Same Input node hits assert error when attempting to re-quantize bias. about sparseml HOT 4 CLOSED

Hi @vixadd, thank you for raising this issue. Could you point us towards a specific open-sourced model or implementation we could look at further? Additionally, what was the method you used for conversion to ONNX? We can then go through and verify the issue from our side to work on a specific fix for this issue.

Thanks,
Mark

from sparseml.

Hi @markurtz

• Thank you for the reply. Unfortunately, the implementation of the model itself is something I cannot share, however, I can share a general overview of what it is/does.
• I use tf2onnx to convert the keras model to an onnx model in my build.
• The model takes multiple Conv 2D layers and pulls in MaxPooling2D to downsample.
• I've indexed each of the conv2d layers. You can see a 3_1 after the identifier. For some reason sparseml is not pulling the other layers, it keeps trying to re-calculate index 3_1 in the list of Convolutions that take place.

from sparseml.

Since this has gone a little stale, I'll put up an example, and just leave it at that.

import tensorflow.keras as tfk

def build_encoder():
    layers = []
    for l in rang(0, 10):

        layer = []
        for b in range(0, 5):
            layer.append( tfk.layers.Conv2D(16, 8, 64, use_bias=False, name=f"encoder/conv2d_{l}_{b}" )

        layers += [layer]

Building something like that into your model will cause this error.

from sparseml.

Hi @vixadd, thank you for putting up that pull request! The code has been merged in now and will be going out with our 0.8 release. It will be available on nightly starting tonight, so I'm going to close this out for now. If any issues come up feel free to reopen.

Thanks,
Mark

from sparseml.