qat训练模型的问题

[dejavvuu]

1. qat_export 中如何去掉量化反量化的算子
2. qat_export 中qat_mAP比训练中量化的模型qat训练时的val map低，已经对比了各种参数，没找到map下降的原因 照理说是一样的模型权重

[lippman1125]

@dejavvuu

1. you can remove quant/de-quant ops from onnx graph before deployment.
2. ~0.1 mAP drop is normal.

[dejavvuu]

@dejavvuu

1. you can remove quant/de-quant ops from onnx graph before deployment.
2. ~0.1 mAP drop is normal.

我的意思是

1. 应该有你们实现好的脚本，希望能提供
2. 这个我认为是同样的权重，但map表现不同是为什么，能解释下吗

qat in train Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.896 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.989 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.966 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.765 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.884 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.927 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.474 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.914 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.916 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.802 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.911 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.945 qat _export Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.896 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.985 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.963 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.748 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.881 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.931 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.476 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.913 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.914 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.789 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.910 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.948

[dejavvuu]

可以提供相关脚本吗

[aitangbodan]

我也想问这个问题remove_qdq和sim是怎么操作的？有大佬知道吗？

[zahidzqj]

0.2.1版本中是怎么操作的，我转完trt后检测框不对，后来我在这里看个到了代码中关于remove_qdq的操作： https://github.com/meituan/YOLOv6/pull/653/commits/54deb7b233e322d9f57ba1c723b7013c53dd08eb#diff-0e63f33ad384f9d1c2694580813fca6e9fb69e5fae1b8e1c30e896e3484b8b2e 将该部分代码加入工程，然后检测框就正常了。

[DrChungAlbert]

在邊緣平台推論應該是要remove_qdq的onnx 對嗎 請問yolov6n_v2_reopt_qat_35.0_bs1_graph_opt_remove_qdq.sim.onnx 是否就是 remove_qdq的操作產出的? 有個.sim是否只是重新命名?