NotImplementedError: (Unimplemented) Delete Weight Dequant Linear Op Pass is not supported for per-channel quantization

[nyh3174680732]

问题描述 / Problem Description

模型在自动压缩之前可以正常运行，自动压缩之后开始报错，错误大概是这个错误表明你在尝试使用不支持的操作。根据错误消息，可能是在尝试使用Delete Weight Dequant Linear Op Pass来处理权重去量化线性操作，但该操作不支持通道量化，不知道怎么解决 完全按照https://github.com/PaddlePaddle/PaddleSlim/tree/develop/example/auto_compression/detection#4%E9%A2%84%E6%B5%8B%E9%83%A8%E7%BD%B2教程走的。

运行环境 / Runtime Environment

• Paddle:PaddlePaddle 2.5.2

复现代码 / Reproduction Code

完整报错 / Complete Error Message

#predict.py %cd ~ !python predict.py /home/aistudio/work/data/val/test.txt result.json /home/aistudio Model Arch: DETR --transform op: Resize --transform op: NormalizeImage --transform op: Permute Traceback (most recent call last): File "/home/aistudio/predict.py", line 221, in main(infer_txt, result_path, det_model_path, threshold) File "/home/aistudio/predict.py", line 204, in main detector = Detector(pred_config, det_model_path) File "/home/aistudio/predict.py", line 126, in init self.predictor, self.config = load_predictor(model_dir) File "/home/aistudio/predict.py", line 78, in load_predictor predictor = create_predictor(config) NotImplementedError: (Unimplemented) Delete Weight Dequant Linear Op Pass is not supported for per-channel quantization (at ../paddle/fluid/framework/ir/delete_weight_dequant_linear_op_pass.cc:130)

可能解决方案 / Possible solutions

附件 / Appendix