TFJS converter mismanages EfficientDet models 2

[partus]

Efficientdet can't be used , it relates to #3903, but seems to be a completely different one
For this model

Converted the following way

tensorflowjs_converter \
    --input_format=tf_hub  \
    --output_format=tfjs_graph_model \
    --signature_name=serving_default \
    --saved_model_tags=serve \
    ./efficientdet_d0_1 ./web_model_tfhub

Loaded and executed as follows in a browser

const model = await tf.loadGraphModel("web_model_tfhub/model.json");
t = tf.browser.fromPixels(img).expandDims(0)
let output = await model.executeAsync(t)

Results in

Uncaught Error: FusedConv2d and DepthwiseConv2d with BiasAdd must have one extra argument: bias.
    at UO (tfjs@2.7.0:17)
    at tfjs@2.7.0:17
    at tfjs@2.7.0:17
    at tfjs@2.7.0:17
    at e.t.scopedRun (tfjs@2.7.0:17)
    at e.t.tidy (tfjs@2.7.0:17)
    at lx (tfjs@2.7.0:17)
    at tfjs@2.7.0:17
    at KO (tfjs@2.7.0:17)
    at p (tfjs@2.7.0:17)

Same when the model is taken from the official repo.

Tried both tfjs version 2.7.0 , which was released just after #3903 was resolved , and 3.6.0, the result is the same

[pyu10055]

@partus I was able to convert and run the TFJS model without errors. Please make sure your tensorflowjs pip is the same version (3.6.0) as the tfjs library.

[partus]

What versions of other packages are relevant? : I am running it on Arch linux, hence the most recent tensorflow + python as well, e.g. tensorflow with python 3.9... is not officially supported. Which are the important dependencies so that I will I will put it into a cantainer, Hopefully veryify that it’s fine. Then with the next step we might verify the issue in a container with upgraded versions

[Linh0704]

@partus Have you resolved this issue, I've countered the same one. Change tfjs version and tensorflowjs won't work
Uncaught Error: FusedConv2d and DepthwiseConv2d with BiasAdd must have one extra argument: bias.

[partus]

@Linh0704 , @pyu10055 I was able to convert and run the tfhub version with python 3.8 instead of 3.9, and tfjs 3.6.0 , but I was unable to run the model from the official repo. Did not dive into it further as decided to run the model on a server.

[rthadur]

Closing the issue , please mention@ if you need further help to reopen. Thank you

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[npip99]

@rthadur I get the same issue right now, on npm tensorflowjs@3.13.0, with pip version of tensorflow==2.7, tensorflowjs==3.13.0

tensorflowjs_converter --input_format=tfjs_layers_model --output_format=tfjs_graph_model
    ./models/MyModel/model.json ./models/MyGraphModel

let graphModel = await tf.loadGraphModel('file://' + path.resolve('models', 'MyGraphModel', 'model.json'));
graphModel.predict(...);