, self.best_bleu4)
if not is_best:
self.epochs_since_improvement += 1
print("\nEpochs since last improvement: %d\n" % (self.epochs_since_improvement,))
else:
self.epochs_since_improvement = 0
# save checkpoint
save_checkpoint(
epoch = epoch,
epochs_since_improvement = self.epochs_since_improvement,
encoder = self.encoder,
decoder = self.decoder,
encoder_optimizer = self.encoder_optimizer,
decoder_optimizer = self.decoder_optimizer,
caption_model = self.caption_model,
bleu4 = recent_bleu4,
is_best = is_best
)
