Error: Unkown value 3,3,3,3 for padding Enum in tf.loadModel

Hello,

I have exported successfully my Keras model using the tensorflowjs_converter without any error or warning. However, when I try to load the model using tf.loadModel(‘model.json’), I get the following error:

Uncaught (in promise) Error: Unkown value 3,3,3,3 for padding Enum
    at new t (tf.min.js:1)
    at i (tf.min.js:1)
    at i (tf.min.js:1)
    at i (tf.min.js:1)
    at i (tf.min.js:1)
    at Object.i [as convertPythonicToTs] (tf.min.js:1)
    at tf.min.js:1
    at r (tf.min.js:1)
    at Object.next (tf.min.js:1)
    at tf.min.js:1

I use the following Keras layers: BatchNormalization, Conv2D, Activation(‘relu’), ZeroPadding2D, MaxPooling2D. Also the following Keras operation: keras.layers.add. Is this due to a layer not implemented yet in tensorflowjs? I am using tensorflow 1.7, Keras 2.1.5 and tensorflowjs 0.1.0 Also, I found here the operations and layers supported from Tensorflow, where could I find the same for Keras?

Thanks a lot

[ericdnielsen]

I’ve identified two issues:

1. ZeroPadding2D wasn’t implemented.
2. The error message is bad, because all layers that were implemented use the padding keyword to map to a string (same, valid, casual) so we converted padding to an Enum internally for stronger type. The ZeroPadding2D layers uses ‘padding’ as a key to a multi-dimensional array, and thus breaks the mapping of value to enum.

[scaelles]

Perfect, thanks a lot for fixing it so quickly!