未使用 Gradient Tape 和 apply_gradients() 更新权重

问题描述

我正在构建一个带有自定义损失函数的 DNN，我正在使用 TensorFlow.kerasenter code here 中的 Gradient Tape 训练这个 DNN。代码运行没有任何错误，但是，就我可以检查 DNN 的权重而言，权重根本没有更新。我完全遵循 TensorFlow 网站上的建议并搜索答案，但仍然不明白是什么原因。这是我的代码：

import numpy as np

import tensorflow as tf
from tensorflow.keras.layers import Input,Dense,LeakyReLU,Concatenate
from tensorflow.keras.models import Model
from tensorflow.keras import backend as K
from tensorflow.keras import optimizers

# Generate a random train data
c0_train = np.array([30 * np.random.uniform() for i in range(10000)])

# Build a simple DNN
c0_input = Input(shape=(1,),name='c0')
hidden_1 = Dense(100)(c0_input)
activation_1 = LeakyReLU(alpha=0.1)(hidden_1)
hidden_2 = Dense(100)(activation_1)
activation_2 = LeakyReLU(alpha=0.1)(hidden_2)
hidden_3 = Dense(100)(activation_2)
activation_3 = LeakyReLU(alpha=0.1)(hidden_3)
x0_output = Dense(1,name='x0')(activation_3)

model = Model(inputs=c0_input,outputs=x0_output)

# Calculating the loss function 
def cal_loss(c0_input):
  x0_output = model(c0_input)
  loss = tf.reduce_mean(
      tf.multiply(c0_input,tf.square(tf.subtract(x0_output,c0_input))))
  return loss

# Compute the gradient calculation
@tf.function
def compute_loss_grads(c0_input):
  with tf.GradientTape() as tape:
    loss = cal_loss(c0_input)
  grads = tape.gradient(loss,model.trainable_variables)
  return loss,grads

# Optimizer
opt = optimizers.Adam(learning_rate=0.01)

# Start looping
for epoch in range(50):
  print('Epoch = ',epoch)
  # Compute the loss and gradients
  [loss,grads] = compute_loss_grads(tf.cast(c0_train,tf.float32))
  # Adjust the weights of the model
  opt.apply_gradients(zip(grads,model.trainable_variables))

我已经使用 model.get_weights() 检查了模型的权重，它们在运行循环之前和之后看起来完全一样。那么这里的问题是什么？还有一个问题，我如何打印出每个 epoch 的损失？

解决方法

权重确实发生了变化。您可以进行如下检查；建立模型后保存您的权重文件（这些是初始权重）。

model = Model(inputs=c0_input,outputs=x0_output)
a_weg = model.get_weights()

现在，运行您的训练循环。训练完成后，得到新的权重如下，对比前后对比。

b_weg = model.get_weights()

a_weg[:1]
[array([[ 0.03541631,-0.02134866,0.17080751,0.10538128,0.1361396,0.08645812,0.114059,0.216836,-0.22464292,-0.21979895,-0.23927784,-0.00685263,0.2167016,0.09989142,-0.17772573,0.16095945,-0.10120587,-0.22456157,-0.22947621,0.04009536,0.01029667,-0.18134505,-0.11318983,0.10220072,0.10100928,b_weg[:1]
[array([[ 0.05140253,0.00969543,0.15155758,0.07171137,0.15917814,0.10883425,0.11428417,0.17012525,-0.25049415,-0.20693016,-0.20231842,0.005939,0.19197173,0.07405043,-0.14260964,0.12490476,-0.11532102,-0.24605738,-0.25135723,0.01863468,0.0311144,-0.20050383,-0.11864465,0.07961675,0.11557189,

这就是您可以在每个时期打印损失分数的方式。

# Start looping
for epoch in range(5):
  # Compute the loss and gradients
  [loss,grads] = compute_loss_grads(tf.cast(c0_train,tf.float32))
  # Adjust the weights of the model
  opt.apply_gradients(zip(grads,model.trainable_variables))
  print('Epoch = ',epoch,' - loss = ',loss.numpy())

Epoch =  0  - loss =  5962.977
Epoch =  1  - loss =  3042.2874
Epoch =  2  - loss =  2877.9978
Epoch =  3  - loss =  2607.5347
Epoch =  4  - loss =  2173.3213