How to apply this package into Quantization models?

[Tallisgo]

First thanks for your hard work. Recently, I train a model on quantized mobilenet_v3_large ( based on torchvision ). I pruned this model and quantize the pruned model. I found precision dropped a lot and is not stable. Can you give me some suggestions?

[VainF]

Hello, @dongL-Wu. Is it possible to apply pruning and quantization separately in your case? For example, we can prune & finetune a model and then quantize it.

[Tallisgo]

Thanks for your reply @VainF. Maybe I just do this. The pipeline is below:

1. train and save a model
2. initial a model with the saved weight
3. use torch_pruning to prune
4. use quantization (pytorch provide) to quantize

` from torchvision.models.quantization import mobilenet_v3_large

对每个卷积剪枝

for m in model.modules(): if isinstance(m, nn.Conv2d): prune_conv(m, amount=0.2)

` Maybe prune every Conv is not valid and I need to analyse importance with this file?

[VainF]

I think the model should be finetuned after pruning to recover its accuracy.