Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.
And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.
Some questions about quantization
See original GitHub issue您好，我在量化RepVGG到INT8时遇到一些问题： 完全按照您在README中提到的量化流程，先插入BN后QAT量化，掉点非常严重。以大模型（b3g4）为例，在ImageNet分类任务上量化后仅有20左右的accuracy。 但假如将https://github.com/DingXiaoH/RepVGG/blob/ce669bad170bf5d7c2a9eb060e90f39e7e109130/quantization/repvgg_quantized.py#L50代码修改为：` torch.quantization.fuse_modules(m, [‘conv’], inplace=True)` 可以直接运行QAT量化，则掉点只有2.7%，调参后掉点大约1.7%。（参数为epochs=20 batch_size:与训练时bs保持一致 lr=1e-3 weight_decay=55e-6） 但对于分类任务而言量化后掉点也有点多，不知道这样做是否正确？请问您有其他量化方法的见解吗？
Issue Analytics
- State:
- Created 2 years ago
- Comments:9
Top Results From Across the Web
Top Related Medium Post
Top Related StackOverflow Question
Troubleshoot Live Code
Top Related Reddit Thread
Top Related Hackernoon Post
Top Related Tweet
Top Related Dev.to Post
Top Related Hashnode Post
@twmht 我后来尝试直接用PTQ量化效果会好很多，TensorRT实现PTQ量化，用交叉熵做校准loss，RepVGG_A0模型在ImageNet分类任务上掉点仅0.16%。 我后来思考QAT量化掉点多的原因可能是：在QAT阶段用的模型是融合后的单分支推理模型，用这个模型进行QAT的finetune会使得模型退化，类似VGG的精度。但我并没有尝试用训练阶段模型进行QAT量化。
@hobbitlzy 没有用insert_bn。