def epoch_ready(self, model, epoch_num, train_logpplx, val_logpplx):
if epoch_num == 0:
print(' '*lib.ProgressBar.width(5), end=' | \t')
else:
print(' | ', end='\t')
print(round(train_logpplx, 3), round(val_logpplx, 3), lib.format_duration(self.epoch_timer.get_duration()), sep='\t')
self.training_prog = None
print(
i,
'<<FOUND HYPERPARAMS THAT RESULTED IN ERRORS LAST TIME>>'
)
num_hyperpars += 1
opt.tell(prepare_hyperpar_for_tell(next_hyperpar), cost)
if best_cost is None or cost < best_cost:
best_hyperpar = next_hyperpar
best_cost = cost
already_seen.add(tuple(next_hyperpar))
print(i,
*next_hyperpar,
cost,
lib.format_duration(duration),
'******' if cost == best_cost else '',
sep='\t')
for _ in range(
i, config.hyperpar_num_random_evals + config.hyperpar_num_evals):
i += 1
num_hyperpars = 1
while True:
t = lib.Timer()
next_hyperpar = standardize_hyperpar(
opt.ask(num_hyperpars)[-1]
) #This allows us to get different hyperparameters every time the previous hyperparameters resulted in <<SEEN>>, <<NAN>>, or <<EMPTY>>
num_hyperpars += 1
mean_grad_max_wrt_multimodalvec = np.mean(
np.mean(grad_max_wrt_multimodalvec, axis=1))
full_duration = full_timer.get_duration()
with open(config.results_dir + '/imageimportance/' +
architecture + '/results_langmod.txt',
'a',
encoding='utf-8') as f:
print(dataset_name,
run,
sent_len - 1,
token_index,
mean_grad_next_wrt_prefix,
mean_grad_max_wrt_prefix,
mean_grad_next_wrt_prevtoken,
mean_grad_max_wrt_prevtoken,
mean_grad_next_wrt_firsttoken,
mean_grad_max_wrt_firsttoken,
mean_grad_next_wrt_multimodalvec,
mean_grad_max_wrt_multimodalvec,
full_duration,
sep='\t',
file=f)
already_seen.add((architecture, dataset_name, run,
sent_len, token_index))
print(lib.format_duration(full_duration))
print()
print(lib.formatted_clock())
capgen_generation_stats['Bleu_2'],
capgen_generation_stats['Bleu_3'],
capgen_generation_stats['Bleu_4'],
capgen_generation_stats['METEOR'],
capgen_generation_stats['ROUGE_L'],
capgen_generation_stats['CIDEr'],
capgen_generation_stats['SPICE'],
capgen_generation_stats['WMD'],
capgen_retrieval_stats['R@1'],
capgen_retrieval_stats['R@5'],
capgen_retrieval_stats['R@10'],
capgen_retrieval_stats['median_rank'],
capgen_retrieval_stats['R@1_frac'],
capgen_retrieval_stats['R@5_frac'],
capgen_retrieval_stats['R@10_frac'],
capgen_retrieval_stats['median_rank_frac'],
capgen_fit_stats['num_epochs'],
capgen_train_duration,
capgen_full_duration,
sep='\t',
file=f)
already_seen.add(
(corpus, frozen_prefix, corpus_size_factor_exponent, run))
print()
print(
lib.format_duration(langmod_full_duration +
capgen_full_duration))
print()
print(lib.formatted_clock())
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/'+corpus+'/'+dir_name+'/1_model.hdf5'))
(sents_logprobs, tokens_logprobs) = model.get_sents_logprobs(max_batch_size=config.val_batch_size, index_sents=dataset.val.index_sents)
val_langmod_prob_stats = evaluation.get_probability_stats(sents_logprobs, dataset.val.index_sents.lens)
langmod_prob_stats = evaluation.get_probability_stats(sents_logprobs, dataset.val.index_sents.lens, capgen_num_unknowns_per_sent, capgen_num_out_of_vocab_tokens)
print('Done!')
print(lib.format_duration(full_timer.get_duration()))
with open(config.results_dir+'/langmodtrans/'+corpus+'/val_pplx.txt', 'a', encoding='utf-8') as f:
print(
corpus,
frozen_prefix,
corpus_size_factor_exponent,
run,
langmod_prob_stats['mean_prob'],
langmod_prob_stats['median_prob'],
langmod_prob_stats['geomean_prob'],
langmod_prob_stats['mean_pplx'],
langmod_prob_stats['median_pplx'],
langmod_prob_stats['geomean_pplx'],
sep='\t', file=f
)