def main():
tf.logging.set_verbosity(tf.logging.INFO)
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--depth-multiplier', default=0.5, type=float)
parser.add_argument('--scale-depthwise', default=1.0, type=float)
parser.add_argument('--scale-pointwise', default=1.0, type=float)
parser.add_argument('--scale-dense', default=1.0, type=float)
parser.add_argument('--adj-power', default=None, type=float)
parser.add_argument('--cutoff', default=None, type=float)
parser.add_argument('--checkpoint-path', default=None, type=str)
parser.add_argument('--train-dir', default=None, type=str)
parser.add_argument('--dataset', default=None, type=str)
args = parser.parse_args()
scale = _make_scale(args.scale_depthwise,
args.scale_pointwise,
args.scale_dense,
args.adj_power,
args.cutoff,
args.depth_multiplier)
network_fn = partial(mobilenet_v1,
num_classes=1001,
depth_multiplier=args.depth_multiplier,
scope=_make_scope(scale))
model_fn = make_model_fn(network_fn, None)
input_fn = make_input_fn(imagenet_data.get_split(args.dataset),
inception_preprocessing.preprocess_image,
n_epochs=1, image_size=224, batch_size=64)
estimator = tf.estimator.Estimator(model_fn, model_dir=args.train_dir)
estimator.evaluate(input_fn, checkpoint_path=args.checkpoint_path)
def evaluate(args, model_fn):
input_fn = make_input_fn(imagenet_data.get_split(args.dataset_path,
shuffle=False),
partial(inception_preprocessing.preprocess_image,
is_training=False),
batch_size=64,
image_size=224)
estimator = tf.estimator.Estimator(model_fn, model_dir=args.train_dir)
estimator.evaluate(input_fn)
def train(args, model_fn):
input_fn = make_input_fn(imagenet_data.get_split(args.dataset_path, shuffle=True),
partial(inception_preprocessing.preprocess_image, is_training=False),
batch_size=64, image_size=224)
estimator = tf.estimator.Estimator(model_fn, model_dir=args.train_dir,
params={'warm_start': args.warm_start})
estimator.train(input_fn, max_steps=args.max_steps,
saving_listeners=[
CommitMaskedValueHook(args.max_steps,
variables_fn=lambda: tf.get_collection(tf.GraphKeys.WEIGHTS))
])
def evaluate(args, model_fn):
input_fn = make_input_fn(imagenet_data.get_split(args.dataset_path, shuffle=False),
partial(inception_preprocessing.preprocess_image, is_training=False),
image_size=224,
n_epochs=1)
estimator = tf.estimator.Estimator(model_fn, model_dir=args.train_dir)
estimator.evaluate(input_fn)
total = 0
for v in tf.all_variables():
total += np.prod(v.get_shape().as_list())
print (total)
print ((total * 4) / (1024 ** 2))
def train(args, model_fn):
input_fn = make_input_fn(imagenet_data.get_split(args.dataset_path,
shuffle=True),
partial(inception_preprocessing.preprocess_image,
is_training=False),
batch_size=64,
image_size=224)
estimator = tf.estimator.Estimator(model_fn,
model_dir=args.train_dir,
params={'warm_start': args.warm_start})
estimator.train(
input_fn,
hooks=[
ExecuteAtSessionCreateHook(op_fn=_quantize_init_op,
name='QuantizeInitHook')
],
saving_listeners=[
quantization.CommitQuantizedValueHook(
args.max_steps,
variables_fn=lambda: tf.get_collection(tf.GraphKeys.WEIGHTS))
])