def validate(model, epoch, val_iterator, src_vocab, tgt_vocab, args, writer):
model.eval()
losses = 0
correct_words = 0
total_words = 0
with torch.no_grad():
for batch_idx, batch in tqdm(enumerate(val_iterator),
total=len(val_iterator)):
device = args.device
src = batch.src.transpose(0, 1).to(device)
tgt = batch.tgt.transpose(0, 1).to(device)
src_mask = padding_mask(src, src_vocab)
tgt_mask = padding_mask(tgt[:, :-1], src_vocab) & subsequent_mask(
tgt[:, :-1]).to(device)
out = model(src, tgt[:, :-1], src_mask, tgt_mask)
labels = tgt[:, 1:].contiguous().view(-1)
loss, n_correct = cal_performance(out, labels, tgt_vocab)
losses += loss.item()
total_words += tgt[:, 1:].ne(
tgt_vocab.stoi[CONSTANTS['pad']]).sum().item()
correct_words += n_correct
print('(Validation) ppl: {ppl: 8.5f}, accuracy: {accu:3.3f} %'.format(
ppl=math.exp(losses / total_words),
accu=100 * correct_words / total_words))
writer.add_scalar('val_loss', losses / total_words, epoch)
def train(model, epoch, train_iterator, optimizer, src_vocab, tgt_vocab, args,
writer):
model.train()
losses = 0
correct_words = 0
total_words = 0
for batch_idx, batch in tqdm(enumerate(train_iterator),
total=len(train_iterator)):
device = args.device
src = batch.src.transpose(0, 1).to(device)
tgt = batch.tgt.transpose(0, 1).to(device)
src_mask = padding_mask(src, src_vocab)
tgt_mask = padding_mask(tgt[:, :-1], tgt_vocab) & subsequent_mask(
tgt[:, :-1]).to(device)
out = model(src, tgt[:, :-1], src_mask, tgt_mask)
optimizer.zero_grad()
labels = tgt[:, 1:].contiguous().view(-1)
loss, n_correct = cal_performance(out, labels, tgt_vocab)
loss.backward()
optimizer.step()
losses += loss.item()
total_words += tgt[:, 1:].ne(
tgt_vocab.stoi[CONSTANTS['pad']]).sum().item()
correct_words += n_correct
print('(Training) ppl: {ppl: 8.5f}, accuracy: {accu:3.3f} %'.format(
ppl=math.exp(losses / total_words),
accu=100 * correct_words / total_words))
writer.add_scalar('train_loss', losses / total_words, epoch)
return correct_words / total_words
def test(model, test_iterator, src_vocab, tgt_vocab, args, writer):
model.eval()
losses = 0
correct_words = 0
total_words = 0
references = []
hypotheses = []
with torch.no_grad():
for batch_idx, batch in tqdm(enumerate(test_iterator),
total=len(test_iterator)):
device = args.device
src = batch.src.transpose(0, 1).to(device)
tgt = batch.tgt.transpose(0, 1).to(device)
src_mask = padding_mask(src, src_vocab)
tgt_mask = padding_mask(tgt[:, :-1], src_vocab) & subsequent_mask(
tgt[:, :-1]).to(device)
out = model(src, tgt[:, :-1], src_mask, tgt_mask)
labels = tgt[:, 1:].contiguous().view(-1)
loss, n_correct = cal_performance(out, labels, tgt_vocab)
losses += loss.item()
total_words += tgt[:, 1:].ne(
tgt_vocab.stoi[CONSTANTS['pad']]).sum().item()
correct_words += n_correct
# Convert target sentence into words
for idxs in tgt.tolist():
references.append([[
idx for idx in idxs
if idx != tgt_vocab.stoi[CONSTANTS['start']]
and idx != tgt_vocab.stoi[CONSTANTS['pad']]
]])
# Convert prediction into a sentence
word_idxs = torch.max(out, dim=-1)[1]
for idxs in word_idxs.tolist():
hypotheses.append([
idx for idx in idxs
if idx != tgt_vocab.stoi[CONSTANTS['start']]
and idx != tgt_vocab.stoi[CONSTANTS['pad']]
])
bleu_1 = corpus_bleu(references, hypotheses, weights=(1, 0, 0, 0))
bleu_2 = corpus_bleu(references, hypotheses, weights=(0.5, 0.5, 0, 0))
print(
'(Test) ppl: {ppl: 8.5f}, accuracy: {accu:3.3f}%, BLEU-1: {bleu1:3.3f}, BLEU-2: {bleu2:3.3f}'
.format(ppl=math.exp(losses / total_words),
accu=100 * correct_words / total_words,
bleu1=bleu_1,
bleu2=bleu_2))
def greedy_decode(model, src, src_mask, max_len, start_symbol):
memory = model.encode(src, src_mask)
ys = torch.ones(1, 1).fill_(start_symbol).type_as(src.data)
for i in range(max_len - 1):
out = model.decode(
memory, src_mask, Variable(ys),
Variable(subsequent_mask(ys.size(1)).type_as(src.data)))
prob = model.generator(out[:, -1])
_, next_word = torch.max(prob, dim=1)
next_word = next_word.data[0]
ys = torch.cat(
[ys, torch.ones(1, 1).type_as(src.data).fill_(next_word)], dim=1)
return ys
def __init__(self, num_chars, num_words, num_classes, bos_token, **kwargs):
super().__init__()
self.transformer_size = kwargs.pop('transformer_size', 512)
self.bos_token = bos_token
with self.init_scope():
self.feature_extractor = ResNet(kwargs.pop('num_layers', 18))
self.transformer = get_conv_feature_encoder_decoder(
num_classes, N=1, model_size=self.transformer_size)
self.classifier = L.Linear(self.transformer_size, num_classes)
self.mask = subsequent_mask(self.transformer_size)
self.num_chars = num_chars
self.num_words = num_words
chainer.global_config.user_text_recognition_grayscale_input = False
def predict_word(self, dec_seq, src_seq, enc_output, n_active_inst):
src_mask = padding_mask(src_seq, self.src_vocab)
tgt_mask = padding_mask(
dec_seq, self.tgt_vocab) & subsequent_mask(dec_seq).to(self.device)
dec_seq = self.model.tgt_embedding(dec_seq) * math.sqrt(
self.model.d_model)
dec_seq = self.model.positional_encoder2(dec_seq)
dec_output = self.model.decoder(dec_seq, enc_output, src_mask,
tgt_mask)
dec_output = dec_output[:, -1, :]
word_prob = F.log_softmax(self.model.linear(dec_output), dim=1)
word_prob = word_prob.view(n_active_inst, self.beam_size, -1)
return word_prob
def __init__(self, vocab_size, max_len, start_symbol, transformer_size=512):
super().__init__()
self.vocab_size = vocab_size
self.max_len = max_len
self.start_symbol = start_symbol
self.transformer_size = transformer_size
with self.init_scope():
model = get_encoder_decoder(
vocab_size,
vocab_size,
N=2,
model_size=transformer_size,
)
self.model = model
self.mask = subsequent_mask(self.transformer_size)
self.classifier = L.Linear(transformer_size, vocab_size)
def __init__(self, *args, **kwargs):
self.transformer_size = kwargs.pop('transformer_size', 512)
super().__init__(*args, **kwargs)
with self.init_scope():
positional_encoding, decoder = build_transform_param_decoder(N=1, model_size=self.transformer_size)
self.positional_encoding = positional_encoding
self.decoder = decoder
self.param_predictor = L.Linear(self.transformer_size, 6)
params = self.param_predictor.b.data
# params[...] = 0
# params[[0, 4]] = 0.8
# self.param_predictor.W.data[...] = 0
self.param_embedder = L.Linear(6, self.transformer_size)
self.mask = subsequent_mask(self.transformer_size)
def make_std_mask(target: Tensor, pad) -> Tensor:
"""
Create a mask for `target` hiding both the padding (specified by `pad`) and the subsequent words
(prevent token at position i to attend to positions > i).
:param target: Tensor to create a mask for.
:param pad: token corresponding to padding elements.
:return: Mask hiding both padding and subsequent elements in target.
"""
# hide padding
target_mask = (target != pad).unsqueeze(-2)
# hide padding and future words
target_mask = (target_mask & subsequent_mask(target.shape[-1]).type_as(target_mask.data))
return target_mask
def test_forward(self):
# Parameters
batch_size = 64
sequence_length = 10
d_k = d_v = d_model = input_size = 512
d_ff = 2048
nb_of_decoder_layers = 6
# Initialize decoder
decoder_layer = DecoderLayer(
size=input_size,
self_attn=MultiHeadAttention(n_head=8,
d_model=d_model,
d_k=d_k,
d_v=d_v,
dropout=0.1),
memory_attn=MultiHeadAttention(n_head=8,
d_model=d_model,
d_k=d_k,
d_v=d_v,
dropout=0.1),
feed_forward=PositionwiseFeedForward(d_model=d_model,
d_ff=d_ff,
dropout=0.1),
dropout=0.1)
decoder = Decoder(layer=decoder_layer, N=nb_of_decoder_layers)
# Initialize input and memory
x = torch.ones((batch_size, sequence_length, input_size))
memory = torch.ones((batch_size, sequence_length, input_size))
# Subsequent mask: mask all words with length > i
decoder_mask = subsequent_mask(sequence_length)
# Forward pass with fool input and memory (same here)
out = decoder.forward(x, memory, decoder_mask, None)
# Unit Tests
self.assertIsInstance(out, torch.Tensor)
self.assertEqual(out.shape, x.shape)
self.assertEqual(out.shape, memory.shape)
self.assertEqual(x.shape, memory.shape)
self.assertEqual(torch.isnan(out).sum(), 0)
def forward(self, tgt_seq, tgt_pos, src_seq, enc_outputs):
slf_attn_mask = torch.gt(
padding_mask(tgt_seq, tgt_seq) + subsequent_mask(tgt_seq), 0)
inter_attn_mask = padding_mask(tgt_seq, src_seq)
embedded = self.emb(tgt_seq)
embedded += self.pos_enc(tgt_pos)
outputs, slf_attns, inter_attns = [], [], []
output = embedded
for layer, enc_output in zip(self.layers, enc_outputs):
output, slf_attn, inter_attn = layer(
output,
enc_output,
slf_attn_mask=slf_attn_mask,
inter_attn_mask=inter_attn_mask)
outputs += [output]
slf_attns += [slf_attn]
inter_attns += [inter_attn]
return outputs, slf_attns, inter_attns
def __init__(self, flist, modules, consts, options):
self.batch_size = len(flist)
self.x = np.zeros((self.batch_size, consts["len_x"]), dtype = np.int64)
self.x_ext = np.zeros((self.batch_size, consts["len_x"]), dtype = np.int64)
self.px = np.zeros((self.batch_size, consts["len_x"]), dtype = np.int64)
self.pxs = np.zeros((self.batch_size, consts["len_x"]), dtype = np.int64)
self.y = np.zeros((self.batch_size, consts["len_y"]), dtype = np.int64)
self.y_ext = np.zeros((self.batch_size, consts["len_y"]), dtype = np.int64)
self.y_inp = np.zeros((self.batch_size, consts["len_y"]), dtype = np.int64)
self.py = np.zeros((self.batch_size, consts["len_y"]), dtype = np.int64)
self.pys = np.zeros((self.batch_size, consts["len_y"]), dtype = np.int64)
self.x_mask = np.zeros((self.batch_size, 1, consts["len_x"]), dtype = np.int64)
self.y_mask = np.zeros((self.batch_size, 1, consts["len_y"]), dtype = np.int64)
self.y_mask_tri = np.zeros((self.batch_size, consts["len_y"], consts["len_y"]), dtype = np.int64)
self.len_x = []
self.len_y = []
self.original_contents = []
self.original_summarys = []
self.x_ext_words = []
self.max_ext_len = 0
w2i = modules["w2i"]
i2w = modules["i2w"]
dict_size = len(w2i)
for idx_doc in xrange(len(flist)):
if len(flist[idx_doc]) == 2:
contents, summarys = flist[idx_doc]
else:
print "ERROR!"
return
content, original_content = contents
summary, original_summary = summarys
self.original_contents.append(original_content)
self.original_summarys.append(original_summary)
xi_oovs = []
send_id = 1
num_word = 0
for idx_word in xrange(len(content)):
# some sentences in duc is longer than len_x
if idx_word == consts["len_x"]:
break
w = content[idx_word]
num_word += 1
if idx_word > 0 and content[idx_word - 1] == "." and num_word >= 10:
send_id += 1
num_word = 1
if w not in w2i: # OOV
if w not in xi_oovs:
xi_oovs.append(w)
self.x_ext[idx_doc, idx_word] = dict_size + xi_oovs.index(w) # 500005, 51000
w = i2w[modules["lfw_emb"]]
else:
self.x_ext[idx_doc, idx_word] = w2i[w]
self.x[idx_doc, idx_word] = w2i[w]
self.x_mask[idx_doc, 0, idx_word] = 1
self.px[idx_doc, idx_word] = idx_word + 1#num_word
self.pxs[idx_doc, idx_word] = send_id
self.len_x.append(np.sum(self.x_mask[idx_doc, :, :]))
self.x_ext_words.append(xi_oovs)
if self.max_ext_len < len(xi_oovs):
self.max_ext_len = len(xi_oovs)
if options["has_y"]:
send_id = 1
num_word = 0
for idx_word in xrange(len(summary)):
w = summary[idx_word]
num_word += 1
if idx_word > 0 and summary[idx_word - 1] == "." and num_word >= 10:
send_id += 1
num_word = 1
if w not in w2i:
if w in xi_oovs:
self.y_ext[idx_doc, idx_word] = dict_size + xi_oovs.index(w)
else:
self.y_ext[idx_doc, idx_word] = w2i[i2w[modules["lfw_emb"]]] # unk
w = i2w[modules["lfw_emb"]]
else:
self.y_ext[idx_doc, idx_word] = w2i[w]
self.y[idx_doc, idx_word] = w2i[w]
if (idx_word + 1) < len(summary):
self.y_inp[idx_doc, idx_word + 1] = w2i[w] # teacher forcing
self.py[idx_doc, idx_word] = idx_word #num_word # 1st:0
self.pys[idx_doc, idx_word] = send_id
if not options["is_predicting"]:
self.y_mask[idx_doc, 0, idx_word] = 1
len_summ = len(summary)
self.len_y.append(len_summ)
self.y_mask_tri[idx_doc,:len_summ, :len_summ] = subsequent_mask(len_summ)
else:
self.y = self.y_mask = self.y_mask_tri=None
max_len_x = int(np.max(self.len_x))
max_len_y = int(np.max(self.len_y))
self.x = self.x[:, 0:max_len_x]
self.x_ext = self.x_ext[:, 0:max_len_x]
self.x_mask = self.x_mask[:, :, 0:max_len_x]
self.px = self.px[:, 0:max_len_x]
self.pxs = self.pxs[:, 0:max_len_x]
self.y = self.y[:, 0:max_len_y]
self.y_ext = self.y_ext[:, 0:max_len_y]
self.y_inp = self.y_inp[:, 0:max_len_y]
self.y_mask = self.y_mask[:, :, 0:max_len_y]
self.y_mask_tri = self.y_mask_tri[:, 0:max_len_y, 0:max_len_y]
self.py = self.py[:, 0:max_len_y]
self.pys = self.pys[:, 0:max_len_y]
def greedy_decode(self,
src: torch.Tensor,
src_mask: torch.Tensor,
trg_vocab,
start_symbol="<s>",
stop_symbol="</s>",
max_length=100) -> torch.Tensor:
"""
Returns the prediction for `src` using greedy decoding for simplicity:
- Feed `src` (after embedding) in the Encoder to get the "memory",
- Feed an initial tensor (filled with start_symbol) in the Decoder, with the "memory" and the appropriate corresponding mask
- Get the predictions of the model, makes a max to get the next token, cat it to the previous prediction and iterate
:param src: sample for which to produce predictions.
:param src_mask: Associated `src` mask
:param trg_vocab: Vocabulary set of the target sentences.
:type trg_vocab: torchtext.vocab.Vocab
:param start_symbol: Symbol used as initial value for the Decoder. Should correspond to start_token="<s>" in the dataset vocab).
:param stop_symbol: Symbol used to represent an end of sentence, e.g. "</s>" (in the dataset vocab).
:param max_length: Maximum sequence length of the prediction.
"""
# 0. Ensure inference mode
self.eval()
# 1. Embed src
embedded = self.src_embeddings(src.type(LongTensor))
# 2. Encode embedded inputs
memory = self.encoder(src=embedded, mask=src_mask)
# 3. Create initial input for decoder
decoder_in = torch.ones(
src.shape[0], 1).type(FloatTensor) * trg_vocab.stoi[start_symbol]
for i in range(max_length):
# 4. Embed decoder_in
decoder_in_embed = self.trg_embeddings(decoder_in.type(LongTensor))
# 5. Go through decoder
out = self.decoder(x=decoder_in_embed,
memory=memory,
self_mask=subsequent_mask(decoder_in.shape[1]),
memory_mask=src_mask)
# 6. classifier: TODO: Why only last word?
logits = self.classifier(out[:, -1])
# 7. Get predicted token for each sample in the batch
_, next_token = logits.max(dim=1, keepdim=True)
# 8. Concatenate predicted token with previous predictions
decoder_in = torch.cat(
[decoder_in, next_token.type(FloatTensor)], dim=1)
# cast to int tensors
decoder_in = decoder_in.type(IntTensor)
# 9. retrieve words from tokens in the target vocab
translation = ""
for i in range(decoder_in.shape[1]):
sym = trg_vocab.itos[decoder_in[0, i]]
translation += sym + " "
if sym == trg_vocab.stoi[stop_symbol]:
break
# 10. return prediction
return translation
def make_std_mask(tgt, pad):
""" Create a mask to hide padding and future words. """
tgt_mask = (tgt != pad).unsqueeze(-2)
tgt_mask = tgt_mask & Variable(
subsequent_mask(tgt.size(-1)).type_as(tgt_mask.data))
return tgt_mask
