Weights were not updated using Gradient Tape and apply_gradients()

Question

I am building a DNN with a custom loss function and I am training this DNN using Gradient Tape in TensorFlow.kerasenter code here. The code runs without any errors, however, as far as I can check the weights of the DNN, the weights were not being updated at all. I followed exactly what recommends from the TensorFlow website and search for the answers but still don't understand what is the reason. Here is my code:

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))

I have checked the weights of the model using model.get_weights() and they look exactly the same before and after running the loop. So what is the problem here? And one more question, how can I print out the loss for every epoch?

Answer 1

The weight does change. You can check as follows; after building the model save your weights file (these are initial weight).

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

Now, run your training loop. After finishing the training, get the new weights as follows and compare before and after.

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,

And this is how you can print your loss score in every epoch.

# 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