def build_model(self):
"""Builds video ssl pretraining model."""
common_input_shape = [
d1 if d1 == d2 else None
for d1, d2 in zip(self.task_config.train_data.feature_shape,
self.task_config.validation_data.feature_shape)
]
input_specs = tf.keras.layers.InputSpec(shape=[None] + common_input_shape)
logging.info('Build model input %r', common_input_shape)
model = factory_3d.build_model(
self.task_config.model.model_type,
input_specs=input_specs,
model_config=self.task_config.model,
num_classes=self.task_config.train_data.num_classes)
return model
def build_model(self):
"""Builds video classification model."""
common_input_shape = self._get_feature_shape()
input_specs = tf.keras.layers.InputSpec(shape=[None] +
common_input_shape)
logging.info('Build model input %r', common_input_shape)
l2_weight_decay = self.task_config.losses.l2_weight_decay
# Divide weight decay by 2.0 to match the implementation of tf.nn.l2_loss.
# (https://www.tensorflow.org/api_docs/python/tf/keras/regularizers/l2)
# (https://www.tensorflow.org/api_docs/python/tf/nn/l2_loss)
l2_regularizer = (tf.keras.regularizers.l2(l2_weight_decay / 2.0)
if l2_weight_decay else None)
model = factory_3d.build_model(self.task_config.model.model_type,
input_specs=input_specs,
model_config=self.task_config.model,
num_classes=self._get_num_classes(),
l2_regularizer=l2_regularizer)
return model