def main(args):
if args.dataset not in _available_datasets:
raise NotImplementedError
dataset = build_dataset(
name=args.dataset,
shape=[args.height, args.width],
)
model = build_mobilenetv3(
args.model_type,
input_shape=(args.height, args.width, dataset["channels"]),
num_classes=dataset["num_classes"],
width_multiplier=args.width_multiplier,
)
if args.optimizer not in _available_optimizers:
raise NotImplementedError
model.load_weights(args.model_path)
model.compile(
optimizer=_available_optimizers.get(args.optimizer)(args.lr),
loss="categorical_crossentropy",
metrics=["accuracy"],
)
model.evaluate(
dataset["test"].make_one_shot_iterator(),
steps=(dataset["num_test"] // args.valid_batch_size) + 1,
)
def main(args):
_available_datasets = {
"mnist": mnist,
"cifar10": cifar10,
}
if args.dataset not in _available_datasets:
raise NotImplementedError
train_data, num_train_data, test_data, num_test_data = _available_datasets.get(args.dataset)(
args.train_batch_size,
args.valid_batch_size,
args.height,
args.width,
)
model = build_mobilenetv3(
args.model_type,
input_shape=(args.height, args.width, args.channels),
num_classes=args.num_classes,
width_multiplier=args.width_multiplier,
l2_reg=args.l2_reg,
)
_available_optimizers = {
"rmsprop": tf.train.RMSPropOptimizer,
"adam": tf.train.AdamOptimizer,
"sgd": tf.train.GradientDescentOptimizer,
}
if args.optimizer not in _available_optimizers:
raise NotImplementedError
model.compile(
optimizer=_available_optimizers.get(args.optimizer)(args.lr),
loss="categorical_crossentropy",
metrics=["accuracy"],
)
callbacks = [
tf.keras.callbacks.TensorBoard(log_dir=args.logdir),
]
model.fit(
train_data.make_one_shot_iterator(),
steps_per_epoch=(num_train_data//args.train_batch_size)+1,
epochs=args.num_epoch,
validation_data=test_data.make_one_shot_iterator(),
validation_steps=(num_test_data//args.valid_batch_size)+1,
callbacks=callbacks,
)
model.save_weights(f"mobilenetv3_{args.model_type}_{args.dataset}_{args.num_epoch}.h5")
def model_fn(features, labels, mode, params):
model = build_mobilenetv3(model_type=params['model_type'],
input_shape=params['input_shape'],
num_classes=params['num_classes'],
batch_size=params['batch_size'])
predictions = model(features,
training=(mode == tf.estimator.ModeKeys.TRAIN))
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode, predictions=predictions)
if params['optimizer'] == 'adam':
optimizer = tf.compat.v1.train.AdamOptimizer()
elif params['optimizer'] == 'adagrad':
optimizer = tf.compat.v1.train.AdagradOptimizer()
else:
raise ValueError('Optimizer "{}" not supported.'.format(
params['optimizer']))
if params['use_tpu']:
optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer)
loss = tf.keras.losses.SparseCategoricalCrossentropy()(tf.argmax(
labels, 1), predictions)
train_op = optimizer.minimize(
loss,
var_list=model.trainable_variables,
global_step=tf.compat.v1.train.get_or_create_global_step())
def metric_fn(y_true, y_pred):
predicted_classes = tf.argmax(y_pred, 1)
accuracy = tf.compat.v1.metrics.accuracy(labels=tf.argmax(y_true, 1),
predictions=predicted_classes,
name="acc_op")
return {"accuracy": accuracy}
return tf.compat.v1.estimator.tpu.TPUEstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op,
eval_metrics=(metric_fn, [labels, predictions]))
def main(args):
_available_datasets = {
"mnist": mnist,
"cifar10": cifar10,
}
if args.dataset not in _available_datasets:
raise NotImplementedError
_, _, test_data, num_test_data = _available_datasets.get(args.dataset)(
args.valid_batch_size,
args.valid_batch_size,
args.height,
args.width,
)
model = build_mobilenetv3(
args.model_type,
input_shape=(args.height, args.width, args.channels),
num_classes=args.num_classes,
width_multiplier=args.width_multiplier,
)
_available_optimizers = {
"rmsprop": tf.train.RMSPropOptimizer,
"adam": tf.train.AdamOptimizer,
"sgd": tf.train.GradientDescentOptimizer,
}
if args.optimizer not in _available_optimizers:
raise NotImplementedError
model.load_weights(args.model_path)
model.compile(
optimizer=_available_optimizers.get(args.optimizer)(args.lr),
loss="categorical_crossentropy",
metrics=["accuracy"],
)
model.evaluate(
test_data.make_one_shot_iterator(),
steps=(num_test_data//args.valid_batch_size)+1,
)
def main(args):
if args.dataset not in _available_datasets:
raise NotImplementedError
dataset = build_dataset(name=args.dataset,
shape=(args.height, args.width),
train_batch_size=args.train_batch_size,
valid_batch_size=args.valid_batch_size)
model = build_mobilenetv3(
args.model_type,
input_shape=(args.height, args.width, dataset["channels"]),
num_classes=dataset["num_classes"],
width_multiplier=args.width_multiplier,
l2_reg=args.l2_reg,
)
if args.optimizer not in _available_optimizers:
raise NotImplementedError
model.compile(
optimizer=_available_optimizers.get(args.optimizer)(args.lr),
loss="categorical_crossentropy",
metrics=["accuracy"],
)
callbacks = [
tf.keras.callbacks.TensorBoard(log_dir=args.logdir),
]
model.fit(
dataset["train"].make_one_shot_iterator(),
steps_per_epoch=(dataset["num_train"] // args.train_batch_size) + 1,
epochs=args.num_epoch,
validation_data=dataset["test"],
validation_steps=(dataset["num_test"] // args.valid_batch_size) + 1,
callbacks=callbacks,
)
model.save_weights(
f"mobilenetv3_{args.model_type}_{args.dataset}_{args.num_epoch}.h5")