search

Xilinx / Vitis-AI

Vitis AI is Xilinx’s development stack for AI inference on Xilinx hardware platforms, including both edge devices and Alveo cards.
https://www.xilinx.com/ai
Apache License 2.0
1.48k stars 627 forks source link

Warnings during quantization of efficientnetv2 using vitis ai 3.0 pytorch quantizer . #1389

Open amitpant7 opened 10 months ago

amitpant7 commented 10 months ago

I have been trying to quantized channel pruned efficientnetv2 using vitis ai pytorch quantizer. But i have been facing these issues

[VAIQ_NOTE]: =>Successfully convert 'EfficientNet' to xmodel.(quantize_result/EfficientNet_int.xmodel) /opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/pytorch_nndct/nn/modules/prim_ops.py:116: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! if not (list(self.node.out_tensors[0].shape[1:]) == list(input.size())[1:]): /opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/pytorch_nndct/quantization/torchquantizer.py:53: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! if inf.sum() > 0 or nan.sum() > 0: /opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/pytorch_nndct/nn/modules/fix_ops.py:68: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! tensor.storage().size() != tensor.numel()): /opt/vitis_ai/conda/envs/vitis-ai-wego-torch/lib/python3.7/site-packages/pytorch_nndct/nn/modules/adaptive_avg_pool.py:42: TracerWarning: Converting a tensor to a Python integer might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs! input_size = [int(dim) for dim in input.shape[2:]]

The model do get quantized ,the onnx quantized version is working fine but when we deploy xmodel to fpga by compiling, we are not getting the accurate outputs and the model is classifying all classes to same label.

github to quantization: https://github.com/amitpant7/Quantizing-Efficientnetv2-using-Vitis-AI-Pytorch/tree/compiled/efficientnet

PedroPDelgado commented 9 months ago

Having a similar problem. Any news?

As for the last TracerWarning, changing from an AdaptiveAvgPool2d to AvgPool2d with a fixed kernel size (if that is ok for your use case) makes the warning go away.