lartpang / zoomnet Goto Github PK

你好，我有一些复现的疑问，我采用COD10k中camouflaged的3040张数据和CAMO的1000张数据进行训练40轮次，训练测试结果与下载训练好的测试模型对比差了好几个点，是训练配置或是训练的数据集存在问题呢？

Hello, when I was running The test code of COD task, a RuntimeError: Unexpected EOF, Expected 578749 more bytes occurred. The File might be corrupted. That is, the pre-trained model is corrupted. After many confirmations, the COD pre-training model I downloaded was 130,135,900 bytes in size. I wonder what you think about this question.

I noticed that you use the "py_sod_metrics" module in the file "utils/recorder/metric_caller.py", but I can't install it in the pip. Can you give me the source code or installation package of this module?

Thank you for your nice work. I am a novice, I would like to ask how to draw the figure 1 of the paper.

Why does it always occupy gpu:0 when I train on other graphics cards

您好，论文中阐述HMU的过程是讲每次处理分裂成三组的第一个，也就是g＇¹与下一组g相加，但是代码里的操作是用g＇²相加，最后α的产生也与论文有出入，请问具体是参照哪一个，谢谢

Hello, I found '_ConfigDict' object has no attribute 'test' when I was executing the test code. I wonder if I could ask for your help

您好，在您的dataset代码中，multi-scale方法如下图：

这是否表示，不同的scale是直接从原始图像分辨率直接resize得到的。也就是说，对于1.5的scale，网络的输入分辨率实际是384*1.5=576，也即是您的网络使用了576*576的大分辨率训练。

想请教问一下在ZoomNet.py中的class ZoomNet_CK是起什么作用的

您好，我最近在做复现工作，数据集这块一直也没弄好。因为我相对来说还是小白，希望你们可以提供一些复现步骤的信息(数据集配置)。十分感谢

I encountered the following error when running, what is going on?

Traceback (most recent call last):
File "main.py", line 355, in
main()
File "main.py", line 340, in main
model_ema = training(model=model, cfg=cfg)
File "main.py", line 204, in training
probs, loss, loss_str = model(
File "/home/nls1/anaconda3/envs/zjb4/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/data/zjb/camo/ZoomNet-main/methods/module/base_model.py", line 19, in forward
results = self.train_forward(*args, **kwargs)
File "/data/zjb/camo/ZoomNet-main/methods/zoomnet/zoomnet.py", line 224, in train_forward
output = self.body(
File "/data/zjb/camo/ZoomNet-main/methods/zoomnet/zoomnet.py", line 199, in body
l_trans_feats = self.encoder_translayer(l_scale)
File "/data/zjb/camo/ZoomNet-main/methods/zoomnet/zoomnet.py", line 195, in encoder_translayer
trans_feats = self.translayer(en_feats)
File "/home/nls1/anaconda3/envs/zjb4/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/data/zjb/camo/ZoomNet-main/methods/zoomnet/zoomnet.py", line 47, in forward
assert isinstance(xs, (tuple, list))
AssertionError

你好，根据您提供的checkpoint，我发现您在论文中report的似乎是max E_measure，但是其他论文报的应该是mean E_measure，这似乎是不是有点不合理

运行main.py让它训练，代码中最后会进行测试，并把测试结果输出，

这个指标和论文相差的有点大，请问一下这是怎么回事

想请问一下在HMU模块上，文章中描述的是每个group的第一个去跟下一个层融合，而对应代码里取的是outs[group_id-1][1]，但是这里的应该为0才跟原文的意**对应吧。
for group_id in range(1, self.num_groups - 1):
branch_out = self.interact[str(group_id)](torch.cat([xs[group_id], outs[group_id - 1][1]], dim=1))
outs.append(branch_out.chunk(3, dim=1))
以及包括后面这两部分的内容， o[0]这里的0是否应该为1才对呢？
out = torch.cat([o[0] for o in outs], dim=1)
gate = self.gate_genator(torch.cat([o[-1] for o in outs], dim=1))

Hi,
Thanks for sharing the code. I am interested in your paper.
I followed your instructions and ran the code with the default setting, but I cannot reproduce the results you proposed in the paper.
The following are the results I reproduced:
CHAMELEON: 'MAE': 0.032, CAMO: 'MAE': 0.075, COD10K: 'MAE': 0.033
I am wondering how to reproduce your results.
Thank you.

在您的论文中，您提到，辅助尺度的图像是由主尺度的图像rescale得到的，然而您的代码是通过rescale原图像来获得1.5，1，0.5这三个尺度的图像的。如果按照您原文叙述，下采样得到主尺度然后上采样得到的1.5倍尺度的图像，这损失的细节要比直接对原图像下采样多。这样感觉1.5倍的尺度并不是辅助尺度了，而是主尺度。很感激您能解释一下我的疑惑，谢谢！

Hi. I'm impressed by your awesome zoomnet method. And I have a question about the code.
On sod_zoomnet config, there's a config name ms.exta_size. And the config is used on baseSODdataset. And the annotation explain the value as "extra-scale for multi-scale learning". But I can't find the usage of these extra-scale information.
Because the model gets the data consist of only 1.5, 1.0, 0.5.
Therefore, could you explain the details and usage of extra-scale size? Thank you.

请问Zoomnet的整体网络参数是多少M呢

你好，我用COD10K-v3的测试集进行测试的时候，指标差距很大，用NC4K测试的指标是正常的。请问是什么原因？
“COD10K-V3”:
{'Smeasure': '0.851', 'wFmeasure': '0.369', 'MAE': '0.082', 'adpEm': '0.863', 'meanEm': '0.875', 'maxEm': '0.885', 'adpFm': '0.375', 'meanFm': '0.388', 'maxFm': '0.395'}
"NC4K":
{'Smeasure': '0.853', 'wFmeasure': '0.784', 'MAE': '0.043', 'adpEm': '0.907', 'meanEm': '0.896', 'maxEm': '0.912', 'adpFm': '0.814', 'meanFm': '0.818', 'maxFm': '0.828'}
是因为没有使用测试集中的背景图片以及非伪装目标的图片吗？

我在代码里指定了用别的显卡运行
然后输入这条指令
python main.py --model-name=ZoomNet --config=configs/zoomnet/zoomnet.py --datasets-info ./configs/base/dataset/dataset_configs.json --info demo

结果发现它还是会占用卡0的一点点显存，这是怎么回事？我想要全部搬运到其他显卡上跑实验，请问要怎么操作？

您好！我是一个小白，最近我在进行复现模型的工作，但是我遇到了一个解决不了的报错：
C:\code\pythonProject\Scripts\python.exe C:/code/pythonProject/ZoomNet-main/main.py
C:\code\pythonProject\henv\Lib\site-packages\numpy_distributor_init.py:30: UserWarning: loaded more than 1 DLL from .libs:
C:\code\pythonProject\henv\Lib\site-packages\numpy.libs\libopenblas.FB5AE2TYXYH2IJRDKGDGQ3XBKLKTF43H.gfortran-win_amd64.dll
C:\code\pythonProject\henv\Lib\site-packages\numpy.libs\libopenblas64__v0.3.21-gcc_10_3_0.dll
warnings.warn("loaded more than 1 DLL from .libs:"
Traceback (most recent call last):
File "C:\code\pythonProject\ZoomNet-main\main.py", line 355, in
main()
File "C:\code\pythonProject\ZoomNet-main\main.py", line 304, in main
with open(cfg.path.cfg_copy, encoding="utf-8", mode="w") as f:
OSError: [Errno 22] Invalid argument: 'C:\code\pythonProject\ZoomNet-main\output\ZoomNet_BS8_LR0.05_E40_H384_W384_OPMsgd_OPGMfinetune_SCf3_AMP\cfg_2023-04-23 00:50:04.555687.py'

Process finished with exit code 1

请问一下，这个是什么问题呢？

想知道文中关于pr曲线和F曲线的绘制方法
我使用 sklearn.metrics 的 precision_recall_curve 函数
将数据集每张图的像素信息保存为 N*1 维数组再送入此函数中，但是对大型数据集来说速度非常慢
请问能否告知您的方法

你好，我想使用你们已经训练好的模型继续训练几轮epoch的时候，进行到这一步的时候
if cfg.resume_from:
params_in_checkpoint = io.load_specific_params(
load_path=cfg.resume_from, names=["model", "optimizer", "scaler", "start_epoch"]
)
报出了这个错误
"There is not 'net_state' in ./output/ForSharing/cod_zoomnet_r50_bs8_e40_2022-03-04.pth: ['shared_encoder.conv1.weight',

请问这个该如何解决呢？

您好！请问一下，就是关于训练数据集，是要把camo中GT部分删除剩下和image对应的1000张、挑出cod10k3040张原始图和对应的gt，然后合在一起吗？