SwinTransformer3D

[wangjingg]

Does this framework support SwinTransformer3D quantification?

[SuperVan-Young]

Currently that's not available, sorry for that.

Some slight modification could get things work, if:

1. You have calibration dataset for the model and you know how to preprocess the pictures.
2. You'll need to manually substitude matrix multiplication, from an operation to a torch.nn.module, so that the framework can set quantization parameters in activation's quantization. But if you just want to quantize weights, you can skip this step.

[wangjingg]

我使用该框架成功量化了Swin3D的matmul和Linear算子, 但无法量化conv3D, 请问, 我需要如何修改才能让它工作?

[SuperVan-Young]

我们这篇工作主要考虑Vision transformer，所以没有实现conv3D算子的量化和封装。ViT的embedding操作是使用conv2D实现的，这部分代码在layers.conv里，您可以考虑仿照conv2D实现conv3D，从代码的核心逻辑上二者没有什么区别，只要把相关操作的维度增加一下就好了。conv2d当中每一步操作的维度如何变换，我在注释中逐行写出来了，可以用作参考。如果您已经实现了conv3D算子，您还要在utils.net_wrap里加上对conv3D的封装操作。