问题描述
我已经加载了训练有素的模型，并且试图在第一个Conv2D层上进行激活。问题是，当我尝试预测该层的输出时，显然会发现卷积中的某些滤波器（大多数）的输出不正确，而对于其中一些滤波器，它会根据权重为我提供正确的输出。 / p>
为了说明这一点，我只向网络输入一个全零的输入，因此，我希望只得到每个层的偏差。代码如下：
import os
import tensorflow
import numpy as np
# Load Trained model
model = tensorflow.keras.models.load_model('viewmodel.h5')
model.summary()
# Choose layer
layer_name = 'conv2d'
layer = model.get_layer(name=layer_name)
feature_extractor = tensorflow.keras.Model(inputs=model.inputs,outputs=layer.output)
# Bias of my feature extractor
b_f = feature_extractor.get_weights()[1]
# Predict output (It is supposed to get all bias)
input_map = np.zeros((1,70,56,2)) # tensorflow.zeros((1,2)) yields the same problem
activation = feature_extractor.predict(input_map)
activation = activation.reshape(68,54,32) # Remove the 1 in front (shape is (1,68,32)
print('--------------------------------------------------')
print('Bias 0 is: {}'.format(b_f[0]))
print('Activation in filter 0 is: ')
print(activation[:,:,0]) # First filter
print('--------------------------------------------------')
print('Bias 9 is: {}'.format(b_f[9]))
print('Activation in filter 9 is: ')
print(activation[:,9]) # Tenth filter
print('--------------------------------------------------')
但是我的结果是：
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_1 (InputLayer) [(None,2)] 0
_________________________________________________________________
conv2d (Conv2D) (None,32) 608
_________________________________________________________________
average_pooling2d (AveragePo (None,34,27,32) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None,32,25,16) 4624
_________________________________________________________________
conv2d_2 (Conv2D) (None,30,23,16) 2320
_________________________________________________________________
average_pooling2d_1 (Average (None,15,11,16) 0
_________________________________________________________________
flatten (Flatten) (None,2640) 0
_________________________________________________________________
dense (Dense) (None,150) 396150
_________________________________________________________________
dropout (Dropout) (None,150) 0
_________________________________________________________________
dense_1 (Dense) (None,1) 151
=================================================================
Total params: 403,853
Trainable params: 403,853
Non-trainable params: 0
_________________________________________________________________
WARNING:tensorflow:10 out of the last 11 calls to <function Model.make_predict_function.<locals>.predict_function at 0x7ff629d1f0d0> triggered tf.function retracing. Tracing is expensive and the excessive number of tracings Could be due to (1) creating @tf.function repeatedly in a loop,(2) passing tensors with different shapes,(3) passing Python objects instead of tensors. For (1),please define your @tf.function outside of the loop. For (2),@tf.function has experimental_relax_shapes=True option that relaxes argument shapes that can avoid unnecessary retracing. For (3),please refer to https://www.tensorflow.org/tutorials/customization/performance#python_or_tensor_args and https://www.tensorflow.org/api_docs/python/tf/function for more details.
--------------------------------------------------
Bias 0 is: -0.7553322911262512
Activation in filter 0 is:
[[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]
...
[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]
[0. 0. 0. ... 0. 0. 0.]]
--------------------------------------------------
Bias 9 is: 0.42768770456314087
Activation in filter 9 is:
[[0.4276877 0.4276877 0.4276877 ... 0.4276877 0.4276877 0.4276877]
[0.4276877 0.4276877 0.4276877 ... 0.4276877 0.4276877 0.4276877]
[0.4276877 0.4276877 0.4276877 ... 0.4276877 0.4276877 0.4276877]
...
[0.4276877 0.4276877 0.4276877 ... 0.4276877 0.4276877 0.4276877]
[0.4276877 0.4276877 0.4276877 ... 0.4276877 0.4276877 0.4276877]
[0.4276877 0.4276877 0.4276877 ... 0.4276877 0.4276877 0.4276877]]
--------------------------------------------------
我希望过滤器0充满其偏置（-0.7553322911262512），但这不是正在发生的事情...我也尝试过馈入tf.zeros ...也许是由于每次警告我执行代码后，数字增加了1（不知道警告的含义是什么，或者与我尝试的运气不符的问题有关，因为我似乎理解这只会影响我的速度代码？）
在这里我会很感激的：）
解决方法
暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！
如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。
小编邮箱:dio#foxmail.com (将#修改为@）