print_loss_total += loss
plot_loss_total += loss
if iter % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_total = 0
print('%s (%d %d%%) %.4f' % (timeSince(start, iter / n_iters),
iter, iter / n_iters * 100, print_loss_avg))
if iter % plot_every == 0:
plot_loss_avg = plot_loss_total/plot_every
plot_losses.append(plot_loss_avg)
plot_loss_total = 0
showPlot(plot_losses)
plt.savefig('base-lstm-loss')
torch.save(encoder.state_dict(), 'encoder.pth')
torch.save(decoder.state_dict(), 'decoder.pth')
if __name__ == "__main__":
hidden_size = 256
n_iters = 910000
teacher_forcing_ratio = 0.5
input_lang, output_lang, pairs = loader.prepareData('eng', 'fra', True)
encoder = seq2seq.Encoder(input_lang.n_words, hidden_size).to(device)
decoder = seq2seq.Decoder(hidden_size, output_lang.n_words).to(device)
trainiters(encoder, decoder, n_iters)
if load_corpus:
with open(corpus_path, 'rb') as f:
input_lang, output_lang, pairs = pickle.load(f)
else:
input_lang, output_lang, pairs = corpus.read_file(
'ENG', 'FRA', 'data/eng-fra.txt', True)
pairs = corpus.filter_pairs(pairs)
for pair in pairs:
input_lang.add_sentence(pair[0])
output_lang.add_sentence(pair[1])
with open(corpus_path, 'wb') as f:
pickle.dump((input_lang, output_lang, pairs), f)
print(f'{len(pairs)} pairs, {input_lang.n_words} source, {output_lang.n_words} target')
# Load model
encoder = seq2seq.Encoder(input_lang.n_words, hidden_size, embed_size, n_layers)
decoder = seq2seq.Decoder(hidden_size, embed_size, output_lang.n_words, n_layers)
model = seq2seq.Seq2seq(encoder, decoder).to(device)
if load_model:
model.load_state_dict(torch.load(model_path))
# train.train(model, (input_lang, output_lang, pairs), batch_size, n_epochs, learning_rate,
# teaching_rate, clip, model_path)
else:
def init_weights(m):
for name, param in m.named_parameters():
nn.init.uniform_(param.data, -0.08, 0.08)
model.apply(init_weights)
# Test or train
if mode == 'train':
train.train(model, (input_lang, output_lang, pairs), batch_size, n_epochs, learning_rate,
if __name__ == "__main__":
config = argparser()
embedding_size = config.embedding_size
hidden_size = config.hidden_size
teacher_forcing_ratio = config.teacher_forcing_ratio
n_iters = config.n_iters
input_lang, output_lang, pairs = loader.prepareData('eng', 'fra', True)
input_emb_matrix, output_emb_matrix = np.load(
'input_emb_matrix.npy'), np.load('output_emb_matrix.npy')
print('Embedding-matrix shape: {}, {}'.format(input_emb_matrix.shape,
output_emb_matrix.shape))
encoder = seq2seq.Encoder(input_size=input_lang.n_words,
embedding_size=embedding_size,
hidden_size=hidden_size,
embedding_matrix=input_emb_matrix,
n_layers=config.n_layers,
dropout_p=config.dropout_p).to(device)
decoder = seq2seq.AttnDecoder(output_size=output_lang.n_words,
embedding_size=embedding_size,
hidden_size=hidden_size,
embedding_matrix=output_emb_matrix,
n_layers=config.n_layers,
dropout_p=config.dropout_p).to(device)
trainiters(pairs, encoder, decoder, n_iters)
print('BLEU: {:.4}'.format(sum(scores) / n))
if __name__ == "__main__":
'''
Evaluation is mostly the same as training,
but there are no targets so we simply feed the decoder’s predictions back to itself for each step.
Every time it predicts a word we add it to the output string,
and if it predicts the EOS token we stop there.
'''
hidden_size = 300
input_lang, output_lang, pairs = loader.prepareData('eng', 'fra', True)
input_emb_matrix, output_emb_matrix = np.load(
'input_emb_matrix.npy'), np.load('output_emb_matrix.npy')
print('Embedding-matrix shape: {}, {}'.format(input_emb_matrix.shape,
output_emb_matrix.shape))
encoder = seq2seq.Encoder(input_lang.n_words, hidden_size,
input_emb_matrix).to(device)
decoder = seq2seq.Decoder(hidden_size, output_lang.n_words,
output_emb_matrix).to(device)
encoder.load_state_dict(torch.load('encoder.pth'))
encoder.eval()
decoder.load_state_dict(torch.load('decoder.pth'))
decoder.eval()
evaluateRandomly(encoder, decoder, pairs, int(len(pairs) * 0.3))
BATCH_SIZE = 64 # try bigger batch for faster training
train = ds_train.take(BUFFER_SIZE) # 1.5M samples
print("Dataset sample taken")
train_dataset = train.map(s2s.tf_encode)
# train_dataset = train_dataset.shuffle(BUFFER_SIZE) – optional
train_dataset = train_dataset.batch(BATCH_SIZE, drop_remainder=True)
print("Dataset batching done")
steps_per_epoch = BUFFER_SIZE // BATCH_SIZE
embedding_dim = 128
units = 256 # from pointer generator paper
EPOCHS = 6
encoder = s2s.Encoder(vocab_size, embedding_dim, units, BATCH_SIZE)
decoder = s2s.Decoder(vocab_size, embedding_dim, units, BATCH_SIZE)
# Learning rate scheduler
lr_schedule = tf.keras.optimizers.schedules.InverseTimeDecay(
0.001,
decay_steps=steps_per_epoch * (EPOCHS / 2),
decay_rate=2,
staircase=False)
optimizer = tf.keras.optimizers.Adam(lr_schedule)
if args.checkpoint is None:
dt = datetime.datetime.today().strftime("%Y-%b-%d-%H-%M-%S")
checkpoint_dir = './training_checkpoints-' + dt
else: