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)
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'])
if args.cuda:
go_slice = utils.move_to_cuda(go_slice)
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.cpu().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):
# 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 = argparse.Namespace(**{**vars(args), **vars(state_dict['args'])})
utils.init_logging(args)
# Load dictionaries
src_dict = Dictionary.load(os.path.join(args.data, 'dict.{}'.format(args.source_lang)))
logging.info('Loaded a source dictionary ({}) with {} words'.format(args.source_lang, len(src_dict)))
tgt_dict = Dictionary.load(os.path.join(args.data, 'dict.{}'.format(args.target_lang)))
logging.info('Loaded a target dictionary ({}) with {} words'.format(args.target_lang, len(tgt_dict)))
# Load dataset
test_dataset = Seq2SeqDataset(
src_file=os.path.join(args.data, 'test.{}'.format(args.source_lang)),
tgt_file=os.path.join(args.data, 'test.{}'.format(args.target_lang)),
src_dict=src_dict, tgt_dict=tgt_dict)
test_loader = torch.utils.data.DataLoader(
test_dataset, num_workers=args.num_workers, collate_fn=test_dataset.collater,
batch_sampler=BatchSampler(
test_dataset, args.max_tokens, args.batch_size, args.distributed_world_size,
args.distributed_rank, shuffle=False, seed=args.seed))
# Build model and criterion
model = models.build_model(args, src_dict, tgt_dict).cuda()
model.load_state_dict(state_dict['model'])
logging.info('Loaded a model from checkpoint {}'.format(args.checkpoint_path))
translator = SequenceGenerator(
model, tgt_dict, beam_size=args.beam_size, maxlen=args.max_len, stop_early=eval(args.stop_early),
normalize_scores=eval(args.normalize_scores), len_penalty=args.len_penalty, unk_penalty=args.unk_penalty,
)
progress_bar = tqdm(test_loader, desc='| Generation', leave=False)
for i, sample in enumerate(progress_bar):
sample = utils.move_to_cuda(sample)
with torch.no_grad():
hypos = translator.generate(sample['src_tokens'], sample['src_lengths'])
for i, (sample_id, hypos) in enumerate(zip(sample['id'].data, hypos)):
src_tokens = utils.strip_pad(sample['src_tokens'].data[i, :], tgt_dict.pad_idx)
has_target = sample['tgt_tokens'] is not None
target_tokens = utils.strip_pad(sample['tgt_tokens'].data[i, :], tgt_dict.pad_idx).int().cpu() if has_target else None
src_str = src_dict.string(src_tokens, args.remove_bpe)
target_str = tgt_dict.string(target_tokens, args.remove_bpe) if has_target else ''
if not args.quiet:
print('S-{}\t{}'.format(sample_id, src_str))
if has_target:
print('T-{}\t{}'.format(sample_id, colored(target_str, 'green')))
# Process top predictions
for i, hypo in enumerate(hypos[:min(len(hypos), args.num_hypo)]):
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'].int().cpu(),
tgt_dict=tgt_dict,
remove_bpe=args.remove_bpe,
)
if not args.quiet:
print('H-{}\t{}'.format(sample_id, colored(hypo_str, 'blue')))
print('P-{}\t{}'.format(sample_id, ' '.join(map(lambda x: '{:.4f}'.format(x), hypo['positional_scores'].tolist()))))
print('A-{}\t{}'.format(sample_id, ' '.join(map(lambda x: str(x.item()), alignment))))
# Score only the top hypothesis
if has_target and i == 0:
# Convert back to tokens for evaluation with unk replacement and/or without BPE
target_tokens = tgt_dict.binarize(target_str, word_tokenize, add_if_not_exist=True)
def main(args):
""" Main function. Visualizes attention weight arrays as nifty heat-maps. """
mpl.rc('font', family='VL Gothic')
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):
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))
# Only visualize the first 10 sentence pairs
if i >= 10:
break
# 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>']
# Generate heat-maps
attn_map = attn_map.squeeze(dim=0).transpose(1,
0).cpu().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 translation function' """
# Load arguments from checkpoint
torch.manual_seed(args.seed) # sets the random seed from pytorch random number generators
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] # returns number of rows from sample['src_tokens']
searches = [BeamSearch(args.beam_size, args.max_len - 1, tgt_dict.unk_idx) for i in range(batch_size)]
# beam search with beamsize, max seq length and unkindex --> do this B times
with torch.no_grad(): # disables gradient calculation
# 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?
# encoder_out = self.encoder(src_tokens, src_lengths) decoder_out = self.decoder(tgt_inputs, encoder_out)
go_slice = \
torch.ones(sample['src_tokens'].shape[0], 1).fill_(tgt_dict.eos_idx).type_as(sample['src_tokens'])
# vector of ones of length sample['src_tokens'] rows and 1 col filled with eos_idx casted to type 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) # decoder out = decoder(tgt_inputs, 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)
# returns largest k elements (here beam_size+1) of the input torch.log(torch.softmax(decoder_out,
# dim=2) in dimension -1 + 1 is taken because the input is given in logarithmic notation
# Create number of beam_size beam search nodes for every input sentence
for i in range(batch_size):
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)
# add normalization here according to paper
lp = normalize(node.length)
score = node.eval()/lp
# Add diverse
score = diverse(score, j)
# __QUESTION 3: Why do we add the node with a negative score?
searches[i].add(-score, 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
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)
# Add length normalization
lp = normalize(node.length)
score = node.eval()/lp
# add diverse
score = diverse(score, j)
search.add_final(-score, 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)
# Add length normalization
lp = normalize(node.length)
score = node.eval()/lp
# add diverse
score = diverse(score, j)
search.add(-score, 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 1 best sentences
#best_sents = torch.stack([search.get_best()[1].sequence[1:].cpu() for search in searches])
# segment 3 best oneliner
best_sents = torch.stack([n[1].sequence[1:] for s in searches for n in s.get_best()])
# segment into n best sentences
#for s in searches:
# for n in s.get_best():
# best_sents = torch.stack([n[1].sequence[1:].cpu()])
print('n best sents', best_sents)
# concatenates a sequence of tensors, gets the one best here, so we should use the n-best (3 best) here
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] # predicts first eos token
if len(first_eos) > 0: # checks if the first eos token is not the beginning (position 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]
# here: adapt so that it takes the 3-best (aka n-best), % used for no overflow
for ii, sent in enumerate(output_sentences):
# all_hyps[int(sample['id'].data[ii])] = sent
# variant for 3-best
all_hyps[(int(sample['id'].data[int(ii / 3)]), int(ii % 3))] = sent
# Write to file (write 3 best per sentence together)
if args.output is not None:
with open(args.output, 'w') as out_file:
for sent_id in range(len(all_hyps.keys())):
# variant for 1-best
# out_file.write(all_hyps[sent_id] + '\n')
# variant for 3-best
out_file.write(all_hyps[(int(sent_id / 3), int(sent_id % 3))] + '\n')
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'])
go_slice = \
torch.ones(sample['src_tokens'].shape[0], 1).fill_(tgt_dict.eos_idx).type_as(sample['src_tokens'])
# Compute the decoder output at the first time step
decoder_out, _ = model.decoder(go_slice, encoder_out)
# __QUESTION 1: What happens here and what do 'log_probs' and 'next_candidates' contain?
decoder_out = length_normalization(
decoder_out) #applies length normalization
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):
for j in range(args.beam_size):
# __QUESTION 2: Why do we need backoff candidates?
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
# __QUESTION 3: What happens internally when we add a new beam search node?
node = BeamSearchNode(searches[i], emb, lstm_out, final_hidden,
final_cell, mask,
torch.cat(
(go_slice[i], next_word)), log_p, 1)
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 1
decoder_out = length_normalization(
decoder_out) #length normalization function
log_probs, next_candidates = torch.topk(torch.log(
torch.softmax(length_normalization(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):
# see __QUESTION 2
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: Why do we treat nodes that generated the end-of-sentence token differently?
# Store the node as final if EOS is generated
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)
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:] 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')
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 = {}
count = 0
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):
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]
# For task 3 length normalization
# To calculate the score after length normalization
lp = (math.pow( (5 + log_probs.shape[1]), args.alpha ))/math.pow( (5+1), args.alpha)
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?
# For task 3 and task 4 diversity promoting beam search
# When alpha set to 0 and gamma set to 0, the is the original code
# When alpha set to non-zero and gamma set to 0, this is for task 3
# When alpha set to 0 or non-zero and gamma non-zero, this is for task 4
searches[i].add(-(node.eval()/lp-(j+1)*args.gamma), 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)
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]
# For task 3 length normalization
# To calculate the score after length normalization
lp = (math.pow( (5 + log_probs.shape[1]), args.alpha ))/math.pow( (5+1), args.alpha)
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
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)
# For task 4 diversity promoting beam search.
# Gamma is the weight to control the influences of rank on the score.
# (j+1) is the rank for the current candidate.
search.add_final(-(node.eval()/lp-(j+1)*args.gamma), 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)
# For task 4 diversity promoting beam search.
# Gamma is the weight to control the influences of rank on the score.
# (j+1) is the rank for the current candidate.
search.add(-(node.eval()/lp-(j+1)*args.gamma), node)
# print ("loop")
# __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()
# From line 239 to line 244, the code is for task 4 diversity promoting beam search.
# To get the n-best lists
# top_n_sent = []
# for search in searches :
# top_n = search.get_top_n(args.beam_size)
# for i in range(args.beam_size) :
# top_n_sent.append(top_n[i][1].sequence[1:])
# best_top_sents = torch.stack(top_n_sent)
# Line 248, the code is for task 4 diversity promoting beam search.
# To get the n-best lists
# decoded_batch = best_top_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
# From line 270 to line 272, the code is for task 4 diversity promoting beam search.
# To get the n-best lists
# for sent in enumerate(output_sentences):
# all_hyps[int(count)] = sent
# count = count+1
# 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')
