A deterministic GPU implementation of fused batch-norm backprop, when training is disabled, is not currently available

Question

I am designing a CNN classifier for image classification with reproducibility. I am using the GPU of the Google Colab for this. To ensure the result is reproducible, I am enabling the TensorFlow ops deterministic using the "tf.config.experimental.enable_op_determinism()" command and getting the reproducible output. The problem is that when I am trying to create a gradient-based saliency map, I get an error. The error says,

"UnimplementedError: {{function_node _wrapped__FusedBatchNormGradV3_device/job:localhost/replica:0/task:0/device:GPU:0}} A deterministic GPU implementation of fused batch-norm backprop, when training is disabled, is not currently available. [Op:FusedBatchNormGradV3] name: "

Here is my code for creating the saliency map,

def salency_map(model, sample_index):

    # Choose a random sample from the test set
    #sample_index = np.random.randint(0, len(X_test))
    sample_image = x_test[sample_index][np.newaxis]
    #sample_image = x_test[sample_index]
    sample_label = y_test[sample_index]
    sample_class_index = np.argmax(sample_label)
    Y_true = np.argmax(y_test[sample_index])

    Y_pred_classes = np.argmax(model.predict(sample_image), axis=1)

    print("Actual Class: ", labels[int(Y_true)])
    print("Predicted Class: ", labels[int(Y_pred_classes)])

    # Initialize GradCAM object
    explainer = GradCAM()
    # Explain the model predictions on the sample image
    grid = explainer.explain((sample_image, None), model, class_index=sample_class_index)

    # Visualize the saliency map
    plt.figure(figsize=(4, 2))
    plt.subplot(1, 2, 1)
    plt.title("Original Image")
    plt.imshow(sample_image.squeeze())


    plt.subplot(1, 2, 2)
    plt.title("Saliency Map")
    plt.imshow(grid)
    plt.colorbar()

    plt.show()

How to overcome this issue?

After training the model, I tried to enable the non-deterministic mode again. But it did not work.