如何在訓練期間添加具有不同標准的高斯噪聲？

Question

我正在使用 keras 和 tensorflow 訓練 CNN。 我想在訓練期間將高斯噪聲添加到我的輸入數據中，並在進一步的步驟中降低噪聲的百分比。 我現在做什么，我使用：

from tensorflow.python.keras.layers import Input, GaussianNoise, BatchNormalization
inputs = Input(shape=x_train_n.shape[1:])
bn0 = BatchNormalization(axis=1, scale=True)(inputs)
g0 = GaussianNoise(0.5)(bn0) 

GaussianNoise 采用的變量是噪聲分布的標准偏差，我無法為其分配動態值，我如何添加例如噪聲，然后根據我所處的時代減少該值？

Answer 1

您可以簡單地設計一個自定義callback ，在訓練一個 epoch 之前更改stddev 。

參考：

https://www.tensorflow.org/api_docs/python/tf/keras/layers/GaussianNoise

https://www.tensorflow.org/guide/keras/custom_callback

from tensorflow.keras.layers import Input, Dense, Add, Activation
from tensorflow.keras.models import Model
import tensorflow as tf
import numpy as np
import random


from tensorflow.python.keras.layers import Input, GaussianNoise, BatchNormalization
inputs = Input(shape=100)
bn0 = BatchNormalization(axis=1, scale=True)(inputs)
g0 = GaussianNoise(0.5)(bn0) 
d0 = Dense(10)(g0)
model = Model(inputs, d0)

model.compile('adam', 'mse')
model.summary()


class MyCustomCallback(tf.keras.callbacks.Callback):

  def on_epoch_begin(self, epoch, logs=None):
    self.model.layers[2].stddev = random.uniform(0, 1)
    print('updating sttdev in training')
    print(self.model.layers[2].stddev)


X_train = np.zeros((10,100))
y_train = np.zeros((10,10))

noise_change = MyCustomCallback()
model.fit(X_train, 
          y_train, 
          batch_size=32, 
          epochs=5, 
          callbacks = [noise_change])

Model: "model_5"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_6 (InputLayer)         [(None, 100)]             0         
_________________________________________________________________
batch_normalization_5 (Batch (None, 100)               400       
_________________________________________________________________
gaussian_noise_5 (GaussianNo (None, 100)               0         
_________________________________________________________________
dense_5 (Dense)              (None, 10)                1010      
=================================================================
Total params: 1,410
Trainable params: 1,210
Non-trainable params: 200
_________________________________________________________________
Epoch 1/5
updating sttdev in training
0.984045691131548
1/1 [==============================] - 0s 1ms/step - loss: 1.6031
Epoch 2/5
updating sttdev in training
0.02821459469022025
1/1 [==============================] - 0s 742us/step - loss: 1.5966
Epoch 3/5
updating sttdev in training
0.6102984511769268
1/1 [==============================] - 0s 1ms/step - loss: 1.8818
Epoch 4/5
updating sttdev in training
0.021155188690323512
1/1 [==============================] - 0s 1ms/step - loss: 1.2032
Epoch 5/5
updating sttdev in training
0.35950227285165115
1/1 [==============================] - 0s 2ms/step - loss: 1.8817

<tensorflow.python.keras.callbacks.History at 0x7fc67ce9e668>