"MAE at the 353 epoch BaiduDisk download with Extraction code: yfwb". 求问这个模型是part_A的吗? 我直接用这个测试part_A, 得到 mae:413.119863447

Hello! Thanks for your implementation of the network. Unfortunately for non - chinese people it's hard to download from pan.baidu so could you please upload your pretrained model on another cloud like dropbox or google drive? Because I've been training on Shanghai Part A but can't get lower than 66 MAE even after 600 epochs.
I've also been training on Shanghai Part B and convergence of the model seems to be much much slower. At 600 epochs MAE is around 28, much worse than the 7.8 reported in the paper. Even using the model trained on Part A for testing on Part B gives better results (something like 23 MAE). For training on Part B I just switched the optimizer to Adam instead of SGD and set batch size to 4 (because any more than that and a "cuda out of memory error" occurs, even though I'm training on a RTX 2080 Ti with 11GB of memory). Should I stick with it and just train for more epochs or change some other hyper parameters of the optimizer to train faster?

The paper sain'We therefore exploit them as input to auxiliary networks with weights θjsa that compute the weights ωj assigned to each one of the S different scales we use.',But I didn't see 'auxiliary networks with weights θjsa' in your code ,Is that a normal conv?Can you help me?

Since the amount of data is not so big, 300, can I set batch_size bigger?

如何将测试生成的密度图和GT输出出来？

Hello, I would like to ask if this code can calculate MSE. It seems that only MAE is calculated in test and train.

Is CPU enough for this project? I see there is no CUDA required.

Must I run python -m visdom.server? Because it is not so convenient to run the command line on the server.

Paper get MAE=62.3 in ShTechA, but I only get 69.02. Are there any mistakes in my code?

thanks for your code. but when i run ,i cant got same result as you say mae=62.3,at epoch 353.

can you give me some advice.
all code not change expect the path of data.

I am creating an algorithm that counts people in real time, and I am interested in using your repository, however, I have a question. When performing an inference with the command

et_dmap = model (img)
the result of the amount of people in the image would be

et_dmap.data.sum() ?
I was in doubt because I am using the .pth provided by you and I'm having a bit strange results.

FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type.
from ._conv import register_converters as _register_converters
在生成.npy的时候，生成不了

训练了多次以后，中断了，如何从断点开始训练呢？

when I train ，apperance “OSError: cannot find loader for this HDF5 file，how to solve this problem
”

你好，为什么我用的你的代码，跑了352个epoch(shanghaiTech Part_A)，mae81左右，没有达到论文说的62.3

what is the project's torch visition

cannnet.py第75行，谢谢你帮我解决除0错误

I found that if I use this method to generate the density maps, the count of the density map is different from the original .mat file. Is there something wrong? For example, the IMG_2 in ShanghaiTechA train data, the count of GT_IMG_2.mat is 707. But the count of the density map generated by your code is 698.
Thanks!

你好，我发现用你的代码生成的密度图和原始的.mat文件里的数目有些偏差，比如ShanghaiTech A里第二张图，原始的.mat文件中的label数是707，但用你的code生成的密度图sum之后是698。不知道是否有什么问题呢？还有你尝试过别的sigma和k取值的影响吗。
感谢!

FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type.
from ._conv import register_converters as _register_converters

请问您这个是不是只有train test，没有val？

第一次接触这类程序，请不要介意我的小白问题。.npy文件已成功生成，显卡是gtx1080，直接跑的代码里的数据集。在训练过程，每次读满两个进度条之后就出现[Errno 111] Connection refused的错误，然后又重新开始读两个进度条。不知道是我哪里的设置有问题，还是说本来这个训练的过程就需要一天以上的时间。

resize the img and target to the desired size?
It's not a common approach and not a ideal handle intuitively.
The image size of SHT-A is quite Variegated, will that make the MAE bigger?

Hello, thank you for your implements for Part_A dataset. I notice that there are some little differences between your model(cannet.py) and the official https://github.com/weizheliu/Context-Aware-Crowd-Counting/blob/master/model.py
The official code add one convolution layer to reduce channel numbers from 1024 to 512 before sending feature maps into backend, but you input it without any other operations.
I wonder if this way works better or for some other reasons. Thank you!

Q- Can i load pretrained weights for training of part A?

When using data_preparation / k_nearest_gaussian_kernel.py, I can't find 'IMG_1.npy'. Has anyone encountered this problem? Thanks for answering.

I am sorry if it comes out wrong. I have run the project successfully and tested on images from Shanghai dataset. I would like to see the number of person predicted. Is it possible?

I need 6mins to run a epoch which is much slower than yours. I use GTX1080. I don't change the code except the root path. Do you know the possible reasons? Thank you.