Loaded Frozen model running forever

To get help from the community, we encourage using Stack Overflow and the tensorflow.js tag.

TensorFlow.js version

tfjs@0.15.1

Browser version

Chrome 72.0.3626.121

Describe the problem or feature request

The loaded frozen model take very long time (over 30 minutes. actually I never saw it complete predicting) even if it’s only about medium size (40MB). And it take me only about 3 seconds to forward pass the model using Python on MAC.

Since there is no message available when I called model.predict(), I’m asking for help to try to identify where should be wrong.

const model = await tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL);
console.log("Model loaded.");
console.log(model);
const cat = document.getElementById('cat');
t = tf.fromPixels(cat);
t = t.toFloat();
t = tf.expandDims(t, 0);

console.log("Start transformation by model...")
model.predict(t).dispose(); # get stuck here forever

Code to reproduce the bug / link to feature request

I had uploaded the code and the model artifacts for reproducing the issue here: Github repo: tfjs-debug-2. You can clone the repo and start a Python HttP server by python httpserver.py and visit localhost:1234.

Any suggestion or advice are appreciated!

[dsmilkov]

Great! tf.fromPixels became tf.browser.fromPixels in 1.0. See the “Breaking changes” section in our release notes: https://github.com/tensorflow/tfjs/releases/tag/v1.0.0

[dsmilkov]

Hi,

If you are short on time, you can produce a custom tfjs bundle with that PR and test to see that it fixes your problem.

• checkout the strided_slice_fix branch from tfjs-core
• yarn && yarn publish-local
• checkout the https://github.com/tensorflow/tfjs “union” repo
• yarn && yarn link-local @tensorflow/tfjs-core && yarn rollup -c
• the bundle should be in tfjs/dist/tf.(min).js

NOTE This will give you tfjs 1.x, which means it will only run the new format (model.json) and you have to use the new tf.loadGraphModel API instead of tf.loadFrozenModel. If you need 0.15.x, you’ll have to checkout the 0.15.x branch from tfjs-core, manually apply the fix in tensorflow/tfjs-core#1622 and follow the steps above.