TF.2 model(training=True) prediction is changed when apply_gradient is not applied

Question

I have a question about TF.2 trainable variables..

When I checked the model weights in loops, I found that loss is changed despite of not applying apply_gradient function.

So I stored the variables and loss but I didn't find the weights are changed in loops. Can you explain that for me..?

(That means.. same input/weight -> predictions are different.... T.T )

### model  ###
class MyModel(Model):
  def __init__(self):
    super(MyModel, self).__init__()
    #self.flatten = Flatten(input_shape=(32, 28, 28)) # flatten, input  
    #self.input_layer = Input( shape=(32, 784) )
    
    self.hidden_256_1 = Dense(256, activation='relu', name='hidden1')
    self.hidden_256_2 = Dense(256, activation='relu', name='hidden2')
    self.drop_out = Dropout(0.5)
    self.dense_final = Dense(10)

  def call(self, x):
    # x = self.flatten(x)
    # input_x = self.input_layer(x)
    x = self.hidden_256_1(x)
    x = self.drop_out(x)
    x = self.hidden_256_2(x)
    x = self.drop_out(x)

    return self.dense_final(x)

model = MyModel()

### print loss from same input/weight ###
vars_list = []
for i in range(5):
    predictions = model(x_train[:10].reshape(10,-1), training=True)
    loss = loss_object(y_train[:10], predictions)

    vars = model.trainable_weights
    vars_list.append( vars )
    print(loss)

tf.Tensor(2.3669665, shape=(), dtype=float32)
tf.Tensor(2.4489748, shape=(), dtype=float32)
tf.Tensor(2.4756782, shape=(), dtype=float32)
tf.Tensor(2.3325222, shape=(), dtype=float32)
tf.Tensor(2.341004, shape=(), dtype=float32)

# print weights of loop 1
print(vars_list[0][4])

<tf.Variable 'my_model/dense_3/kernel:0' shape=(256, 10) dtype=float32, numpy=
array([[-0.09301923,  0.03969355, -0.01272634, ..., -0.0475869 ,
         0.00238203,  0.0469147 ],
       [ 0.03952365, -0.08654938,  0.10042584, ..., -0.0594193 ,
        -0.01550625, -0.14284368],
       [-0.06716147,  0.07387908,  0.07724436, ...,  0.08974642,
        -0.11946932, -0.0196999 ],
       ...,
       [-0.12894998,  0.10549605, -0.04042872, ..., -0.07101952,
         0.14050862, -0.13851884],
       [ 0.05723909, -0.00989152,  0.02192681, ...,  0.11859939,
         0.06149144,  0.1294086 ],
       [ 0.0371833 ,  0.09864074,  0.04617405, ..., -0.01336585,
         0.10135004, -0.07977842]], dtype=float32)>

### print weights of loop 2

print(vars_list[1][4])

<tf.Variable 'my_model/dense_3/kernel:0' shape=(256, 10) dtype=float32, numpy=
array([[-0.09301923,  0.03969355, -0.01272634, ..., -0.0475869 ,
         0.00238203,  0.0469147 ],
       [ 0.03952365, -0.08654938,  0.10042584, ..., -0.0594193 ,
        -0.01550625, -0.14284368],
       [-0.06716147,  0.07387908,  0.07724436, ...,  0.08974642,
        -0.11946932, -0.0196999 ],
       ...,
       [-0.12894998,  0.10549605, -0.04042872, ..., -0.07101952,
         0.14050862, -0.13851884],
       [ 0.05723909, -0.00989152,  0.02192681, ...,  0.11859939,
         0.06149144,  0.1294086 ],
       [ 0.0371833 ,  0.09864074,  0.04617405, ..., -0.01336585,
         0.10135004, -0.07977842]], dtype=float32)>