def do_predict(settings, args):
#
if args.ckpt_loading == "latest":
dir_ckpt = settings.model_dir
else:
dir_ckpt = settings.model_dir_best
#
pb_file = os.path.join(dir_ckpt, "model_frozen.pb")
#
# model
model = settings.ModelClass(settings)
model.prepare_for_prediction_with_pb(pb_file)
#
# data
if args.data == "test":
file_raw = os.path.join(args.dir_examples, "data_examples_test.txt")
elif args.data == "train":
file_raw = os.path.join(args.dir_examples, "data_examples_train.txt")
elif args.data == "valid":
file_raw = os.path.join(args.dir_examples, "data_examples_valid.txt")
#
data_raw = data_utils.load_from_file_raw(file_raw)
#
batch_stder = lambda x: data_utils.get_batch_std(x, settings)
data_batcher = DataBatcher(data_raw,
batch_stder,
settings.batch_size_eval,
single_pass=True)
#
# predict
count = 0
while True:
batch = data_batcher.get_next_batch()
#
if batch is None: break
if count == settings.max_batches_eval: continue #
#
count += 1
print(count)
#
print("batch data:")
print(batch["input_y"])
print("batch data end")
#
results = model.predict_with_pb_from_batch(batch)["logits"]
#
print("results:")
print(np.argmax(results[0], -1))
print("results end")
print()
#
#
settings.logger.info('prediction finished, with total num_batches: %d' %
count)
def do_eval(settings, args):
#
if args.ckpt_loading == "latest":
dir_ckpt = settings.model_dir
else:
dir_ckpt = settings.model_dir_best
#
# model
model = settings.ModelClass(settings)
model.prepare_for_train_and_valid(dir_ckpt)
model.assign_dropout_keep_prob(1.0)
#
# data
file_raw = os.path.join(args.dir_examples, "data_examples_test.txt")
data_raw = data_utils.load_from_file_raw(file_raw)
#
batch_stder = lambda x: data_utils.get_batch_std(x, settings)
data_batcher = DataBatcher(data_raw,
batch_stder,
settings.batch_size_eval,
single_pass=True)
#
# eval
eval_score, loss_aver, metric_aver = eval_process(
model, data_batcher, settings.max_batches_eval, mode_eval=True)
#
print('loss_aver, metric_aver: %g, %g' % (loss_aver, metric_aver))
settings.logger.info('loss_aver, metric_aver: %g, %g' %
(loss_aver, metric_aver))
settings.logger.info('{}'.format(eval_score))
#
settings.result_dir = os.path.join(settings.base_dir, "result")
if not os.path.exists(settings.base_dir):
os.mkdir(settings.base_dir)
if not os.path.exists(settings.result_dir):
os.mkdir(settings.result_dir)
#
# model & vocab
settings.model_tag = "multi_doc_qa"
settings.vocab = vocab
#
# mode
if args.mode == "train":
example_gen = lambda single_pass: example_generator(train_files, True, settings.max_p_len, single_pass)
batch_stder = lambda items: do_batch_std(items, vocab, settings)
batcher = DataBatcher(example_gen, batch_stder, settings.batch_size, single_pass=False)
#
settings.is_train = True
settings.check_settings()
settings.logger.info("{}".format(args))
#
model = ModelDocQA(settings)
model.prepare_for_train_and_valid(settings.model_dir)
#
model_utils.do_train(model, batcher, settings)
model.close_logger()
#
elif args.mode == "eval":
example_gen = lambda single_pass: example_generator(dev_files, True, settings.max_p_len, single_pass)
batch_stder = lambda items: do_batch_std(items, vocab, settings)
batcher = DataBatcher(example_gen, batch_stder, settings.batch_size, single_pass=True) # batch_size
def do_train_and_valid(settings, args):
#
if args.ckpt_loading == "latest":
dir_ckpt = settings.model_dir
else:
dir_ckpt = settings.model_dir_best
#
# model
model = settings.ModelClass(settings)
model.prepare_for_train_and_valid(dir_ckpt)
#
# data
file_raw = os.path.join(args.dir_examples, "data_examples_train.txt")
data_raw = data_utils.load_from_file_raw(file_raw)
#
batch_stder = lambda x: data_utils.get_batch_std(x, settings)
data_batcher = DataBatcher(data_raw,
batch_stder,
settings.batch_size,
single_pass=False)
#
eval_period = settings.valid_period_batch
file_raw_eval = os.path.join(args.dir_examples, "data_examples_valid.txt")
data_raw_eval = data_utils.load_from_file_raw(file_raw_eval)
#
# train
loss = 10000.0
best_metric_val = 0
# last_improved = 0
lr = 0.0
#
count = 0
model.settings.logger.info("")
while True:
#
# eval
if count % eval_period == 0:
settings.logger.info("training curr batch, loss, lr: %d, %g, %g" %
(count, loss, lr))
#
model.save_ckpt(settings.model_dir, settings.model_name, count)
model.assign_dropout_keep_prob(1.0)
#
settings.logger.info('evaluating after num_batches: %d' % count)
eval_batcher = DataBatcher(data_raw_eval,
batch_stder,
settings.batch_size,
single_pass=True)
#
eval_score, loss_aver, metric_val = eval_process(
model,
eval_batcher,
settings.max_batches_eval,
mode_eval=False)
settings.logger.info(
"eval loss_aver, metric, metric_best: %g, %g, %g" %
(loss_aver, metric_val, best_metric_val))
#
# save best
if metric_val >= best_metric_val: # >=
best_metric_val = metric_val
# last_improved = count
# ckpt
settings.logger.info('a new best model, saving ...')
model.save_ckpt_best(settings.model_dir_best,
settings.model_name, count)
#
#
if lr < settings.learning_rate_minimum and count > settings.warmup_steps:
settings.logger.info(
'current learning_rate < learning_rate_minimum, stop training'
)
break
#
model.assign_dropout_keep_prob(settings.keep_prob)
settings.logger.info("")
#
#
# train
batch = data_batcher.get_next_batch()
# if batch is None: break
count += 1
# print(count)
#
result_dict = model.run_train_one_batch(batch) # just for train
loss = result_dict["loss_optim"]
lr = result_dict["lr"]
#
# print(loss)
# model.logger.info("training curr batch, loss, lr: %d, %g, %g" % (count, loss, lr)
#
#
settings.logger.info("training finshed with total num_batches: %d" % count)
def do_debug(settings, args):
#
if args.ckpt_loading == "latest":
dir_ckpt = settings.model_dir
else:
dir_ckpt = settings.model_dir_best
#
# model
model = ModelWrapper(settings, settings.model_graph)
model.prepare_for_train_and_valid(dir_ckpt)
model.assign_dropout_keep_prob(1.0)
#
# data
file_raw = os.path.join(args.dir_examples, "data_examples_test.txt")
data_raw = data_utils.load_from_file_raw(file_raw)
#
batch_stder = lambda x: data_utils.get_batch_std(x, settings)
data_batcher = DataBatcher(data_raw,
batch_stder,
settings.batch_size_eval,
single_pass=True)
#
# eval
list_batches_result = []
#
loss_aver, metric_aver = 0.0, 0.0
count = 0
while True:
batch = data_batcher.get_next_batch()
#
if batch is None: break
if count == 1000000: continue #
#
count += 1
# print(count)
#
results, loss, metric = model.run_eval_one_batch(batch)
loss_aver += loss
metric_aver += metric
# print(loss)
# print(metric)
#
print(count)
print("batch data:")
print(batch[-1])
#
print("results:")
print(np.argmax(results[0], -1))
print()
#
item = batch[0], batch[1], np.argmax(results[0], -1)
list_batches_result.append(item)
#
#
dir_result = "data_check_result"
if not os.path.exists(dir_result): os.mkdir(dir_result)
#
file_path = os.path.join(
dir_result, "list_batches_result_%d.pkl" % settings.batch_size_eval)
#
with open(file_path, 'wb') as fp:
pickle.dump(list_batches_result, fp)
#
loss_aver /= count
metric_aver /= count
#
print('loss_aver, metric_aver: %g, %g' % (loss_aver, metric_aver))
model.logger.info('loss_aver, metric_aver: %g, %g' %
(loss_aver, metric_aver))