[wasm] Add SquaredDifference kernel.

[santhiya-v]

tfjs-node / tfjs-backend-wasm version

1.7.2

Problem

I am running my model using tfjs-node code and it works fine(tensorflow backend). When I try to use 'wasm' backend (for performance), I get following error :

Error: 'squaredDifference' not yet implemented or not found in the registry. Did you forget to import the kernel?

Model :

I am using the transformer model (depthwise separable convolutions) from https://github.com/reiinakano/arbitrary-image-stylization-tfjs

Is there any way support for this kernel in 'wasm' can be achieved?

Thanks.

[annxingyuan]

Hi @santhiya-v - thanks for reporting this - we aim to add kernels to our WASM backend gradually in response to user need, which we assess by looking at issues like this one :)

However you say you want to use the wasm backend for performance - I wonder whether you might get better performance with our default WebGL backend instead.

[janakg]

@annxingyuan thanks for the response.

We are trying in both node and browser. In a browser with WebGL mode, toPixel() takes a long time in subsequent calls.

So we are trying in node with wasm backend for faster predict() , but we are limited by wasm's kernel support.

Please let us know if we are missing something.

[alvinsunyixiao]

@annxingyuan - Hi, I have put up a PR #3784. Could you help review it? Thanks!

[rthadur]

Related PR is merged now , closing this issue.Thank you