def main(args):
""" Main translation function' """
# Load arguments from checkpoint
torch.manual_seed(args.seed)
state_dict = torch.load(
args.checkpoint_path,
map_location=lambda s, l: default_restore_location(s, 'cpu'))
args_loaded = argparse.Namespace(**{
**vars(args),
**vars(state_dict['args'])
})
args_loaded.data = args.data
args = args_loaded
utils.init_logging(args)
# Load dictionaries
src_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.source_lang)))
logging.info('Loaded a source dictionary ({:s}) with {:d} words'.format(
args.source_lang, len(src_dict)))
tgt_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.target_lang)))
logging.info('Loaded a target dictionary ({:s}) with {:d} words'.format(
args.target_lang, len(tgt_dict)))
# Load dataset
test_dataset = Seq2SeqDataset(
src_file=os.path.join(args.data, 'test.{:s}'.format(args.source_lang)),
tgt_file=os.path.join(args.data, 'test.{:s}'.format(args.target_lang)),
src_dict=src_dict,
tgt_dict=tgt_dict)
test_loader = torch.utils.data.DataLoader(test_dataset,
num_workers=1,
collate_fn=test_dataset.collater,
batch_sampler=BatchSampler(
test_dataset,
9999999,
args.batch_size,
1,
0,
shuffle=False,
seed=args.seed))
# Build model and criterion
model = models.build_model(args, src_dict, tgt_dict)
if args.cuda:
model = model.cuda()
model.eval()
model.load_state_dict(state_dict['model'])
logging.info('Loaded a model from checkpoint {:s}'.format(
args.checkpoint_path))
progress_bar = tqdm(test_loader, desc='| Generation', leave=False)
# Iterate over the test set
all_hyps = {}
for i, sample in enumerate(progress_bar):
with torch.no_grad():
# Compute the encoder output
encoder_out = model.encoder(sample['src_tokens'],
sample['src_lengths'])
go_slice = \
torch.ones(sample['src_tokens'].shape[0], 1).fill_(tgt_dict.eos_idx).type_as(sample['src_tokens'])
prev_words = go_slice
next_words = None
for _ in range(args.max_len):
with torch.no_grad():
# Compute the decoder output by repeatedly feeding it the decoded sentence prefix
decoder_out, _ = model.decoder(prev_words, encoder_out)
# Suppress <UNK>s
_, next_candidates = torch.topk(decoder_out, 2, dim=-1)
best_candidates = next_candidates[:, :, 0]
backoff_candidates = next_candidates[:, :, 1]
next_words = torch.where(best_candidates == tgt_dict.unk_idx,
backoff_candidates, best_candidates)
prev_words = torch.cat([go_slice, next_words], dim=1)
# Segment into sentences
decoded_batch = next_words.numpy()
output_sentences = [
decoded_batch[row, :] for row in range(decoded_batch.shape[0])
]
assert (len(output_sentences) == len(sample['id'].data))
# Remove padding
temp = list()
for sent in output_sentences:
first_eos = np.where(sent == tgt_dict.eos_idx)[0]
if len(first_eos) > 0:
temp.append(sent[:first_eos[0]])
else:
temp.append([])
output_sentences = temp
# Convert arrays of indices into strings of words
output_sentences = [tgt_dict.string(sent) for sent in output_sentences]
# Save translations
assert (len(output_sentences) == len(sample['id'].data))
for ii, sent in enumerate(output_sentences):
all_hyps[int(sample['id'].data[ii])] = sent
# Write to file
if args.output is not None:
with open(args.output, 'w') as out_file:
for sent_id in range(len(all_hyps.keys())):
out_file.write(all_hyps[sent_id] + '\n')
def main(args):
""" Main function. Visualizes attention weight arrays as nifty heat-maps. """
#mpl.rc('font', family='VL Gothic Regular')
mpl.font_manager._rebuild()
font_path = '/usr/share/fonts/truetype/vlgothic/VL-Gothic-Regular.ttf'
fp = FontProperties(fname=font_path, size=14)
rcParams['font.family'] = fp.get_name()
torch.manual_seed(42)
state_dict = torch.load(
args.checkpoint_path,
map_location=lambda s, l: default_restore_location(s, 'cpu'))
args = argparse.Namespace(**{**vars(args), **vars(state_dict['args'])})
utils.init_logging(args)
# Load dictionaries
src_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.source_lang)))
print('Loaded a source dictionary ({:s}) with {:d} words'.format(
args.source_lang, len(src_dict)))
tgt_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.target_lang)))
print('Loaded a target dictionary ({:s}) with {:d} words'.format(
args.target_lang, len(tgt_dict)))
# Load dataset
test_dataset = Seq2SeqDataset(
src_file=os.path.join(args.data, 'test.{:s}'.format(args.source_lang)),
tgt_file=os.path.join(args.data, 'test.{:s}'.format(args.target_lang)),
src_dict=src_dict,
tgt_dict=tgt_dict)
vis_loader = torch.utils.data.DataLoader(test_dataset,
num_workers=1,
collate_fn=test_dataset.collater,
batch_sampler=BatchSampler(
test_dataset,
None,
1,
1,
0,
shuffle=False,
seed=42))
# Build model and optimization criterion
model = models.build_model(args, src_dict, tgt_dict)
if args.cuda:
model = model.cuda()
model.load_state_dict(state_dict['model'])
print('Loaded a model from checkpoint {:s}'.format(args.checkpoint_path))
# Store attention weight arrays
attn_records = list()
# Iterate over the visualization set
for i, sample in enumerate(vis_loader):
# Only visualize the first 10 sentence pairs
if i >= 35:
break
if args.cuda:
sample = utils.move_to_cuda(sample)
if len(sample) == 0:
continue
# Perform forward pass
output, attn_weights = model(sample['src_tokens'],
sample['src_lengths'],
sample['tgt_inputs'])
attn_records.append((sample, attn_weights))
# Generate heat-maps and store them at the designated location
if not os.path.exists(args.vis_dir):
os.makedirs(args.vis_dir)
for record_id, record in enumerate(attn_records):
# Unpack
sample, attn_map = record
src_ids = utils.strip_pad(sample['src_tokens'].data, tgt_dict.pad_idx)
tgt_ids = utils.strip_pad(sample['tgt_inputs'].data, tgt_dict.pad_idx)
# Convert indices into word tokens
src_str = src_dict.string(src_ids).split(' ') + ['<EOS>']
tgt_str = tgt_dict.string(tgt_ids).split(' ') + ['<EOS>']
print(src_str)
# Generate heat-maps
attn_map = attn_map.squeeze(dim=0).transpose(1, 0).detach().numpy()
attn_df = pd.DataFrame(attn_map, index=src_str, columns=tgt_str)
sns.heatmap(attn_df,
cmap='Blues',
linewidths=0.25,
vmin=0.0,
vmax=1.0,
xticklabels=True,
yticklabels=True,
fmt='.3f')
plt.yticks(rotation=0)
plot_path = os.path.join(args.vis_dir,
'sentence_{:d}.png'.format(record_id))
plt.savefig(plot_path, dpi='figure', pad_inches=1, bbox_inches='tight')
plt.clf()
print(
'Done! Visualized attention maps have been saved to the \'{:s}\' directory!'
.format(args.vis_dir))
def main(args):
""" Main training function. Trains the translation model over the course of several epochs, including dynamic
learning rate adjustment and gradient clipping. """
logging.info('Commencing training!')
torch.manual_seed(42)
utils.init_logging(args)
# Load dictionaries
src_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.source_lang)))
logging.info('Loaded a source dictionary ({:s}) with {:d} words'.format(
args.source_lang, len(src_dict)))
tgt_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.target_lang)))
logging.info('Loaded a target dictionary ({:s}) with {:d} words'.format(
args.target_lang, len(tgt_dict)))
# Load datasets
def load_data(split):
return Seq2SeqDataset(
src_file=os.path.join(args.data,
'{:s}.{:s}'.format(split, args.source_lang)),
tgt_file=os.path.join(args.data,
'{:s}.{:s}'.format(split, args.target_lang)),
src_dict=src_dict,
tgt_dict=tgt_dict)
train_dataset = load_data(
split='train') if not args.train_on_tiny else load_data(
split='tiny_train')
valid_dataset = load_data(split='valid')
# Build model and optimization criterion
model = models.build_model(args, src_dict, tgt_dict)
logging.info('Built a model with {:d} parameters'.format(
sum(p.numel() for p in model.parameters())))
criterion = nn.CrossEntropyLoss(ignore_index=src_dict.pad_idx,
reduction='sum')
if args.cuda:
model = model.cuda()
criterion = criterion.cuda()
# Instantiate optimizer and learning rate scheduler
optimizer = torch.optim.Adam(model.parameters(), args.lr)
# Load last checkpoint if one exists
state_dict = utils.load_checkpoint(args, model, optimizer) # lr_scheduler
last_epoch = state_dict['last_epoch'] if state_dict is not None else -1
# Track validation performance for early stopping
bad_epochs = 0
best_validate = float('inf')
for epoch in range(last_epoch + 1, args.max_epoch):
train_loader = \
torch.utils.data.DataLoader(train_dataset, num_workers=1, collate_fn=train_dataset.collater,
batch_sampler=BatchSampler(train_dataset, args.max_tokens, args.batch_size, 1,
0, shuffle=True, seed=42))
model.train()
stats = OrderedDict()
stats['loss'] = 0
stats['lr'] = 0
stats['num_tokens'] = 0
stats['batch_size'] = 0
stats['grad_norm'] = 0
stats['clip'] = 0
# Display progress
progress_bar = tqdm(train_loader,
desc='| Epoch {:03d}'.format(epoch),
leave=False,
disable=False)
# Iterate over the training set
for i, sample in enumerate(progress_bar):
if args.cuda:
sample = utils.move_to_cuda(sample)
if len(sample) == 0:
continue
model.train()
output, _ = model(sample['src_tokens'], sample['src_lengths'],
sample['tgt_inputs'])
loss = \
criterion(output.view(-1, output.size(-1)), sample['tgt_tokens'].view(-1)) / len(sample['src_lengths'])
loss.backward()
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(),
args.clip_norm)
optimizer.step()
optimizer.zero_grad()
# Update statistics for progress bar
total_loss, num_tokens, batch_size = loss.item(
), sample['num_tokens'], len(sample['src_tokens'])
stats['loss'] += total_loss * len(
sample['src_lengths']) / sample['num_tokens']
stats['lr'] += optimizer.param_groups[0]['lr']
stats['num_tokens'] += num_tokens / len(sample['src_tokens'])
stats['batch_size'] += batch_size
stats['grad_norm'] += grad_norm
stats['clip'] += 1 if grad_norm > args.clip_norm else 0
progress_bar.set_postfix(
{
key: '{:.4g}'.format(value / (i + 1))
for key, value in stats.items()
},
refresh=True)
logging.info('Epoch {:03d}: {}'.format(
epoch, ' | '.join(key + ' {:.4g}'.format(value / len(progress_bar))
for key, value in stats.items())))
# Calculate validation loss
valid_perplexity = validate(args, model, criterion, valid_dataset,
epoch)
model.train()
# Save checkpoints
if epoch % args.save_interval == 0:
utils.save_checkpoint(args, model, optimizer, epoch,
valid_perplexity) # lr_scheduler
# Check whether to terminate training
if valid_perplexity < best_validate:
best_validate = valid_perplexity
bad_epochs = 0
else:
bad_epochs += 1
if bad_epochs >= args.patience:
logging.info(
'No validation set improvements observed for {:d} epochs. Early stop!'
.format(args.patience))
break
def main(args):
""" Main translation function' """
# Load arguments from checkpoint
torch.manual_seed(args.seed)
state_dict = torch.load(
args.checkpoint_path,
map_location=lambda s, l: default_restore_location(s, 'cpu'))
args_loaded = argparse.Namespace(**{
**vars(args),
**vars(state_dict['args'])
})
args_loaded.data = args.data
args = args_loaded
utils.init_logging(args)
# Load dictionaries
src_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.source_lang)))
logging.info('Loaded a source dictionary ({:s}) with {:d} words'.format(
args.source_lang, len(src_dict)))
tgt_dict = Dictionary.load(
os.path.join(args.data, 'dict.{:s}'.format(args.target_lang)))
logging.info('Loaded a target dictionary ({:s}) with {:d} words'.format(
args.target_lang, len(tgt_dict)))
# Load dataset
test_dataset = Seq2SeqDataset(
src_file=os.path.join(args.data, 'test.{:s}'.format(args.source_lang)),
tgt_file=os.path.join(args.data, 'test.{:s}'.format(args.target_lang)),
src_dict=src_dict,
tgt_dict=tgt_dict)
test_loader = torch.utils.data.DataLoader(test_dataset,
num_workers=1,
collate_fn=test_dataset.collater,
batch_sampler=BatchSampler(
test_dataset,
9999999,
args.batch_size,
1,
0,
shuffle=False,
seed=args.seed))
# Build model and criterion
model = models.build_model(args, src_dict, tgt_dict)
if args.cuda:
model = model.cuda()
model.eval()
model.load_state_dict(state_dict['model'])
logging.info('Loaded a model from checkpoint {:s}'.format(
args.checkpoint_path))
progress_bar = tqdm(test_loader, desc='| Generation', leave=False)
# Iterate over the test set
all_hyps = {}
for i, sample in enumerate(progress_bar):
# Create a beam search object or every input sentence in batch
batch_size = sample['src_tokens'].shape[0]
searches = [
BeamSearch(args.beam_size, args.max_len - 1, tgt_dict.unk_idx)
for i in range(batch_size)
]
with torch.no_grad():
# Compute the encoder output
encoder_out = model.encoder(sample['src_tokens'],
sample['src_lengths'])
# __QUESTION 1: What is "go_slice" used for and what do its dimensions represent?
go_slice = \
torch.ones(sample['src_tokens'].shape[0], 1).fill_(tgt_dict.eos_idx).type_as(sample['src_tokens'])
if args.cuda:
go_slice = utils.move_to_cuda(go_slice)
# Compute the decoder output at the first time step
decoder_out, _ = model.decoder(go_slice, encoder_out)
# __QUESTION 2: Why do we keep one top candidate more than the beam size?
log_probs, next_candidates = torch.topk(torch.log(
torch.softmax(decoder_out, dim=2)),
args.beam_size + 1,
dim=-1)
# Create number of beam_size beam search nodes for every input sentence
for i in range(batch_size):
div_para = 1
div_vec = [-1, -2, -3, -4] * div_para
div_tens = torch.tensor(div_vec, dtype=torch.long)
log_probs[i, :, :] += div_tens
for j in range(args.beam_size):
best_candidate = next_candidates[i, :, j]
backoff_candidate = next_candidates[i, :, j + 1]
best_log_p = log_probs[i, :, j]
backoff_log_p = log_probs[i, :, j + 1]
next_word = torch.where(best_candidate == tgt_dict.unk_idx,
backoff_candidate, best_candidate)
log_p = torch.where(best_candidate == tgt_dict.unk_idx,
backoff_log_p, best_log_p)
log_p = log_p[-1]
# Store the encoder_out information for the current input sentence and beam
emb = encoder_out['src_embeddings'][:, i, :]
lstm_out = encoder_out['src_out'][0][:, i, :]
final_hidden = encoder_out['src_out'][1][:, i, :]
final_cell = encoder_out['src_out'][2][:, i, :]
try:
mask = encoder_out['src_mask'][i, :]
except TypeError:
mask = None
node = BeamSearchNode(searches[i], emb, lstm_out, final_hidden,
final_cell, mask,
torch.cat(
(go_slice[i], next_word)), log_p, 1)
# __QUESTION 3: Why do we add the node with a negative score?
searches[i].add(-node.eval(), node)
# Start generating further tokens until max sentence length reached
for _ in range(args.max_len - 1):
# Get the current nodes to expand
nodes = [n[1] for s in searches for n in s.get_current_beams()]
if nodes == []:
break # All beams ended in EOS
# Reconstruct prev_words, encoder_out from current beam search nodes
prev_words = torch.stack([node.sequence for node in nodes])
encoder_out["src_embeddings"] = torch.stack(
[node.emb for node in nodes], dim=1)
lstm_out = torch.stack([node.lstm_out for node in nodes], dim=1)
final_hidden = torch.stack([node.final_hidden for node in nodes],
dim=1)
final_cell = torch.stack([node.final_cell for node in nodes],
dim=1)
encoder_out["src_out"] = (lstm_out, final_hidden, final_cell)
try:
encoder_out["src_mask"] = torch.stack(
[node.mask for node in nodes], dim=0)
except TypeError:
encoder_out["src_mask"] = None
with torch.no_grad():
# Compute the decoder output by feeding it the decoded sentence prefix
decoder_out, _ = model.decoder(prev_words, encoder_out)
# see __QUESTION 2
log_probs, next_candidates = torch.topk(torch.log(
torch.softmax(decoder_out, dim=2)),
args.beam_size + 1,
dim=-1)
# Create number of beam_size next nodes for every current node
for i in range(log_probs.shape[0]):
for j in range(args.beam_size):
best_candidate = next_candidates[i, :, j]
backoff_candidate = next_candidates[i, :, j + 1]
best_log_p = log_probs[i, :, j]
backoff_log_p = log_probs[i, :, j + 1]
next_word = torch.where(best_candidate == tgt_dict.unk_idx,
backoff_candidate, best_candidate)
log_p = torch.where(best_candidate == tgt_dict.unk_idx,
backoff_log_p, best_log_p)
log_p = log_p[-1]
next_word = torch.cat((prev_words[i][1:], next_word[-1:]))
# Get parent node and beam search object for corresponding sentence
node = nodes[i]
search = node.search
# __QUESTION 4: How are "add" and "add_final" different? What would happen if we did not make this distinction?
# Store the node as final if EOS is generated
a = 0.6
if next_word[-1] == tgt_dict.eos_idx:
node = BeamSearchNode(
search, node.emb, node.lstm_out, node.final_hidden,
node.final_cell, node.mask,
torch.cat((prev_words[i][0].view([1]), next_word)),
node.logp / node.length**a, node.length)
search.add_final(-node.eval(), node)
# Add the node to current nodes for next iteration
else:
node = BeamSearchNode(
search, node.emb, node.lstm_out, node.final_hidden,
node.final_cell, node.mask,
torch.cat((prev_words[i][0].view([1]), next_word)),
node.logp + log_p, node.length + 1)
search.add(-node.eval(), node)
# __QUESTION 5: What happens internally when we prune our beams?
# How do we know we always maintain the best sequences?
for search in searches:
search.prune()
# Segment into sentences
best_sents = torch.stack(
[search.get_best()[1].sequence[1:].cpu() for search in searches])
decoded_batch = best_sents.numpy()
output_sentences = [
decoded_batch[row, :] for row in range(decoded_batch.shape[0])
]
# __QUESTION 6: What is the purpose of this for loop?
temp = list()
for sent in output_sentences:
first_eos = np.where(sent == tgt_dict.eos_idx)[0]
if len(first_eos) > 0:
temp.append(sent[:first_eos[0]])
else:
temp.append(sent)
output_sentences = temp
# Convert arrays of indices into strings of words
output_sentences = [tgt_dict.string(sent) for sent in output_sentences]
for ii, sent in enumerate(output_sentences):
all_hyps[int(sample['id'].data[ii])] = sent
# Write to file
if args.output is not None:
with open(args.output, 'w') as out_file:
for sent_id in range(len(all_hyps.keys())):
out_file.write(all_hyps[sent_id] + '\n')