Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

[tfjs-node] Inconsistent behavior for `BatchMatMul` kernel

See original GitHub issue

kernel BatchMatMul does not have same functionality in WebGL and NodeJS backends

This causes newer variation of FaceMesh-with-Attention model to fail when executing in NodeJS environments

Official implementation for a model was added via issue https://github.com/tensorflow/tfjs/issues/6002 and PR https://github.com/tensorflow/tfjs-models/pull/1002
Reference model is https://tfhub.dev/mediapipe/tfjs-model/face_landmarks_detection/attention_mesh/1

Error is:

Message: In[0] and In[1] has different ndims: [1,4,4] vs. [4,4]
    at NodeJSKernelBackend.executeSingleOutput (@tensorflow+tfjs-node@3.18.0/node_modules/@tensorflow/tfjs-node/dist/nodejs_kernel_backend.js:214:43)
    at Object.kernelFunc (@tensorflow+tfjs-node@3.18.0/node_modules/@tensorflow/tfjs-node/dist/kernels/BatchMatMul.js:33:24)
    at kernelFunc (@tensorflow+tfjs-core@3.18.0/node_modules/@tensorflow/tfjs-core/dist/tf-core.node.js:4574:32)

Same model executes without issues in browser environments using WebGL backend

Environment: TFJS 3.18.0 on Ubuntu 21.10 (WSL2) with NodeJS 16.15.0