recognition_model.py 文件源码

def create_base(self, inputs, is_training):
        """Creates a base part of the Model (no gradients, losses or summaries)."""

        with tf.name_scope('Model'):
            with slim.arg_scope([slim.fully_connected], activation_fn=tf.nn.relu,
                                # weights_regularizer=slim.l2_regularizer(0.01),
                                # weights_initializer=initializers.xavier_initializer(seed=self._config.random_seed),
                                # biases_initializer=tf.constant_initializer(0.1)
                                ):
                # first fully connected layer
                net = slim.fully_connected(inputs, self._config.mlp_params['hidden_sizes'][0], scope='fc1')

                # dropout1
                net = slim.dropout(net, self._config.keep_prob, is_training=is_training, scope='dropout1')

                # second fully connected layer
                net = slim.fully_connected(net, self._config.mlp_params['hidden_sizes'][1], scope='fc2')

                # dropout2
                net = slim.dropout(net, self._config.keep_prob, is_training=is_training,  scope='dropout2')

                # final fully-connected dense layer
                logits = slim.fully_connected(net, self._config.num_classes, activation_fn=None, scope='fc3')

                with tf.name_scope('output'):
                    predicted_classes = tf.to_int32(tf.argmax(logits, dimension=1), name='y')

        return logits, predicted_classes