def objective(hyperpar):
[
init_method,
max_init_weight,
embed_size,
rnn_size,
post_image_size,
pre_output_size,
post_image_activation,
rnn_type,
optimizer,
learning_rate,
normalize_image,
weights_reg_weight,
image_dropout_prob,
post_image_dropout_prob,
embedding_dropout_prob,
rnn_dropout_prob,
max_gradient_norm,
minibatch_size,
beam_width,
] = hyperpar
with model_neural_trad.TradNeuralModel(
vocab_size=vocab.size,
init_method=init_method,
max_init_weight=max_init_weight,
embed_size=embed_size,
rnn_size=rnn_size,
post_image_size=post_image_size,
pre_output_size=pre_output_size,
post_image_activation=post_image_activation,
rnn_type=rnn_type,
architecture='merge',
optimizer=optimizer,
learning_rate=learning_rate,
normalize_image=normalize_image,
weights_reg_weight=weights_reg_weight,
image_dropout_prob=image_dropout_prob,
post_image_dropout_prob=post_image_dropout_prob,
embedding_dropout_prob=embedding_dropout_prob,
rnn_dropout_prob=rnn_dropout_prob,
max_gradient_norm=max_gradient_norm,
freeze_prefix_params=True,
) as model:
model.compile_model()
result = list()
for _ in range(config.hyperpar_num_runs):
model.init_params()
model.set_prefix_params(best_prefix_params)
model.fit(dataset,
config.hyperpar_dir + '/langmodtrans/' + corpus +
'/2_model.hdf5',
max_batch_size=config.val_batch_size,
minibatch_size=minibatch_size,
max_epochs=config.hyperpar_max_epochs,
early_stop_patience=config.early_stop_patience)
(index_sents, logprobs) = model.generate_sents_beamsearch(
max_batch_size=config.val_batch_size,
images=test_images,
beam_width=beam_width,
lower_bound_len=config.lower_bound_len,
upper_bound_len=config.upper_bound_len,
temperature=config.temperature)
text_sents = index_sents.decompile_sents(vocab).sents
wmd = evaluation.get_wmd_score(test_sents, text_sents)[0]
result.append(wmd)
return -np.mean(result)
langmod_minibatch_size = int(hyperpars['minibatch_size'])
langmod_beam_width = int(hyperpars['beam_width'])
correct_sents = datasources['test'].tokenize_sents().text_sents.sents
with model_neural_trad.TradNeuralModel(
vocab_size=langmod_vocab.size,
init_method=langmod_init_method,
max_init_weight=langmod_max_init_weight,
embed_size=langmod_embed_size,
rnn_size=langmod_rnn_size,
post_image_size=langmod_post_image_size,
pre_output_size=langmod_pre_output_size,
post_image_activation=langmod_post_image_activation,
rnn_type=langmod_rnn_type,
architecture='langmod',
optimizer=langmod_optimizer,
learning_rate=langmod_learning_rate,
normalize_image=langmod_normalize_image,
weights_reg_weight=langmod_weights_reg_weight,
image_dropout_prob=langmod_image_dropout_prob,
post_image_dropout_prob=langmod_post_image_dropout_prob,
embedding_dropout_prob=langmod_embedding_dropout_prob,
rnn_dropout_prob=langmod_rnn_dropout_prob,
max_gradient_norm=langmod_max_gradient_norm,
freeze_prefix_params=False,
) as model:
model.compile_model()
model.set_params(
model.load_params(config.results_dir + '/langmodtrans/' +
dataset_name + '/' + lm_dir_name +
'/1_model.hdf5'))
def objective(hyperpar):
[
init_method,
max_init_weight,
embed_size,
rnn_size,
post_image_size,
pre_output_size,
post_image_activation,
rnn_type,
optimizer,
learning_rate,
normalize_image,
weights_reg_weight,
image_dropout_prob,
post_image_dropout_prob,
embedding_dropout_prob,
rnn_dropout_prob,
max_gradient_norm,
minibatch_size,
beam_width,
] = hyperpar
with model_neural_trad.TradNeuralModel(
vocab_size=vocab.size,
init_method=init_method,
max_init_weight=max_init_weight,
embed_size=embed_size,
rnn_size=rnn_size,
post_image_size=post_image_size,
pre_output_size=pre_output_size,
post_image_activation=post_image_activation,
rnn_type=rnn_type,
architecture='langmod',
optimizer=optimizer,
learning_rate=learning_rate,
normalize_image=normalize_image,
weights_reg_weight=weights_reg_weight,
image_dropout_prob=image_dropout_prob,
post_image_dropout_prob=post_image_dropout_prob,
embedding_dropout_prob=embedding_dropout_prob,
rnn_dropout_prob=rnn_dropout_prob,
max_gradient_norm=max_gradient_norm,
freeze_prefix_params=False,
) as model:
model.compile_model()
result = list()
for _ in range(config.hyperpar_num_runs):
model.init_params()
model.fit(dataset,
config.hyperpar_dir + '/langmodtrans/' + corpus +
'/1_model.hdf5',
max_batch_size=config.val_batch_size,
minibatch_size=minibatch_size,
max_epochs=config.hyperpar_max_epochs,
early_stop_patience=config.early_stop_patience)
(logpplx, num_inf_pplx) = evaluation.get_loggeomean_perplexity(
model.get_sents_logprobs(
max_batch_size=config.val_batch_size,
index_sents=test_index_sents)[0],
test_index_sents.lens)
if num_inf_pplx > 0:
raise InfinitePerplexityError()
result.append(logpplx)
return np.mean(result)
]
else:
correct_sents = datasources['test'].shuffle(seed=0).take(
one_per_group=True).tokenize_sents().text_sents.sents
model = model_neural_trad.TradNeuralModel(
vocab_size=vocab.size,
init_method=init_method,
max_init_weight=max_init_weight,
embed_size=embed_size,
rnn_size=rnn_size,
post_image_size=post_image_size,
pre_output_size=pre_output_size,
post_image_activation=post_image_activation,
rnn_type=rnn_type,
architecture=architecture,
optimizer=optimizer,
learning_rate=learning_rate,
normalize_image=normalize_image,
weights_reg_weight=weights_reg_weight,
image_dropout_prob=image_dropout_prob,
post_image_dropout_prob=post_image_dropout_prob,
embedding_dropout_prob=embedding_dropout_prob,
rnn_dropout_prob=rnn_dropout_prob,
max_gradient_norm=max_gradient_norm,
freeze_prefix_params=False,
)
model.compile_model()
model.set_params(
model.load_params(config.results_dir + '/whereimage/' +
architecture + '/' + dir_name +
'/model.hdf5'))
