def train(train, dev, model, model_dir, batch_size, glove, beam_size,
samples_per_epoch, val_samples, cmodel, epochs):
if not os.path.exists(model_dir):
os.makedirs(model_dir)
hypo_len = model.get_layer('hypo_input').input_shape[1] -1
ne = model.get_layer('noise_embeddings')
vae = model.get_layer('vae_output')
g_train = train_generator(train, batch_size, hypo_len,
'class_input' in model.input_names, ne, vae)
saver = ModelCheckpoint(model_dir + '/weights.hdf5', monitor = 'hypo_loss', mode = 'min', save_best_only = True)
#saver = ModelCheckpoint(model_dir + '/weights{epoch:02d}.hdf5')
#es = EarlyStopping(patience = 4, monitor = 'hypo_loss', mode = 'min')
csv = CsvHistory(model_dir + '/history.csv')
gtest = gm.gen_test(model, glove, batch_size)
noise_size = ne.output_shape[-1] if ne else model.get_layer('expansion').input_shape[-1]
cb = ValidateGen(dev, gtest, beam_size, hypo_len, val_samples, noise_size, glove, cmodel, True, True)
hist = model.fit_generator(g_train, samples_per_epoch = samples_per_epoch, nb_epoch = epochs,
callbacks = [cb, saver, csv])
return hist
def train(train, dev, model, model_dir, batch_size, glove, beam_size,
samples_per_epoch, val_samples, cmodel, epochs):
if not os.path.exists(model_dir):
os.makedirs(model_dir)
hypo_len = model.get_layer('hypo_input').input_shape[1] - 1
ne = model.get_layer('noise_embeddings')
vae = model.get_layer('vae_output')
g_train = train_generator(train, batch_size, hypo_len, 'class_input'
in model.input_names, ne, vae)
saver = ModelCheckpoint(model_dir + '/weights.hdf5',
monitor='hypo_loss',
mode='min',
save_best_only=True)
#saver = ModelCheckpoint(model_dir + '/weights{epoch:02d}.hdf5')
#es = EarlyStopping(patience = 4, monitor = 'hypo_loss', mode = 'min')
csv = CsvHistory(model_dir + '/history.csv')
gtest = gm.gen_test(model, glove, batch_size)
noise_size = ne.output_shape[-1] if ne else model.get_layer(
'expansion').input_shape[-1]
cb = ValidateGen(dev, gtest, beam_size, hypo_len, val_samples, noise_size,
glove, cmodel, True, True)
hist = model.fit_generator(g_train,
samples_per_epoch=samples_per_epoch,
nb_epoch=epochs,
callbacks=[cb, saver, csv])
return hist
cmodel = cm.attention_model(c_hidden_size, glove)
if os.path.exists(orig_cmodel_dir):
cmodel.load_weights(orig_cmodel_dir + 'model.weights')
if method == 'orig_class':
ca.train(train, dev, cmodel, orig_cmodel_dir, batch_size)
if method == 'train_gen':
gtrain = gm.gen_train(len(train[0]), g_hidden_size, latent_size, glove, hypo_len, version)
ga.train(train, dev, gtrain, dir_name, batch_size, glove, beam_size,
epoch_size, dev_sample_size, cmodel, gen_epochs)
if method == 'augment':
gtrain = gm.gen_train(len(train[0]), g_hidden_size, latent_size, glove, hypo_len, version)
gtrain.load_weights(dir_name + '/weights.hdf5')
gtest = gm.gen_test(gtrain, glove, batch_size)
augment.new_generate_save(dev, dir_name, augment_file_size, gtest, beam_size, hypo_len,
latent_size, cmodel, wi, 'dev', len(dev[0]), aug_threshold)
augment.new_generate_save(train, dir_name, augment_file_size, gtest, beam_size, hypo_len,
latent_size, cmodel, wi, 'train', len(train[0]), aug_threshold)
if method == 'train_class':
for t in thresholds:
if type(t) == str and t[0] == 'a':
aug_train, aug_dev = augment.load_dataset(dir_name, True, 2**30, 2**30, wi, prem_len, hypo_len)
aug_dev = aa.filter_adverserial(aug_dev, t, len(dev[0]), dir_name, glove, a_hidden_size)
aug_train = aa.filter_adverserial(aug_train, t, len(train[0]), dir_name, glove, a_hidden_size)
else:
aug_train, aug_dev = augment.load_dataset(dir_name, t, len(train[0]), len(dev[0]), wi, prem_len, hypo_len)
aug_cmodel = cm.attention_model(c_hidden_size, glove)
g_hidden_size) + '-' + str(latent_size)
cmodel = cm.attention_model(c_hidden_size, glove)
cmodel.load_weights('models/cmodel/model.weights')
if method == 'train_gen':
gtrain = gm.gen_train(len(train[0]), g_hidden_size, latent_size, glove,
hypo_len, version)
ga.train(train, dev, gtrain, dir_name, batch_size, glove, beam_size,
epoch_size, dev_sample_size, cmodel, gen_epochs)
if method == 'augment':
gtrain = gm.gen_train(len(train[0]), g_hidden_size, latent_size, glove,
hypo_len, version)
gtrain.load_weights(dir_name + '/weights.hdf5')
gtest = gm.gen_test(gtrain, glove, batch_size)
augment.new_generate_save(dev, dir_name, augment_file_size, gtest,
beam_size, hypo_len, latent_size, cmodel, wi,
'dev', len(dev[0]), aug_threshold)
augment.new_generate_save(train, dir_name, augment_file_size, gtest,
beam_size, hypo_len, latent_size, cmodel, wi,
'train', len(train[0]), aug_threshold)
if method == 'train_class':
for t in thresholds:
if type(t) == str and t[0] == 'a':
aug_train, aug_dev = augment.load_dataset(
dir_name, True, 2**30, 2**30, wi, prem_len, hypo_len)
aug_dev = aa.filter_adverserial(aug_dev, t, len(dev[0]),
dir_name, glove, a_hidden_size)
aug_train = aa.filter_adverserial(aug_train, t, len(train[0]),
