brevitas
brevitas copied to clipboard
Xilinx
Brevitas: neural network quantization in PyTorch
Supports export to Q(C)DQ and extend QOps with clipping. (C) as in clipping allows to correctly low precision quantization (e.g. 4b) by clipping the output of QuantizeLinear. @cmcgirr-amd
Switch over to a Vitis-AI like implementation and export flow for QuantAvgPool2d for FlexML graph quantization. Keep the conversion of AvgPool to dephwise Conv as an option but simplify the...
Hi brevitas team, I am a little confused: What is the difference between brevitas qat and Vitis AI qat? It seems that the usage of them are completely different.
Thanks for your great work. Is there any plan to reimplement some SOTA QAT method such as LSQ, PACT, EWGS.... on various network like ResNet series, Mobilenet series， Efficient series.........
Hi, If I create a layer using the following code which works. my question is: key works argument `weight_quant_type` could not be found in the parent class(`QuantLinear, QuantWeightBiasInputOutputLayer, QuantWeightMixin, QuantProxyMixin`)....
Hello, I am trying to fuse Conv & BatchNorm layers after training the model using `torch.nn.utils.fusion fuse_conv_bn_eval`. But I got a huge drop in accuracy, which I guess is resulting...
softmax
Hi, Is softmax supported? how to use it in my python code? Is there any document about sofmax in Brevitas?
May I know why 500 epochs should use for cnv1w2a, cnv2w2a and 1000 for cnv1w1a.?
Hi, developers of brevitas. I once worked with vitisAI team in Xilinx for serval months (2020~2021, internship), and now I still work for creating better network quantization tools in Sensetime....
When creating my own integer quantizer, I get the following error: > TypeError: forward() missing 3 required positional arguments: 'zero_point', 'bit_width', and 'x' Using this quantizer: ```python class ActInt(ExtendedInjector): tensor_quant...
