模型鏈接：https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/efficientnet_model.py

Head

分類頭
build：

	#激活函式
	self._relu_fn = (self._block_args.activation_fn
                     or global_params.relu_fn or tf.nn.swish)
 # Head part.
 	#卷積層定義
    self._conv_head = utils.Conv2D(
        filters=round_filters(1280, self._global_params, self._fix_head_stem),
        kernel_size=[1, 1],#1x1
        strides=[1, 1],
        kernel_initializer=conv_kernel_initializer,
        padding='same',
        data_format=self._global_params.data_format,
        use_bias=False)
    #bn定義
    self._bn1 = self._batch_norm(
        axis=channel_axis,
        momentum=batch_norm_momentum,
        epsilon=batch_norm_epsilon)
	#平均池化定義
    self._avg_pooling = tf.keras.layers.GlobalAveragePooling2D(
        data_format=self._global_params.data_format)
    #全連接層定義
    if self._global_params.num_classes:
      self._fc = tf.layers.Dense(
          self._global_params.num_classes,
         kernel_initializer=dense_kernel_initializer)
    else:
      self._fc = None
	#dropout定義
    if self._global_params.dropout_rate > 0:
      self._dropout = tf.keras.layers.Dropout(self._global_params.dropout_rate)
    else:
      self._dropout = None

前向傳播：

    if not features_only:#不只是特征（后續分類）
      # Calls final layers and returns logits.
      with tf.variable_scope('head'):
        #conv-bn1-激活(relu/swish) 
        outputs = self._relu_fn(
            self._bn1(self._conv_head(outputs), training=training))
        self.endpoints['head_1x1'] = outputs
		#全域平均池化-dropout-fc
		#tf的池化        
        if self._global_params.local_pooling:
          shape = outputs.get_shape().as_list()
          kernel_size = [
              1, shape[self._spatial_dims[0]], shape[self._spatial_dims[1]], 1]
          outputs = tf.nn.avg_pool(
              outputs, ksize=kernel_size, strides=[1, 1, 1, 1], padding='VALID')
          self.endpoints['pooled_features'] = outputs
          #dropout
          if not pooled_features_only:
            if self._dropout:
              outputs = self._dropout(outputs, training=training)
            self.endpoints['global_pool'] = outputs
            #全連接 輸出各類
            if self._fc:
              outputs = tf.squeeze(outputs, self._spatial_dims)
              outputs = self._fc(outputs)
            self.endpoints['head'] = outputs
        #自己的池化
        else:
          outputs = self._avg_pooling(outputs)
          self.endpoints['pooled_features'] = outputs
          if not pooled_features_only:
            #dropout
            if self._dropout:
              outputs = self._dropout(outputs, training=training)
            self.endpoints['global_pool'] = outputs
            #fc
            if self._fc:
              outputs = self._fc(outputs)
            self.endpoints['head'] = outputs
    return outputs