CustomMaskWarning

[lukbast]

Hello @mrdbourke,
Yesterday I finished my DeepFood model with 84% of accuracy but I ran on issues on saving it. For some reason I have this warning:

c:\MLandDS\tensorflow\env\lib\site-packages\keras\utils\generic_utils.py:494: CustomMaskWarning: Custom mask layers require a config and must override get_config. When loading, the custom mask layer must be passed to the custom_objects argument.
warnings.warn('Custom mask layers require a config and must override '

It appers when you create own layers by subclassing existing tensorflow classes. Problem is, I didn't do that. I used EfficientNetB0 from tf.keras.applications module and layers from tf.keras.layers.
This is function that I used to creting my model:

def create_model_dropouts_full_training():
    input_shape = (224,224,3)
    base_model = applications.EfficientNetB0(include_top=False)

    inputs = layers.Input(shape=input_shape)
    x = base_model(inputs)
    x = layers.GlobalAveragePooling2D()(x)
    x = layers.Dropout(0.3)(x)
    outputs = layers.Dense(101, activation="softmax")(x)

    return Model(inputs, outputs)

It's summary:

Model: "model_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_4 (InputLayer)         [(None, 224, 224, 3)]     0         
_________________________________________________________________
efficientnetb0 (Functional)  (None, None, None, 1280)  4049571   
_________________________________________________________________
global_average_pooling2d_2 ( (None, 1280)              0         
_________________________________________________________________
dropout_2 (Dropout)          (None, 1280)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 101)               129381    
=================================================================
Total params: 4,178,952
Trainable params: 4,136,929
Non-trainable params: 42,023
_________________________________________________________________

Also ploted architecture will be in attached file.

I'm using TF 2.6.0, Verion 2.4.1 is not available on PIP anymore:

!pip install tensorflow==2.4.1
ERROR: Could not find a version that satisfies the requirement tensorflow==2.4.1 (from versions: 2.5.0rc0, 2.5.0rc1, 2.5.0rc2, 2.5.0rc3, 2.5.0, 2.5.1, 2.6.0rc0, 2.6.0rc1, 2.6.0rc2, 2.6.0)
ERROR: No matching distribution found for tensorflow==2.4.1

Becasue of that, model loses performance (6%) when I save it under .h5 format and reload it. Im planing to deploy it in application in near future. If I wanted to use it n form with best performance I would have also provide entire tensorflow lib which is problematic becasue free hosting servers usually have very limited storage and TF needs some space + weight of the model itself.

I coudn't find solution to this problem anywhere. All sources addreses this issue for subclassed custom layers which is not the case here. Do you have any idea what is going on?

[someshfengde]

since you have an efficient net as your Keras layer while saving model it doesn't get saved with the model
so if you want to load the model then you have to specify a layer while loading

eg tf.keras.models.load_model('model_path', custom_objects = {"KerasLayer" : hub.KerasLayer('your hub model URL')})

[lukbast]

1. My model don't have any KerasLayers from TFhub. It's Functional layer from tf.keras.applications
2. When I load model, it has EfficientNetB0 inside as Functional layer.
   Architecture after loading:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_4 (InputLayer)         [(None, 224, 224, 3)]     0         
_________________________________________________________________
efficientnetb0 (Functional)  (None, None, None, 1280)  4049571   
_________________________________________________________________
global_average_pooling2d_2 ( (None, 1280)              0         
_________________________________________________________________
dropout_2 (Dropout)          (None, 1280)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 101)               129381    
=================================================================
Total params: 4,178,952
Trainable params: 4,136,929
Non-trainable params: 42,023
_________________________________________________________________

This is not a problem of EffNetB0 not being saved. Just think about it, it's not possible that model with 3 layers will be able to achieve 78% accuracy on such a big and complex dataset.

I did passed I like this:

loaded_model = tf.keras.models.load_model("./final_model.h5",
                            custom_objects={"efficientnetb0": tf.keras.applications.EfficientNetB0})

And it gave me nothing. Also tired with KerasLayer and TFhub with no effect.

[mrdbourke]

Hey @lukbast,

Any luck in solving this?

So when you're loading your model back the results are far less than before you saved it?

Is the data you're testing it on the same?

E.g. not shuffled every time or not testing on a different subset?

Have you tried saving it to SavedModel format? Does that work?

[lukbast]

Hi, @mrdbourke

My issue is that some of my layers are considered as "custom", despite the fact that I didn't create any custom layers. In result, some layers configurations are not saved correctly, what results with performance loss (78% after loading vs 84% before loading).

You can create custom layer by subclassing existing it's class from TF and tinker with it, and like I said before I didn't do that. Data is not the issue.

I did some digging, it's an issue with EfficientNet model. It dosen't save custom gradients. Model saved in .h5 format have 6% worse accuracy than it's version before saving, but when I save it in a SavedModel format, it broke completly. It's known issue:
tensorflow/tensorflow#40166

What I can do is to try different model or when I'll try deploy this model I just recreate it using a function and load saved weigths to maintain best accuracy.

[mrdbourke]

Excellent! Superb investigation skills, seems there's a few issues with the pretrained EfficientNet models. Glad you found a solution though.