Quantization model recall drop around 5%.

[Rfank2021]

I have a pytorch roberta base model for a classification task, and I use onnxruntime to quantization, I found precision is almost same, but recall drop around 5%.

My code is similar to https://github.com/huggingface/notebooks/blob/main/examples/onnx-export.ipynb, and I also saw https://medium.com/microsoftazure/faster-and-smaller-quantized-nlp-with-hugging-face-and-onnx-runtime-ec5525473bb7 saying that f1 score should be similar to the original model.

I tried onnxruntime.quantization.qdq_loss_debug. create_weight_matching, but it outputs an empty dict. I believe I already tried most of the config during quantization, is there a way or tool to determine which nodes to exclude during quantization?