search

intel / webml-polyfill

Deprecated, the Web Neural Network Polyfill project has been moved to https://github.com/webmachinelearning/webnn-polyfill
Apache License 2.0
161 stars 42 forks source link

[DNNL] The labels are incorrect for squeezeNet(OpenVINO) #866

Open Christywl opened 5 years ago

Christywl commented 5 years ago

Test Env: Chromium Version: nightly build 75.0.3739.0 (f456c6d) Platform: Linux/Windows/macOS

Expected Result: The labels should be correct.

Actual Result: The labels are incorrect for squeezeNet(OpenVINO):

SqueezeNet(OpenVINO)

How to Reproduce:

  1. Setup test server and start
  2. Visit http://localhost:8080/examples/image_classification
  3. Select FAST_SINGLE_ANSWER
  4. Select squeezeNet(OpenVINO)
pinzhenx commented 5 years ago

This might be an issue in the softmax. If we bypass the softmax, i.e. set SOFTMAX under mkldnn to False, and enable dual-backend, say WASM + FAST, then it works fine.

pinzhenx commented 5 years ago

It seems that it’s the input shape of the softmax that leads to the wrong results: [1,1000] is ok but not the [1,1,1,1000].

https://github.com/intel/webml-polyfill/issues/884