在 tensorflow 的推理过程中如何获得神经元的激活？

Question

我想具体知道如何让神经网络中的神经元被激活（激活函数后每个神经元的输出）

当我在 Tensorflow 2 中的模型推理过程中提供输入时，如何获得顺序模型的所有神经元的激活值？

Answer 1

尝试这样的事情：

intermediate_output = tf.keras.Model(model.input, 
                                     model.get_layer('conv2_block1_3_conv').output)

您可以使用 model.layers 获取图层列表并选择您想要的图层。

list(map(lambda x: x.name, model.layers))

['input_2',
 'conv1_pad',
 'conv1_conv',
 'conv1_bn',
 'conv1_relu',
 'pool1_pad',
 'pool1_pool',
 'conv2_block1_1_conv',
 'conv2_block1_1_bn',
 'conv2_block1_1_relu',
 'conv2_block1_2_conv',
 'conv2_block1_2_bn',
 'conv2_block1_2_relu',
 'conv2_block1_0_conv',
 'conv2_block1_3_conv', ...

完整示例：

import tensorflow as tf
from skimage import data
import matplotlib.pyplot as plt

model = tf.keras.applications.resnet.ResNet50(
    weights='imagenet', include_top=True)

model.build(input_shape=(None, 224, 224, 3))

image = tf.image.resize(data.chelsea(), (224, 224))/255

intermediate_output = tf.keras.Model(model.input, 
                                    model.get_layer('conv2_block1_3_conv').output)

extracted = intermediate_output(image[None, ...])

<tf.Tensor: shape=(1, 56, 56, 256), dtype=float32, numpy=
array([[[[ 0.23296243, -0.19640587, -0.8846314 , ..., -0.3091477 ,
          -0.51000404, -0.00218517],
         [ 0.27896926, -0.22646117, -0.91138077, ..., -0.4151363 ,
          -0.73324907,  0.05706196],
         [ 0.27908558, -0.2267507 , -0.9121696 , ..., -0.41521263,
          -0.73362166,  0.05721636],
         [ 0.04438811,  0.49058744, -1.5047315 , ..., -0.15204512,
          -1.2029954 , -0.29269713],
         [ 0.04450325,  0.4905177 , -1.505692  , ..., -0.15128748,
          -1.2025517 , -0.29254213],
         [ 0.05638191,  0.22808033, -1.5240382 , ...,  0.0052015 ,
          -0.8789809 , -0.19639899]]]], dtype=float32)>

在 tensorflow 的推理过程中如何获得神经元的激活？

How can you get the activations of the neurons during an inference in tensorflow?

tensorflow

neural-network

activation-function