layer.get_weights() is not equal in the same model layers-CodePudding

Why not all the layer weights equal:

import tensorflow as tf
import numpy as np

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(4, kernel_size=3, padding="same", 
                           activation="selu", input_shape=(28,28,1)
    ),
    tf.keras.layers.MaxPool2D(pool_size=2),
    tf.keras.layers.Conv2D(8, kernel_size=3, padding="same"),
    tf.keras.layers.Conv2D(16, kernel_size=3, padding="same"),
    tf.keras.layers.MaxPool2D(pool_size=2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10)
])

model.summary()

a = []
b = []

for layer in model.layers:
    weights = layer.get_weights()
    a.append(weights)
    b.append(weights)

for index, i in enumerate(a):
    a_weights = a[index]
    b_weights = b[index]
    print("index [ {} ]: a_weights == b_weights: {}".format(
        index,
        np.array_equal(a_weights, b_weights)
    ))

Here is the output:

Model: "sequential_10"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_15 (Conv2D)          (None, 28, 28, 4)         40        
                                                                 
 max_pooling2d_15 (MaxPoolin  (None, 14, 14, 4)        0         
 g2D)                                                            
                                                                 
 conv2d_16 (Conv2D)          (None, 14, 14, 8)         296       
                                                                 
 conv2d_17 (Conv2D)          (None, 14, 14, 16)        1168      
                                                                 
 max_pooling2d_16 (MaxPoolin  (None, 7, 7, 16)         0         
 g2D)                                                            
                                                                 
 flatten_10 (Flatten)        (None, 784)               0         
                                                                 
 dense_20 (Dense)            (None, 128)               100480    
                                                                 
 dense_21 (Dense)            (None, 10)                1290      
                                                                 
=================================================================
Total params: 103,274
Trainable params: 103,274
Non-trainable params: 0
_________________________________________________________________
index [ 0 ]: a_weights == b_weights: False
index [ 1 ]: a_weights == b_weights: True
index [ 2 ]: a_weights == b_weights: False
index [ 3 ]: a_weights == b_weights: False
index [ 4 ]: a_weights == b_weights: True
index [ 5 ]: a_weights == b_weights: True
index [ 6 ]: a_weights == b_weights: False
index [ 7 ]: a_weights == b_weights: False

The a_weights == b_weights are not all the "True".

What's the problem?

CodePudding user response：

Notice that the only time a_weights == b_weights is True, is when you are referencing a layer, which does not have any weights. np.array_equal is returning False because you are actually comparing lists of arrays and not the arrays themselves. Each trainable layer of yours has a kernel weight tensor and bias tensor. So actually your code should look like this:

import tensorflow as tf
import numpy as np

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(4, kernel_size=3, padding="same", 
                           activation="selu", input_shape=(28,28,1)
    ),
    tf.keras.layers.MaxPool2D(pool_size=2),
    tf.keras.layers.Conv2D(8, kernel_size=3, padding="same"),
    tf.keras.layers.Conv2D(16, kernel_size=3, padding="same"),
    tf.keras.layers.MaxPool2D(pool_size=2),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(10)
])

model.summary()

a = []
b = []

for layer in model.layers:
    weights = layer.get_weights()
    a.append(weights)
    b.append(weights)

for index, i in enumerate(a):
    a_weights = a[index]
    b_weights = b[index]
    if a_weights:
      print("index [ {} ]: kernel: a_weights == b_weights: {}".format(
          index,
          np.array_equal(a_weights[0], b_weights[0])
      ))
      print("index [ {} ]: bias: a_weights == b_weights: {}".format(
          index,
          np.array_equal(a_weights[1], b_weights[1])
      ))
    else:
      print("index [ {} ]: no layer weights: a_weights == b_weights: {}".format(
          index,
          np.array_equal(a_weights, b_weights)
      ))

...
index [ 0 ]: kernel: a_weights == b_weights: True
index [ 0 ]: bias: a_weights == b_weights: True
index [ 1 ]: no layer weights: a_weights == b_weights: True
index [ 2 ]: kernel: a_weights == b_weights: True
index [ 2 ]: bias: a_weights == b_weights: True
index [ 3 ]: kernel: a_weights == b_weights: True
index [ 3 ]: bias: a_weights == b_weights: True
index [ 4 ]: no layer weights: a_weights == b_weights: True
index [ 5 ]: no layer weights: a_weights == b_weights: True
index [ 6 ]: kernel: a_weights == b_weights: True
index [ 6 ]: bias: a_weights == b_weights: True
index [ 7 ]: kernel: a_weights == b_weights: True
index [ 7 ]: bias: a_weights == b_weights: True

You could also use a_weights == b_weights directly.