[SOLVED] "convolutional_block" gives different output at each run in the local environment

[Lingxiao_Gao]

Environment:
Ubuntu 21.10
TensorFlow: 2.8.0
NVIDIA-SMI 510.54
CUDA Version: 11.6

I am trying to reproduce C4-W2A1 in my own environment, and this happened:

[TMosh]

For the notebooks to work correctly, you have to be extremely careful that you’re using the same versions of all the tools that Coursera uses.

[TMosh]

Do you get similar results if you run the notebook in Coursera’s environment?

[Lingxiao_Gao]

I had all test passed in Coursera Environment

[ai_curious]

it seems to have something to do with reusing X1. Check this out.

np.random.seed(1)
X1 = np.ones((1,4,4,3)) * -1
X2 = np.ones((1,4,4,3)) * 1
X3 = np.ones((1,4,4,3)) * 3

X4 = np.concatenate((X1, X2, X3), axis=0).astype(np.float32)
X5 = np.concatenate((X1, X2, X3), axis=0).astype(np.float32)

X6 = tf.keras.layers.Conv2D(filters=2, 
           kernel_size = 1, 
           strides = (2,2),
           padding='valid',
           kernel_initializer=glorot_uniform(seed=0))(X4)

X7 = tf.keras.layers.Conv2D(filters=2, 
           kernel_size = 1, 
           strides = (2,2),
           padding='valid',
           kernel_initializer=glorot_uniform(seed=0))(X5)

assert(np.array_equiv(X6,X7))

that code passes the assertion

What’s happening with X1 in that block?

[Lingxiao_Gao]

I had both piece failed.

I also tried:

X1 = np.ones((1, 4, 4, 3)) * -1
X2 = np.ones((1, 4, 4, 3)) * 1
X3 = np.ones((1, 4, 4, 3)) * 3
X = np.concatenate((X1, X2, X3), axis = 0).astype(np.float32)

y = Conv2D(filters = 2, kernel_size = 1, strides = (2,2), padding = 'valid')(X)
z = Conv2D(filters = 2, kernel_size = 1, strides = (2,2), padding = 'valid')(X)

assert(np.array_equiv(y,z))

and it called:

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Input In [204], in <cell line: 7>()
      5 y = Conv2D(filters = 2, kernel_size = 1, strides = (2,2), padding = 'valid')(X)
      6 z = Conv2D(filters = 2, kernel_size = 1, strides = (2,2), padding = 'valid')(X)
----> 7 assert(np.array_equiv(y,z))

AssertionError: 

[ai_curious]

Strange. I pasted that code into a cell in my notebook and the assertion evaluated True.

X1 = np.ones((1, 4, 4, 3)) * -1
X2 = np.ones((1, 4, 4, 3)) * 1
X3 = np.ones((1, 4, 4, 3)) * 3
X = np.concatenate((X1, X2, X3), axis = 0).astype(np.float32)

y = tf.keras.layers.Conv2D(filters = 2, kernel_size = 1, strides = (2,2), padding = 'valid')(X)
z = tf.keras.layers.Conv2D(filters = 2, kernel_size = 1, strides = (2,2), padding = 'valid')(X)

assert(np.array_equiv(y,z))

print(y)
print(z)

tf.Tensor(
[[[[ 0.90270627 -0.7846885 ]
   [ 0.90270627 -0.7846885 ]]

  [[ 0.90270627 -0.7846885 ]
   [ 0.90270627 -0.7846885 ]]]
...

tf.Tensor(
[[[[ 0.90270627 -0.7846885 ]
   [ 0.90270627 -0.7846885 ]]

  [[ 0.90270627 -0.7846885 ]
   [ 0.90270627 -0.7846885 ]]]

[Lingxiao_Gao]

I removed all “assertion” tests after exercises in the notebook and succeed to reproduce model with accuracy >0.80 of several runs.

It seems Coursera environment always produce the same seed with np.random.seed(1) , and it isn’t the case for my local environment.