Unknown layer: SpatialDropout2D

The following error appears to me when trying to import a model in tensorflow.js

Uncaught (in promise) Error: Unknown layer: SpatialDropout2D. This may be due to one of the following reasons:

1. The layer is defined in Python, in which case it needs to be ported to TensorFlow.js or your JavaScript code.
2. The custom layer is defined in JavaScript, but is not registered properly with tf.serialization.registerClass().

It seems like SpatialDropout isn't implemented yet. Is anyone working on it?

[syt123450]

Hi @caisq , are you working on it? I would love to implement SpatialDropout2D layer and noise shape.

[jessetrana]

In case it helps anyone with preexisting models: I found this issue because I had trained a model in Keras, used tensorflowjs_converter, and then encountered this error only when trying to infer in JS. In this situation, you can remove the layer in the converted model by simply modifying model.json to remove the layer reference and pull the output of the previous layer to be the input of the next. That way you don't have to do the surgery in Keras as a post-processing step on model saving.

[syt123450]

Hi @caisq , I took a look into this issue, it seems that to implement SpatialDropout2D, we need to make tf.layers.dropout support noiseShape first. While dropout layer only supports default noiseShape now (as dropout op in tfjs-core only support default noiseShape), I sent two PRs to make the dropout op and dropout layer support non-default noiseShape. Could you please take a look into these PRs? Thanks!

• Make dropout op support non-default noiseShape tensorflow/tfjs-core#1782
• Make dropout layer support non-default noiseShape and seed tensorflow/tfjs-layers#556