softmax = Softmax(dim_proj, pt_v_size)
# ensemble the sequence-to-sequence model
seq = SequenceToSequence(encoder=encoder,
decoder=decoder,
output=softmax,
source_v_size=en_v_size,
target_v_size=pt_v_size)
# set optimizer
optimizer = Adadelta()
# set up the model
seq.setup(batch_size=batch_size, optimizer=optimizer)
time2 = time.time()
print 'Initialization took %3.5f seconds. \n' % (time2 - time1)
# perform the optimization
seq.train(train_data,
valid_data,
n_epochs=n_epochs,
n_train_samples=160975,
n_valid_samples=1375,
print_train_info=True,
save_model=True,
filepath='/home/gian/seq_to_seq.hp5y',
keep_old_models=True)
print 'Done!'
target_v_size=pt_v_size)
# set optimizer
optimizer = Adadelta()
# set up the model
seq.setup(batch_size=batch_size, optimizer=optimizer)
time2 = time.time()
print 'Initialization took %3.5f seconds. \n' % (time2 - time1)
model_file = '/home/gian/%s_%shid_prj%s_en%s_pt%s_%s_batch%s.hp5y' % \
(gru1.__class__.__name__, len(encoder), dim_proj,
en_v_size, pt_v_size, optimizer.__class__.__name__,
batch_size)
print model_file
# perform the optimization
seq.train(train_data,
valid_data,
n_epochs=n_epochs,
n_train_samples=160975,
n_valid_samples=1375,
print_train_info=True,
save_model=True,
filepath=model_file,
keep_old_models=True)
print 'Done!'
# ensemble the sequence-to-sequence model
seq = SequenceToSequence(encoder=encoder,
decoder=decoder,
source_v_size=en_v_size,
target_v_size=pt_v_size,
auto_setup=False) # set auto_setup to false to avoid initialization
# (weights will be overwritten anyway)
# load source and target language dictionaries
sr_dict = load_dictionary('/home/gian/datasets/dict.sort.en', max_words=en_v_size)
tr_dict = load_dictionary('/home/gian/datasets/dict.sort.pt', max_words=pt_v_size)
# load the corpora and convert its words to their indexes (corpora must be already tokenized)
sequences1 = load_and_convert_corpora('/home/gian/datasets/fapesp/fapesp-v2.tok.test-a.en', sr_dict)
sequences2 = load_and_convert_corpora('/home/gian/datasets/fapesp/fapesp-v2.tok.test-a.pt', tr_dict)
# prepare the data (add padding values to the end of each sequence so they have the same size)
seq.load_weights('/home/gian/seq_to_seq.hp5y')
seq.train(sequences1,
sequences2,
n_epochs=2,
print_train_info=True)
#
# example = numpy.asarray(sequences1[0])
# seq.translate(example.reshape(1, example.shape[0]))
print 'Done!'
