CoTNet

CoTNet copied to clipboard

export onnx error

3

cuda11.1 + pytorch1.8 + cupy-cuda111, how to fix it, thanks a lot.

MZener

请问aggregation_zeropad.py有什么作用呢？看不太懂。既然论文中说CotLayer 可以即插即用，但这么复杂也无法迁移到其他模型块儿中呀？还望大佬解答

Chang-shuaishuai

Dear predecessors who have solved this problem, thank you very much for your help.

LJIAl

训练过程中的loss值不正常，请问可能是什么原因导致的呢？

1


rhj0103

XhHello

Some questions about your code

1

您好，我对您关于使用卷积来实现Self-Attention，并因此来替代CNN backbone中的bottleneck这种设计非常感兴趣，但是关于Contextual Transformer block我有一些问题，想向您请教： 1. 基于contextual attention matrix w来attention所有的values map V得到attented feature map，为什么要是用LocalConvolution，而不是直接的矩阵乘法，这样设计的原因是什么呢，并且将contextual attention matrix进行reshape分组之后再与value map进行LocalConvolution， 这个LocalConvolution 具体是怎么实现的呢？：[reshape](https://github.com/JDAI-CV/CoTNet/blob/master/models/cotnet.py#85) - [LocalConvolution](https://github.com/JDAI-CV/CoTNet/blob/master/models/cotnet.py#88) 2. 代码中在static key与contextual dynamic key进行fusion之后，为什么又进行了一个类似Self-Attention的操作呢？这个设计的目标又是什么呢？好像在论文中并没有提关于这里的细节：  3. 最后，我想问的是，模型的前向传播过程中没有出现任何关于position encoding或者position bias的设计，是因为采取了卷积操作替代了之前的Self-Attention机制，由于卷积的捕捉local-range信息的能力，就不再需要position...

Huzhen757

Questions about CotLayer's parameters and computational complexity

1

Thank you for the open-source code. Do you have a comparison of the parametric size of conv2d and cotlayer when the kernel size are the same? How do you calculate...

dangf15

When trying `python3 -m torch.distributed.launch --nproc_per_node=8 train.py --folder ./experiments/cot_experiments/CoTNet-50-350epoch` I am getting the error `ImportError: cannot import name '_C' from partially initialized module 'config' (most likely due to a circular...

sunildkumar

pretrained CotNet model for classification

4

I have files on my Google Drive that contain data.pkl which may be a pre-trained model, but I am unable to load it. Do you have any pre-trained models for...

SomayehSe

Can you supply an implement of CoTNet3d, i want to use it in ResNet3d。

chenqqqqqqq