如何在循环中检索中间预测结果

Question

需要计算每一层预测结果的时间，并将模型的输出作为另一层的输入。将每一层视为一个模型并获得中间结果并将这些结果传递给下一层。当我尝试运行时，会引发代码断言错误。

   model = vgg16.VGG16(weights='imagenet')
   filename = 'I:/PhD/SecondYear/Experiment/DNN/image/cat.4001.jpg'
   original_image = load_img(filename, target_size=(224, 224))
   numpy_image = img_to_array(original_image)
   input_image = np.expand_dims(numpy_image, axis=0)

   processed_image_vgg16 = vgg16.preprocess_input(input_image.copy())
   new_input=keras.layers.Input(shape=(224, 224, 3))
   input=new_input

   def profiling(model, test_input):
   next_layer = input
   for layer in model.layers:       
     out= layer(next_layer)
     #next_layer = out
     intermediate_model= keras.Model(next_layer,out)
     start = time.time()
     x = intermediate_model.predict(test_input)
     next_layer = out
     end = time.time() - start
     print(end)

profiling(model,processed_image_vgg16)

Answer 1

我对你的功能做了细微的改动：

def profiling(model, test_input):
  data_input = test_input
  for layer in model.layers: 

    start = time.time()
    im_input = tf.keras.layers.Input(shape=layer.input.shape[1:])
    # Get current layer's input shape and create input tensor.
    im_output = layer(im_input)
    # Apply current layer's operation on current input.
    intermediate_model= keras.models.Model(im_input,im_output) 
    # Create model from input and output 
    data_input = intermediate_model.predict(data_input)
    # predict and update output to data_input variable which will be used in next iteration. 
    end = time.time() - start
    print("Layer: ",type(layer).__name__,end)

profiling(model,processed_image_vgg16)

# output:
# Layer:  InputLayer 0.026311635971069336
# Layer:  Conv2D 0.048013925552368164
# Layer:  Conv2D 0.11106109619140625
# Layer:  MaxPooling2D 0.03328657150268555
# ...

这将为您提供分层子模型的预测时间。

如何在循环中检索中间预测结果

How to retrieve intermediate predictions results in a loop

python-3.x

keras

keras-layer

keras-2

tf.keras