Yaman Umuroglu

Hi! Happy to see the interest around

> One possibility is that we specify how the bit values are packed into an uint8 (as part of the Quantize op's spec) and we don't even need to introduce...

I think we would benefit from having a more structured discussion to consider the various aspects of representing quantization that are brought up in this thread. These aspects are related...

@xiguadong thanks, and agreed. Regarding your 3rd point, in this sentence: > 3rd, as using int32 as save format of uint8, maybe we can save the 2/4bit as int32. For...

I've been debugging the different outputs produced by the ONNX-exported 4-bit MobileNet-v1 and its Brevitas implementation, and it looks like this issue is also the cause of the problem there....

Hi @simplelins, I don't think there is an easy fix for this, unfortunately. You could manually add a data width converter at the stitched IP level to do the 512->8192...

From https://github.com/Xilinx/finn/blob/dev/src/finn/transformation/__init__.py#L30 ``` * Your transformation's apply function should take in a ModelWrapper, and return a tuple with (transformed_model: ModelWrapper, model_was_changed: Bool) * model_was_changed indicates whether your transformation made any...

Hello YZW-explorer, you should use the version of ONNX that your current version of the FINN docker image depends on, e.g. for the master branch: https://github.com/Xilinx/finn/blob/master/requirements.txt#L8 `onnx==1.7.0` In general my...

Hi Neil, thanks for reporting this. It sounds like you've found the root of the problem -- the current `SetFolding` strategy is node-local (i.e. it doesn't consider the neighboring nodes)...

@neilkimn that's great! Does your transformation take into account only cycles (like `SetFolding`) or both cycles+resource limit (like `AllocateResources`)? If it's the former I would have no qualms whatsoever about...