如何在功能性 tensorflow ResNet50 model 中添加一層？

Question

我正在構建一個圖像標題 model，為此我使用ResNet50作為特征提取 model。我已經編寫了代碼，它工作正常：

rs50 = tf.keras.applications.ResNet50(include_top = False, weights = 'imagenet', input_shape = (224, 224, 3))
new_input = rs50.input
hidden_layer = rs50.layers[-1].output

feature_extract = tf.keras.Model(new_input, hidden_layer)

以下是 model 摘要 ( feature_extract.summary() ) 的最后幾行：

 conv5_block3_3_bn (BatchNormal  (None, 7, 7, 2048)  8192        ['conv5_block3_3_conv[0][0]']    
 ization)                                                                                         
                                                                                                  
 conv5_block3_add (Add)         (None, 7, 7, 2048)   0           ['conv5_block2_out[0][0]',       
                                                                  'conv5_block3_3_bn[0][0]']      
                                                                                                  
 conv5_block3_out (Activation)  (None, 7, 7, 2048)   0           ['conv5_block3_add[0][0]']       
                                                                                                  
==================================================================================================
Total params: 23,587,712
Trainable params: 23,534,592
Non-trainable params: 53,120

但是，問題是它生成了 2048 個特征。 我沒有那么多 memory，所以我想將其(None, 7, 7, 2048)更改為(None, 7, 7, 1024)

我該怎么做？

Answer 1

一種方法是找到具有 output 形狀(None, 14, 14, 1024)的最后一層，然后提取模型的層直到該點。 conv4_block6_out層恰好是最后一個塊開始之前的最后一層。 這樣，最后一個塊被完全跳過，節省了更多 memory。然后，應用一個或多個Conv2D或MaxPooling層來獲得形狀(None, 7, 7, 1024) ：

import tensorflow as tf
rs50 = tf.keras.applications.ResNet50(include_top = False, weights = 'imagenet', input_shape = (224, 224, 3))

index = 0
for i, l in enumerate(rs50.layers):
  if 'conv4_block6_out' in l.name:
    index = i

new_input = rs50.input
hidden_layer = rs50.layers[index].output
output = tf.keras.layers.Conv2D(1024, kernel_size=8)(hidden_layer)

feature_extract = tf.keras.Model(new_input, output)
print(feature_extract.output)

KerasTensor(type_spec=TensorSpec(shape=(None, 7, 7, 1024), dtype=tf.float32, name=None), name='conv2d_4/BiasAdd:0', description="created by layer 'conv2d_4'")