def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
indices = search_space_utils.parse_list(FLAGS.indices, int)
ssd = FLAGS.ssd
cost = mobile_cost_model.estimate_cost(indices, ssd)
print('estimated cost: {:f}'.format(cost))
def test_parse_list(self):
self.assertEqual([], search_space_utils.parse_list('', str))
self.assertEqual([], search_space_utils.parse_list(' \t', str))
self.assertEqual(['hello'],
search_space_utils.parse_list('hello', str))
self.assertEqual(['he', 'lo'],
search_space_utils.parse_list('he:lo', str))
self.assertEqual([42], search_space_utils.parse_list('42', int))
self.assertEqual([4, 2], search_space_utils.parse_list('4:2', int))
self.assertEqual([1, 2, 3],
search_space_utils.parse_list('1:2:3', int))
self.assertAllClose([1.5, 2.5],
search_space_utils.parse_list('1.5:2.5', float))
def main(argv):
del argv # Unused.
if FLAGS.use_held_out_test_set:
default_epochs = 360
train_dataset_size = fast_imagenet_input.dataset_size_for_mode('train')
else:
default_epochs = 90
train_dataset_size = fast_imagenet_input.dataset_size_for_mode('l2l_train')
epochs = FLAGS.epochs
if epochs is None:
epochs = default_epochs
weight_decay = FLAGS.weight_decay
if weight_decay is None:
weight_decay = 3e-5 if FLAGS.use_bfloat16 else 4e-5
dropout_rate = FLAGS.dropout_rate
if dropout_rate is None:
# Select a dropout rate automatically. For MnasNet-sized models, dropout
# substantially improves accuracy when training for 360 epochs, but we
# haven't investigated whether it helps when training for 90 epochs. We
# currently enable dropout only for long training runs.
if epochs < 150:
# Disable dropout when training for less than 150 epochs.
dropout_rate = 0
elif FLAGS.ssd in mobile_search_space_v3.MOBILENET_V3_LIKE_SSDS:
# MobileNetV3-based search space, training for at least 150 epochs.
dropout_rate = 0.25
else:
# MobileNetV2-based search space, training for at least 150 epochs.
dropout_rate = 0.15
max_global_step = train_dataset_size * epochs // FLAGS.train_batch_size
params = {
'checkpoint_dir': FLAGS.checkpoint_dir,
'dataset_dir': FLAGS.dataset_dir,
'learning_rate': FLAGS.base_learning_rate * FLAGS.train_batch_size / 256,
'tpu': FLAGS.tpu,
'tpu_zone': FLAGS.tpu_zone,
'gcp_project': FLAGS.gcp_project,
'momentum': FLAGS.momentum,
'weight_decay': weight_decay,
'max_global_step': max_global_step,
'warmup_steps': int(FLAGS.warmup_steps_fraction * max_global_step),
'tpu_iterations_per_loop': FLAGS.tpu_iterations_per_loop,
'train_batch_size': FLAGS.train_batch_size,
'eval_batch_size': FLAGS.eval_batch_size,
'indices': search_space_utils.parse_list(FLAGS.indices, int),
'use_held_out_test_set': FLAGS.use_held_out_test_set,
'use_bfloat16': FLAGS.use_bfloat16,
'filters_multiplier': FLAGS.filters_multiplier,
'dropout_rate': dropout_rate,
'path_dropout_rate': FLAGS.path_dropout_rate,
'ssd': FLAGS.ssd,
}
if FLAGS.mode in ['train', 'train_and_eval']:
_write_params_to_checkpoint_dir(params)
run_model(params, FLAGS.mode)