Ray

This one [https://github.com/microsoft/nni/blob/e8c78bbab03fa651563e21cdf004220f29d9560a/examples/tutorials/quantization_speedup.py](url)

I got the same error, do you find the solution?

Hi, @matudouchen after reducing my pytorch version to 1.9.1, I got another error running quantization_quick_start_mnist.py [07/15/2022-06:54:23] [TRT] [W] DynamicRange(min: -0.424213, max: 2.82149). Dynamic range should be symmetric for better accuracy....

I tried to train the model for 100 epochs on coco128 before, however, the error did not seem to be smaller. So maybe the error is caused by the training...

The code is slightly modified from [https://github.com/ultralytics/yolov5/blob/master/train.py](url), I try to make it simpler and easier to read. In short, I comment out amp scaler and ema in the training code...

Hi, thanks for helping. After the modification, the error did become lower, and the problem described above is solved in the newly generated python script [autoshape_qat.zip](https://github.com/alibaba/TinyNeuralNetwork/files/9347358/model_qat.zip). However, the conversion error...

QAT can run without error but it doesn't seem to lower the conversion error. The model trained weight and generated script are here [model zip](https://github.com/alibaba/TinyNeuralNetwork/files/9347738/model.zip). For a simple input, you...

Thanks for replying. As the [FAQ](https://github.com/alibaba/TinyNeuralNetwork/blob/main/docs/FAQ.md#how-to-handle-the-case-of-inconsistent-training-and-inference-computation-graphs) mentioned, yolov5 has different outputs during training and evaluation. However, when using Tinynn quantizer to convert the evaluation model, it will show error `rewrite...

I change this [line](https://github.com/ultralytics/yolov5/blob/master/models/yolo.py#L62) to make sure the model will go to evaluation mode, and then use the converter to convert it. And I did turn inplace into false. ```python...