[tfjs-converter] @tensorflow/tfjs 1.2.3 fails to load weights from the official GCP bucket due to CORS issues

TensorFlow.js version

1.2.3 & 1.2.2

Browser version

Chrome Version 77.0.3854.3 (Official Build) dev (64-bit) FireFox 69.0b4 (64-bit)

Bug report

Here is a minimalistic repo showing the bug.

Loading one of the PoseNet weights in the erroring demo fails with:

Access to fetch at 'https://storage.googleapis.com/tfjs-models/savedmodel/posenet/mobilenet/quant2/100/group1-shard2of2.bin' from origin 'http://localhost:1234' has been blocked by CORS policy: Response to preflight request doesn't pass access control check: No 'Access-Control-Allow-Origin' header is present on the requested resource. If an opaque response serves your needs, set the request's mode to 'no-cors' to fetch the resource with CORS disabled.

Both Chrome and FireFox allow the JSON file, but deny the loading of the model chunks. Using CORS Everywhere helps in FireFox.

This issue was first spotted in the context of semantic segmentation and text detection models loading weights from a non-official GCP bucket with the following CORS rules suggested by @nsthorat here:

[
 {
   "origin": ["*"],
   "responseHeader": ["Content-Type"],
   "method": ["GET", "HEAD"],
   "maxAgeSeconds": 3600
 }
]

I have tried imposing other rules as well, as described here, but the issue persists.

[nsthorat]

Loading from the official buckets should work now. I’m reverting a change that will fix this for any custom server.

[kangyizhang]

hi @sdll @erikcartman

This is a bug in tfjs@1.2.3 and we will fix it soon.