Adding data_augmentation layer causes errors with model_1 in 05_transfer_learning

[MikaATuomaala]

Hi,

I have problems with "05_transfer_learning". Following code used for model_1:

input_shape = (224, 224, 3)
base_model = tf.keras.applications.EfficientNetV2B0(include_top=False)
base_model.trainable = False

inputs = layers.Input(shape=input_shape, name="input_layer")

x = data_augmentation(inputs, training=True)

where data_augmentation is defined as follows:

data_augmentation = keras.Sequential([
    layers.RandomFlip("horizontal"),
    layers.RandomRotation(0.2),
    layers.RandomZoom(0.2),
    layers.RandomHeight(0.2),
    layers.RandomHeight(0.2),
    # preprocessing.rescasle(1./255) # Resnet50v2 may need this
], name = "data_augmentation")

produces following error:

Input 0 of layer "functional_1" is incompatible with the layer: expected shape=(None, 384, 512, 3), found shape=(None, 224, 224, 3)

Arguments received by Sequential.call():
  • args=('<KerasTensor shape=(None, 224, 224, 3), dtype=float32, sparse=False, name=input_layer>',)
  • kwargs={'training': 'True', 'mask': 'None'}

As far as I can tell, the code is same as in the updated course notes. And I have tried to copy-paste from the updated notes, but that doesn't seem to make any difference.

Tensorflow version is 2.17.1 and the runtime type is T4 GPU

Any ideas what could be causing this? And better yet, how to fix it?

Thanks,
-Mika

[ilsid]

Hi Mika,

I faced the same issue.
Looks like "View a random image" check sets up the input shape based on the passed image's size

# Augment the image
augmented_img = data_augmentation(tf.expand_dims(img, axis=0))

So, you've got (None, 384, 512, 3) shape for the first random image.

I found the following workaround

cloned_da_model = tf.keras.models.clone_model(data_augmentation)
augmented_img = cloned_da_model(tf.expand_dims(img, axis=0))

The original data_augmentation remains untouched and the cloned one is used for the image augmentation

Regards,
Illia

[MikaATuomaala]

Hi Illia,

Thanks for the hint!

Cheers,
Mika