class BiLSTMModel(nn.Module):
def __init__(self):
super(BiLSTMModel, self).__init__()
self.base_rnn = None
self.wd = None
def init_model(self,
wd,
hidden_size,
e_layers,
d_layers,
base_rnn,
pretrained_embeddings=None,
dropout_p=0.1):
self.base_rnn = base_rnn
self.wd = wd
self.dropout_p = dropout_p
if pretrained_embeddings is True:
print("Loading GloVe Embeddings ...")
pretrained_embeddings = load_glove_embeddings(
wd.word2index, hidden_size)
self.encoder = EncoderRNN(wd.n_words,
hidden_size,
n_layers=e_layers,
base_rnn=base_rnn,
pretrained_embeddings=pretrained_embeddings)
self.mlp = torch.nn.Sequential(
torch.nn.Linear(int(hidden_size * 8), int(hidden_size)),
torch.nn.ReLU(), torch.nn.Dropout(dropout_p),
torch.nn.Linear(int(hidden_size), 3), torch.nn.Softmax(dim=1))
self.parameter_list = [
self.encoder.parameters(),
self.mlp.parameters()
]
if USE_CUDA:
self.encoder = self.encoder.cuda()
self.mlp = self.mlp.cuda()
return self
def forward(self, batch, inference=False):
# Convert batch from numpy to torch
if inference is True:
text_batch, text_lengths, hyp_batch, hyp_lengths = batch
else:
text_batch, text_lengths, hyp_batch, hyp_lengths, labels = batch
batch_size = text_batch.size(1)
# Pass the input batch through the encoder
text_enc_fwd_outputs, text_enc_bkwd_outputs, text_encoder_hidden = self.encoder(
text_batch, text_lengths)
hyp_enc_fwd_outputs, hyp_enc_bkwd_outputs, hyp_encoder_hidden = self.encoder(
hyp_batch, hyp_lengths)
last_text_enc_fwd = text_enc_fwd_outputs[-1, :, :]
last_text_enc_bkwd = text_enc_bkwd_outputs[0, :, :]
last_text_enc = torch.cat((last_text_enc_fwd, last_text_enc_bkwd),
dim=1)
last_hyp_enc_fwd = hyp_enc_fwd_outputs[-1, :, :]
last_hyp_enc_bkwd = hyp_enc_bkwd_outputs[0, :, :]
last_hyp_enc = torch.cat((last_hyp_enc_fwd, last_hyp_enc_bkwd), dim=1)
mult_feature, diff_feature = last_text_enc * last_hyp_enc, torch.abs(
last_text_enc - last_hyp_enc)
features = torch.cat(
[last_text_enc, last_hyp_enc, mult_feature, diff_feature], dim=1)
outputs = self.mlp(features) # B x 3
return outputs
def get_loss_for_batch(self, batch):
labels = batch[-1]
outputs = self(batch)
loss_fn = torch.nn.CrossEntropyLoss()
loss = loss_fn(outputs, labels)
return loss
def torch_batch_from_numpy_batch(self, batch):
batch = list(batch)
variable_indices = [
0, 2, 4
] # tuple indices of variables need to be converted
for i in variable_indices:
var = Variable(torch.from_numpy(batch[i]))
if USE_CUDA:
var = var.cuda()
batch[i] = var
return batch
# Trains on a single batch
def train_batch(self, batch, tl_mode=False):
self.train()
batch = self.torch_batch_from_numpy_batch(batch)
loss = self.get_loss_for_batch(batch)
loss.backward()
return loss.item() #loss.data[0]
def validate(self, batch):
self.eval()
batch = self.torch_batch_from_numpy_batch(batch)
return self.get_loss_for_batch(batch).item() #.data[0]
def score(self, data):
batch_size = 1
batches = nli_batches(batch_size, data)
total_correct = 0
for batch in tqdm(batches):
batch = self.torch_batch_from_numpy_batch(batch)
num_correct = self._acc_for_batch(batch)
total_correct += num_correct
acc = total_correct / (len(batches) * batch_size)
return acc
def _acc_for_batch(self, batch):
'''
:param batch:
:return: The number of correct predictions in a batch
'''
self.eval()
outputs = self(batch)
predictions = outputs.max(1)[1]
labels = batch[-1]
num_error = torch.nonzero(labels - predictions)
num_correct = labels.size(0) - num_error.size(0)
return num_correct
def export_state(self, dir, label, epoch=-1):
print("Saving models.")
cwd = os.getcwd() + '/'
enc_out = dir + ENC_1_FILE
mlp_out = dir + MLP_FILE
i2w_out = dir + I2W_FILE
w2i_out = dir + W2I_FILE
w2c_out = dir + W2C_FILE
inf_out = dir + INF_FILE
torch.save(self.encoder.state_dict(), enc_out)
torch.save(self.mlp.state_dict(), mlp_out)
i2w = open(i2w_out, 'wb')
pickle.dump(self.wd.index2word, i2w)
i2w.close()
w2i = open(w2i_out, 'wb')
pickle.dump(self.wd.word2index, w2i)
w2i.close()
w2c = open(w2c_out, 'wb')
pickle.dump(self.wd.word2count, w2c)
w2c.close()
info = open(inf_out, 'w')
using_lstm = 1 if self.base_rnn == nn.LSTM else 0
info.write(
str(self.encoder.hidden_size) + "\n" + str(self.encoder.n_layers) +
"\n" + str(self.wd.n_words) + "\n" + str(using_lstm))
if epoch > 0:
info.write("\n" + str(epoch))
info.close()
files = [enc_out, mlp_out, i2w_out, w2i_out, w2c_out, inf_out]
print("Bundling models")
tf = tarfile.open(cwd + dir + label, mode='w')
for file in files:
tf.add(file)
tf.close()
for file in files:
os.remove(file)
print("Finished saving models.")
def import_state(self, model_file, active_dir=TEMP_DIR, load_epoch=False):
print("Loading models.")
cwd = os.getcwd() + '/'
tf = tarfile.open(model_file)
# extract directly to current model directory
for member in tf.getmembers():
if member.isreg():
member.name = os.path.basename(member.name)
tf.extract(member, path=active_dir)
info = open(active_dir + INF_FILE, 'r')
lns = info.readlines()
hidden_size, e_layers, n_words, using_lstm = [int(i) for i in lns[:4]]
if load_epoch:
epoch = int(lns[-1])
i2w = open(cwd + TEMP_DIR + I2W_FILE, 'rb')
w2i = open(cwd + TEMP_DIR + W2I_FILE, 'rb')
w2c = open(cwd + TEMP_DIR + W2C_FILE, 'rb')
i2w_dict = pickle.load(i2w)
w2i_dict = pickle.load(w2i)
w2c_dict = pickle.load(w2c)
wd = WordDict(dicts=[w2i_dict, i2w_dict, w2c_dict, n_words])
w2i.close()
i2w.close()
w2c.close()
self.base_rnn = nn.LSTM if using_lstm == 1 else nn.GRU
self.wd = wd
self.encoder = EncoderRNN(wd.n_words,
hidden_size,
n_layers=e_layers,
base_rnn=self.base_rnn)
self.mlp = torch.nn.Sequential(
torch.nn.Linear(int(hidden_size * 8), int(hidden_size)),
torch.nn.ReLU(), torch.nn.Dropout(0.1),
torch.nn.Linear(int(hidden_size), 3), torch.nn.Softmax(dim=1))
if not USE_CUDA:
self.encoder.load_state_dict(
torch.load(cwd + TEMP_DIR + ENC_1_FILE,
map_location=lambda storage, loc: storage))
self.mlp.load_state_dict(
torch.load(cwd + TEMP_DIR + MLP_FILE,
map_location=lambda storage, loc: storage))
else:
self.encoder.load_state_dict(
torch.load(cwd + TEMP_DIR + ENC_1_FILE))
self.mlp.load_state_dict(torch.load(cwd + TEMP_DIR + MLP_FILE))
self.encoder = self.encoder.cuda()
self.mlp = self.mlp.cuda()
self.encoder.eval()
self.mlp.eval()
self.parameter_list = [
self.encoder.parameters(),
self.mlp.parameters()
]
tf.close()
print("Loaded models.")
if load_epoch:
return self, epoch
else:
return self
def torch_batch_from_numpy_batch_without_label(self, batch):
batch = list(batch)
variable_indices = [0, 2]
for i in variable_indices:
var = Variable(torch.from_numpy(batch[i]))
if USE_CUDA:
var = var.cuda()
batch[i] = var
return batch
def predict(self, data):
batch_size = 1
batches = nli_batches_without_label(batch_size, data)
predictions = []
for batch in tqdm(batches):
batch = self.torch_batch_from_numpy_batch_without_label(batch)
outputs = self(batch, inference=True)
pred = outputs.max(1)[1]
predictions.append(pred)
return torch.cat(predictions)
def add_new_vocabulary(self, genre):
old_vocab_size = self.wd.n_words
print("Previous vocabulary size: " + str(old_vocab_size))
train_set = nli_preprocessor.get_multinli_text_hyp_labels(
genre=genre
) #nli_preprocessor.get_multinli_training_set(max_lines=args.max_lines)
matched_val_set = nli_preprocessor.get_multinli_matched_val_set(
) #genre_val_set(genre)
unmerged_sentences = []
for data in [train_set, matched_val_set]:
unmerged_sentences.extend([data["text"], data["hyp"]])
all_sentences = list(itertools.chain.from_iterable(unmerged_sentences))
for line in all_sentences:
self.wd.add_sentence(line)
print("New vocabulary size: " + str(self.wd.n_words))
print("Extending the Embedding layer with new vocabulary...")
num_new_words = self.wd.n_words - old_vocab_size
self.encoder.extend_embedding_layer(self.wd.word2index, num_new_words)
self.new_vocab_size = num_new_words
def freeze_source_params(self):
for name, param in self.named_parameters():
if "rnn" in name:
param.requires_grad = False
if ("M_k" in name or "M_v" in name) and "target_4" not in name:
param.requires_grad = False
for name, param in self.named_parameters():
if param.requires_grad is True:
print(name)
encoder_sd = checkpoint['en']
decoder_sd = checkpoint['de']
encoder_optimizer_sd = checkpoint['en_opt']
decoder_optimizer_sd = checkpoint['de_opt']
embedding_sd = checkpoint['embedding']
voc.__dict__ = checkpoint['voc_dict']
# Initialize word embeddings
embedding = nn.Embedding(voc.num_words, hidden_size)
embedding.load_state_dict(embedding_sd)
# Initialize encoder & decoder models
encoder = EncoderRNN(hidden_size, embedding, encoder_n_layers, dropout)
decoder = LuongAttnDecoderRNN(attn_model, embedding, hidden_size,
voc.num_words, decoder_n_layers, dropout)
encoder.load_state_dict(encoder_sd)
decoder.load_state_dict(decoder_sd)
encoder = encoder.to(device)
decoder = decoder.to(device)
# Set dropout layers to eval mode
encoder.eval()
decoder.eval()
# Initialize search module
searcher = GreedySearchDecoder(encoder, decoder)
# Begin chatting (uncomment and run the following line to begin)
evaluateInput(encoder, decoder, searcher, voc)
class Sequence2SequenceNetwork(object):
def __init__(self, config):
self.init_writer()
self.load_configuration(config)
self.load_vocabulary()
self.prepare_data()
self.build_model()
self.load_pretrained_model()
self.train_model()
self.save_model(self.n_epochs)
self.evaluate_all()
self.close_writer()
def init_writer(self):
self.writer = SummaryWriter()
def load_configuration(self, config):
# Load configuration
self.iter_num = 0
self.lr = config['lr']
self.gpu = config['gpu']
self.unit = config['unit']
self.clip = config['clip']
self.beta1 = config['beta1']
self.beta2 = config['beta2']
self.langs = config['langs']
self.fusion = config['fusion']
self.log_tb = config['log_tb']
self.epsilon = config['epsilon']
self.attn_model = config['attn']
self.dropout = config['dropout']
self.emb_mode = config['emb_mode']
self.save_dir = config['save_dir']
self.data_dir = config['data_dir']
self.n_epochs = config['n_epochs']
self.SOS_TOKEN = config['SOS_TOKEN']
self.EOS_TOKEN = config['EOS_TOKEN']
self.MAX_LENGTH = config['MAX_LENGTH']
self.latent_dim = config['latent_dim']
self.batch_size = config['batch_size']
self.model_code = config['model_code']
self.vocab_path = config['vocab_path']
self.hidden_size = config['hidden_size']
self.use_cuda = torch.cuda.is_available()
self.log_tb_every = config['log_tb_every']
self.enc_n_layers = config['enc_n_layers']
self.dec_n_layers = config['dec_n_layers']
self.dec_learning_ratio = config['dec_lr']
self.bidirectional = config['bidirectional']
self.enc_input_dim = config['enc_input_dim']
self.embedding_dim = config['embedding_dim']
self.use_scheduler = config['use_scheduler']
self.use_embeddings = config['use_embeddings']
self.lr_lower_bound = config['lr_lower_bound']
self.teacher_forcing_ratio = config['tf_ratio']
self.load_model_name = config['load_model_name']
self.modality = config[
'modalities'] # no splitting as it's not multimodal case
if self.modality in ['ss-vv', 'v-s']:
self.pretrained_modality = config['pretrained_modality']
self.generate_word_embeddings = config['generate_word_embeddings']
self.device = torch.device(
'cuda:{}'.format(self.gpu) if self.use_cuda else 'cpu')
def load_vocabulary(self):
try:
with open(self.vocab_path, 'rb') as f:
self.vocab = pickle.load(f)
except FileNotFoundError as e: # build vocab if it doesn't exist
self.vocab = buildVocab()
def prepare_data(self):
# Note: The below workaround is used a lot and doing so is okay
# because this script would only be run for unimodal cases
self.pairs = prepareData(self.langs, [self.modality])[self.modality]
num_pairs = len(self.pairs)
self.pairs = self.pairs[:self.batch_size *
(num_pairs // self.batch_size)]
random.shuffle(self.pairs)
self.n_iters = len(self.pairs)
print('\nLoading test data pairs')
self.test_pairs = prepareData(self.langs, [self.modality],
train=False)[self.modality]
random.shuffle(self.test_pairs)
print(random.choice(self.pairs))
if self.use_embeddings:
if self.generate_word_embeddings:
self.embedding_wts = generateWordEmbeddings(
self.vocab, self.emb_mode)
else:
self.embedding_wts = loadWordEmbeddings(self.emb_mode)
def build_model(self):
if self.use_embeddings:
self.embedding = nn.Embedding.from_pretrained(self.embedding_wts)
else:
self.embedding = nn.Embedding(self.vocab.n_words,
self.embedding_dim)
if self.modality == 't': # Need embedding only for t2t mode
self.encoder = EncoderRNN(self.embedding_dim,
self.hidden_size,
self.enc_n_layers,
self.dropout,
self.unit,
self.modality,
self.embedding,
fusion_or_unimodal=True).to(self.device)
else:
# Note: no embedding used here
self.encoder = EncoderRNN(self.enc_input_dim,
self.hidden_size,
self.enc_n_layers,
self.dropout,
self.unit,
self.modality,
fusion_or_unimodal=True).to(self.device)
self.decoder = DecoderRNN(self.attn_model, self.embedding_dim,
self.hidden_size, self.vocab.n_words,
self.unit, self.dec_n_layers, self.dropout,
self.embedding).to(self.device)
self.encoder_optimizer = optim.Adam(self.encoder.parameters(),
lr=self.lr)
self.decoder_optimizer = optim.Adam(self.decoder.parameters(),
lr=self.lr *
self.dec_learning_ratio)
self.epoch = 0 # define here to add resume training feature
self.project_factor = self.encoder.project_factor
self.latent2hidden = nn.Linear(self.latent_dim, self.hidden_size *
self.project_factor).to(self.device)
def load_pretrained_model(self):
if self.load_model_name:
checkpoint = torch.load(self.load_model_name,
map_location=self.device)
print('Loaded {}'.format(self.load_model_name))
self.epoch = checkpoint['epoch']
self.encoder.load_state_dict(checkpoint['en'])
self.decoder.load_state_dict(checkpoint['de'])
self.encoder_optimizer.load_state_dict(checkpoint['en_op'])
self.decoder_optimizer.load_state_dict(checkpoint['de_op'])
self.embedding.load_state_dict(checkpoint['embedding'])
def train_model(self):
best_score = 1e-200
print_loss_total = 0 # Reset every epoch
saving_skipped = 0
for epoch in range(self.epoch, self.n_epochs):
incomplete = False
for iter in range(0, self.n_iters, self.batch_size):
pairs = self.pairs[iter:iter + self.batch_size]
# Skip incomplete batch
if len(pairs) < self.batch_size:
incomplete = True
continue
training_batch = batch2TrainData(self.vocab, pairs,
self.modality)
# Extract fields from batch
input_variable, lengths, target_variable, \
mask, max_target_len, _ = training_batch
if incomplete:
break
# Run a training iteration with the current batch
loss = self.train(input_variable, lengths, target_variable,
mask, max_target_len, iter)
self.writer.add_scalar('{}loss'.format(self.data_dir), loss,
iter)
print_loss_total += loss
print_loss_avg = print_loss_total * self.batch_size / self.n_iters
print_loss_total = 0
print('Epoch: [{}/{}] Loss: {:.4f}'.format(epoch, self.n_epochs,
print_loss_avg))
# evaluate and save the model
curr_score = self.evaluate_all()
self.writer.add_scalar('{}bleu_score'.format(self.data_dir),
curr_score)
if curr_score > best_score:
saving_skipped = 0
best_score = curr_score
self.save_model(epoch)
saving_skipped += 1
if self.use_scheduler and saving_skipped > 3:
saving_skipped = 0
new_lr = self.lr * 0.5
print('Entered the dungeon...')
if new_lr > self.lr_lower_bound: # lower bound on lr
self.lr = new_lr
print('lr decreased to => {}'.format(self.lr))
def train(self, input_variable, lengths, target_variable, mask,
max_target_len, iter):
self.encoder.train()
self.decoder.train()
self.encoder_optimizer.zero_grad()
self.decoder_optimizer.zero_grad()
input_variable = input_variable.to(self.device)
lengths = lengths.to(self.device)
target_variable = target_variable.to(self.device)
mask = mask.to(self.device)
# Initialize variables
loss = 0
print_losses = []
n_totals = 0
# Forward pass through encoder
encoder_outputs, encoder_hidden = self.encoder(input_variable, lengths)
# Create initial decoder input (start with SOS tokens for each sentence)
decoder_input = torch.LongTensor([[self.SOS_TOKEN] * self.batch_size])
decoder_input = decoder_input.to(self.device)
# Set initial decoder hidden state to the encoder's final hidden state
if self.unit == 'gru':
decoder_hidden = encoder_hidden[:self.decoder.n_layers]
else:
decoder_hidden = (encoder_hidden[0][:self.decoder.n_layers],
encoder_hidden[1][:self.decoder.n_layers])
if iter % conf['log_tb_every'] == 0:
# Visualize latent space
if self.unit == 'gru':
vis_hidden = decoder_hidden[-1, :, :]
else:
vis_hidden = decoder_hidden[0][-1, :, :]
self.writer.add_embedding(vis_hidden,
tag='decoder_hidden_{}'.format(iter))
use_teacher_forcing = True if random.random(
) < self.teacher_forcing_ratio else False
if use_teacher_forcing:
for t in range(max_target_len):
decoder_output, decoder_hidden = self.decoder(
decoder_input, decoder_hidden, encoder_outputs)
# Teacher forcing: next input is current target
decoder_input = target_variable[t].view(1, -1)
# Calculate and accumulate loss
mask_loss, nTotal = self.mask_nll_loss(decoder_output,
target_variable[t],
mask[t])
loss += mask_loss
print_losses.append(mask_loss.item() * nTotal)
n_totals += nTotal
else:
for t in range(max_target_len):
decoder_output, decoder_hidden = self.decoder(
decoder_input, decoder_hidden, encoder_outputs)
# No teacher forcing: next input is decoder's own current output
_, topi = decoder_output.topk(1)
decoder_input = torch.LongTensor(
[[topi[i][0] for i in range(self.batch_size)]])
decoder_input = decoder_input.to(self.device)
# Calculate and accumulate loss
mask_loss, nTotal = self.mask_nll_loss(decoder_output,
target_variable[t],
mask[t])
loss += mask_loss
print_losses.append(mask_loss.item() * nTotal)
n_totals += nTotal
loss.backward()
# Clip gradients: gradients are modified in place
torch.nn.utils.clip_grad_norm_(self.encoder.parameters(), self.clip)
torch.nn.utils.clip_grad_norm_(self.decoder.parameters(), self.clip)
self.encoder_optimizer.step()
self.decoder_optimizer.step()
return sum(print_losses) / n_totals
def mask_nll_loss(self, inp, target, mask):
n_total = mask.sum()
cross_entropy = -torch.log(
torch.gather(inp, 1, target.view(-1, 1)).squeeze(1))
loss = cross_entropy.masked_select(mask).sum()
loss = loss.to(self.device)
return loss, n_total.item()
def save_model(self, epoch):
directory = self.save_dir
if not os.path.exists(directory):
os.makedirs(directory)
torch.save(
{
'epoch': epoch,
'en': self.encoder.state_dict(),
'de': self.decoder.state_dict(),
'en_op': self.encoder_optimizer.state_dict(),
'de_op': self.decoder_optimizer.state_dict(),
'embedding': self.embedding.state_dict()
}, '{}{}-{}-{}-{}.pth'.format(directory, self.model_code,
self.modality, self.langs, epoch))
def evaluate_all(self):
self.encoder.eval()
self.decoder.eval()
searcher = GreedySearchDecoder(self.encoder, self.decoder, None,
self.device, self.SOS_TOKEN)
refs = []
hyp = []
for pair in self.test_pairs:
output_words = self.evaluate(self.encoder, self.decoder, searcher,
self.vocab, pair[0])
if output_words:
final_output = []
for x in output_words:
if x == '<EOS>':
break
final_output.append(x)
refs.append([pair[1].split()])
hyp.append(final_output)
bleu_scores = calculateBleuScores(refs, hyp)
print('Bleu score: {bleu_1} | {bleu_2} | {bleu_3} | {bleu_4}'.format(
**bleu_scores))
eg_idx = random.choice(range(len(hyp)))
print(hyp[eg_idx], refs[eg_idx])
return bleu_scores['bleu_4']
def evaluate(self,
encoder,
decoder,
searcher,
vocab,
sentence_or_vector,
max_length=conf['MAX_LENGTH']):
with torch.no_grad():
if self.modality == 't': # `sentence_or_vector` ~> sentence
# Format input sentence as a batch
# words => indexes
indexes_batch = [
indexesFromSentence(vocab, sentence_or_vector)
]
if None in indexes_batch:
return None
for idx, indexes in enumerate(indexes_batch):
indexes_batch[idx] = indexes_batch[idx] + [self.EOS_TOKEN]
# Create lengths tensor
lengths = torch.tensor(
[len(indexes) for indexes in indexes_batch])
# Transpose dimensions of batch to match models' expectations
input_batch = torch.LongTensor(indexes_batch).transpose(0, 1)
else: # `sentence_or_vector` ~> vector
input_batch, lengths = inputVarVec([sentence_or_vector],
self.modality)
# Use appropriate device
input_batch = input_batch.to(self.device)
lengths = lengths.to(self.device)
# Decode sentence with searcher
tokens, scores = searcher(input_batch, lengths, max_length)
# indexes -> words
decoded_words = [
vocab.index2word[token.item()] for token in tokens
]
return decoded_words
def close_writer(self):
self.writer.close()