dk-liang / transcrowd Goto Github PK

how can we generate the attetion weight map?

rt，请教一下
此方法是否可以用于其他类型的密度估计（如鸟群）？

Really quick to fix:

In one of your training calls, there is a "--" missing before "batch_size":
"python train.py --dataset ShanghaiA --save_path ./save_file/ShanghaiA batch_size 24 --model_type 'gap'"

是要训练整整2w个epoch吗，我浮现的结果停在MAE 200左右啊

您好，我使用您训练好的模型进行测试，在test.py中运行，出现unboundLocal Error问题，请问如何解决谢谢！

Hi, thanks for your novel work. Can you disclose pre-trained models for other datasets and data preparation, which I can compare with your work? Thanks again.

Input image size ({H}{W}) doesn't match model ({self.img_size[0]}{self.img_size[1]})

想复现代码，但是报了上面的错误，不知道什么原因

您好，最近想复现您的代码时候，根据readme中的指示，在train阶段发生报错，具体表现为：
RuntimeError: stack expects each tensor to be equal size, but got [3, 384, 384] at entry 0 and [3, 768, 1152] at entry 1
尝试用transformer.Resize(384,384)解决无果，遂来询问。望解答

tuner_params = nni.get_next_parameter()根据这个tuner_params为空，看来nni并没有调整参数呢，为什么要调这个呢

没有提供prepare_nwpu.py 跟描述里面的不一样呢

你好，我正在复现你的代码。有两点疑问：

1. batch_size 为多少？原文中并未提到
2. 文中提到，regression token 有两个线性层，请问第一个线性层采用的权重维度是VIT的code中的representation_size吗？具体是多少呢？另外，连个线性层之间的activate function 用的是‘tanh’还是‘gelu’？预训练模型的head中的参数需要load进来吗？

请问作者尝试调整过里面一些超参数吗？就比如用AdamW优化器等，而且现在出了一些对VIT优化的trick，会不会使您的方法有一定提升？

So im playing with this model around to see exactly how it works at code level, and as far as I know the 'token' model uses a regression token to the input sequence Z0 for the counting, creating a size of HW/K² + 1 input in the regression head (being K the number of patches, HW the dimensions of the image). But i am not able to recognize the explicit difference between the 'token' and 'gap' regression heads inputs in the code.

Could you give me more explanation on how this "regression token" is created and where? and what it is exactly? the paper does not give much enough information about it...

请问作者可以提供attention weight可视化的代码吗？从而方便后续工作分析网络在关注什么

The performance cannot be reproduced. The experiment was conducted on Shanghaitech A, and the mae of the method using the token is 70.4, and the method using the gap is only 74. All experiments performed in accordance with the experimental hyperparameters

In weakly-supervised counting, only the global crowd numbers are available. If you crop one image into several patches, there will be more supervision information, and the comparison may be unfair due to different number of patches.
I wonder the performance with only the global ground truth number of each image.
Thanks!

I am very interested in your proposal and would love to use it.
However, since I cannot create a Baidu account, can you distribute the trained model with other cloud services?

Thank you for your work

Thank you for your work

Now, when I run test.py and train.py, I get the error "UnboundLocalError: local variable'img_return' referenced before assignment".

Can you solve this problem?
I'm running it with the code below.

python test.py --dataset ShanghaiA --pre ./Networks/model_best_A_gap.pth --model_type gap

python train.py --dataset ShanghaiA --save_path ./save_file/ShanghaiA --batch_size 1 --model_type gap