def eval_driver(params, checkpoint_dir, model_dir, input_fn, model_fn):
"""Eval.
Args:
params: hyper-parameters.
checkpoint_dir: where the checkpoints live and where `done` is found.
model_dir: where to dump eval TensorBoard logs.
input_fn: for `Estimator`.
model_fn: for `Estimator`.
"""
estimator = utils.create_estimator(params, model_dir, model_fn)
eval_hooks = []
prev_checkpoint = None
num_mins_waited = 0
while True:
curr_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
if curr_checkpoint is not None and curr_checkpoint != prev_checkpoint:
tf.logging.info('Eval at {0}'.format(curr_checkpoint))
checkpoint_number = int(
curr_checkpoint.split('/')[-1].split('-')[-1])
if (checkpoint_number >= params.start_moving_average
and not params.moving_average):
tf.logging.info('From now on, use moving average for eval')
del estimator
params.set_hparam('moving_average', True)
estimator = utils.create_estimator(params, model_dir, model_fn)
try:
results = estimator.evaluate(input_fn=input_fn,
steps=params.num_eval_steps,
hooks=eval_hooks,
checkpoint_path=curr_checkpoint)
log_ppl = results['log_ppl/{0}'.format(params.task_mode)]
ppl = np.exp(log_ppl)
tf.logging.info('Eval step={0} {1}_ppl={2:<.2f}'.format(
results['global_step'], params.task_mode, ppl))
except Warning as w:
tf.logging.info(w)
except Exception: # pylint: disable=broad-except
traceback.print_exc()
tf.logging.info('Eval failed. Retrying...')
continue
prev_checkpoint = curr_checkpoint
num_mins_waited = 0
elif gfile.Exists(os.path.join(checkpoint_dir, 'done')):
tf.logging.info('Finished')
sys.exit(0)
else:
time.sleep(30)
num_mins_waited += 0.5
tf.logging.info('Waited {0:<.1f} mins'.format(num_mins_waited))
if num_mins_waited >= 120:
sys.exit(0)
def train_driver(params, model_dir, input_fn, model_fn):
"""What we will do for training."""
estimator = utils.create_estimator(params, model_dir, model_fn)
tf.logging.info('Train for {0} steps.'.format(params.num_train_steps))
for trial_id in range(MAX_RETRIES):
try:
estimator.train(input_fn=input_fn,
max_steps=params.num_train_steps)
break
except Warning as w:
tf.logging.info(w)
except Exception as e: # pylint: disable=broad-except
tf.logging.info(e)
traceback.print_exc()
tf.logging.info('Failed {0} times. Retry!'.format(trial_id + 1))
continue
else:
with gfile.GFile(os.path.join(model_dir, 'done'), 'w') as fout:
fout.write('Job failed after {0} retries'.format(MAX_RETRIES))
fout.flush()
sys.exit(1)
with gfile.GFile(os.path.join(model_dir, 'done'), 'w') as fout:
fout.write('Job finished')
fout.flush()