def main(_):
random.seed(FLAGS.seed)
np.random.seed(FLAGS.seed)
dataset = Dataset(FLAGS.dataset)
train_input_fn = dataset.get_input_fn('train',
batch_size=FLAGS.batch_size,
num_epochs=FLAGS.num_epochs,
shuffle=True)
eval_input_fn = dataset.get_input_fn('test',
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
params = {
'vocab_size': dataset.vocab_size,
'embedding_size': FLAGS.embedding_size,
'num_blocks': FLAGS.num_blocks,
'learning_rate_init': FLAGS.learning_rate,
'learning_rate_decay_steps': (10000 // FLAGS.batch_size) * 25,
'learning_rate_decay_rate': 0.5,
'clip_gradients': FLAGS.clip_gradients,
'debug': FLAGS.debug,
}
eval_metrics = {
"accuracy":
tf.contrib.learn.metric_spec.MetricSpec(
tf.contrib.metrics.streaming_accuracy)
}
config = tf.contrib.learn.RunConfig(tf_random_seed=FLAGS.seed,
save_summary_steps=120,
save_checkpoints_secs=600,
keep_checkpoint_max=5,
keep_checkpoint_every_n_hours=1,
log_device_placement=True)
dataset_name = os.path.splitext(os.path.basename(FLAGS.dataset))[0]
timestamp = int(time.time())
model_dir = os.path.join(FLAGS.model_dir, dataset_name, str(timestamp))
estimator = tf.contrib.learn.Estimator(model_dir=model_dir,
model_fn=model_fn,
config=config,
params=params)
experiment = tf.contrib.learn.Experiment(estimator,
train_input_fn,
eval_input_fn,
train_steps=None,
eval_steps=None,
eval_metrics=eval_metrics,
train_monitors=None,
local_eval_frequency=1)
experiment.train_and_evaluate()
def main(_):
dataset = Dataset(FLAGS.dataset)
input_fn = dataset.get_input_fn('test',
batch_size=FLAGS.batch_size,
num_epochs=1,
shuffle=False)
params = {
'vocab_size': dataset.vocab_size,
'embedding_size': FLAGS.embedding_size,
'num_blocks': FLAGS.num_blocks,
'debug': False,
}
config = tf.contrib.learn.RunConfig(save_summary_steps=120,
save_checkpoints_secs=600,
keep_checkpoint_max=5,
keep_checkpoint_every_n_hours=1,
log_device_placement=True)
eval_metrics = {
"accuracy":
tf.contrib.learn.metric_spec.MetricSpec(
tf.contrib.metrics.streaming_accuracy)
}
estimator = tf.contrib.learn.Estimator(model_dir=FLAGS.model_dir,
model_fn=model_fn,
config=config,
params=params)
experiment = tf.contrib.learn.Experiment(estimator,
train_input_fn=None,
eval_input_fn=input_fn,
train_steps=None,
eval_steps=None,
eval_metrics=eval_metrics,
train_monitors=None,
local_eval_frequency=1)
experiment.evaluate(delay_secs=10)