icey-zhang/superyolo

SuperYOLO is accepted by TGRS

superyolo's Issues

train-val size

1、这个train-val size代表train跟Val的大小是一样的吗？
1、你好，您在论文中提到训练的图片分辨率是1024，而测试的时候是512，，训练的时候图片要下采样为512，那么测试的时候图片输入大小是512，经过网络，那不也是变成了原来的1/2,变成256了吗？

A question about the test code

When I enter this line of code, I get the following error - "python test.py --weights weights/YOLOv5s/multi/yolov5s_multi_fold1.pt --input_mode RGB+IR+MF". How should I solve this problem? There are two files in your models folder, "yolors.py" and "SRyolo.py", will these two files help my problem? Thank you and wish you a happy life

训练一直报错，数据集也是下载的您的，怎么回事呢

train: Scanning 'D:\SuperYOLO-main\dataset\VEDAI_1024\images.cache' for images and labels... 0 found, 1089 missing, 0 empty, 0 corrupted: 100%|██████████| 1089/1089 [00:00<?, ?it/s]
Traceback (most recent call last):
File "D:/SuperYOLO-main/train.py", line 673, in
train(hyp, opt, device, tb_writer)
File "D:/SuperYOLO-main/train.py", line 216, in train
image_weights=opt.image_weights, quad=opt.quad, prefix=colorstr('train: '))
File "D:\SuperYOLO-main\utils\datasets.py", line 101, in create_dataloader_sr
prefix=prefix)
File "D:\SuperYOLO-main\utils\datasets.py", line 701, in init
assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {help_url}'
AssertionError: train: No labels in D:\SuperYOLO-main\dataset\VEDAI_1024\images.cache. Can not train without labels. See https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data

超分结果

你好，超分分支，L1损失是由Ground——Truth监督训练吗

关于不同训练模式下的配置修改和运行命令问题

您好！在readme文件中，是否可以较为详细的说明下RGB、IR、RGB+IR+MF、RGB+IR+fusion几种模式下，应该如何修改配置和运行命令？
另外，如果不使用超分辨率分支辅助训练的话，上述几种模式如何设置？谢谢！

推理时找不到对应的数据集文件？

您好，我在尝试使用small_EDSR_MF_fold1.pt预训练权重推理VEDAI测试集时发现并未找到对应的图片文件。报错截图如下：

我使用了您提供的transform.py对数据集进行预处理，并按照自己的路径配置好了SRvedai.yaml，生成的fold01test_write.txt如下所示：

我猜测可能是fold01test_write.txt文件里的图片路径后缀不对造成的，但我仍然不知道该如何修改才能使数据集和图片对应上，烦请您帮忙指点一下啦3Q

solved

训练问题

1.混合训练的时候，train_batch01.jpg上看到，只有RGB图像，没有IR图像呀
2.单独训练IR，train_batch01.jpg上，也是RGB图像，没有IR图像，这是对的吗？还是我理解问题，我理解的是RGB和IR图像都在里面，提供更多的目标样本

Option w.r.t '--multi-scale'

Is it possible to use --multi-scale option in training?

In my case, there is an error of "sr_loss = 0.1*(torch.nn.L1Loss()(ouput~~)" in the line 512 in train.py.

Could you help me in this case?

Thank you so much.

YOLOv5是什么版本？显存占用是多少？推理速度怎么样?

在运行4. Begin to train multi images without SR branch第一行代码时，出现错误

python train.py --cfg models/SRyolo_noFocus_small.yaml --train_img_size 512 --data data/SRvedai.yaml --ch 4 --input_mode RGB+IR File "D:\Anaconda\envs\super\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl
return forward_call(*input, **kwargs)
File "E:\SuperYOLO\models\deeplabedsr.py", line 60, in forward
x_sr= self.sr_decoder(x, low_level_feat,self.factor)
File "D:\Anaconda\envs\super\lib\site-packages\torch\nn\modules\module.py", line 1190, in _call_impl
return forward_call(*input, *kwargs)
File "E:\SuperYOLO\models\sr_decoder_noBN_noD.py", line 38, in forward
x = F.interpolate(x, size=[i(factor//2) for i in low_level_feat.size()[2:]], mode='bilinear', align_corners=True)
File "D:\Anaconda\envs\super\lib\site-packages\torch\nn\functional.py", line 3950, in interpolate
return torch._C._nn.upsample_bilinear2d(input, output_size, align_corners, scale_factors)
RuntimeError: Input and output sizes should be greater than 0, but got input (H: 16, W: 16) output (H: 0, W: 0)
请帮忙看下如何解决这个问题，我将万分感谢

你好，请问运用下载的VEDAI数据集训练时提示无标签怎么解决？

错误：

SRvedai.yaml文件：

已运行transform_vedai.py文件，txt文件如下：

test.py 的输出

你好，项目中存在RuntimeError 这个错误

有预训练模型权重文件吗？

之前发布的SuperYOLO_fold1_best.pt和small_EDSR_fold1.pt是一样的吗

迁移到yolo8出现make_anchor报错

作者可以把这个模型迁移到anchor-free的yolo8上面吗？我在迁移过程中出现了make_anchor的问题：

size只有三个参数

这是啥原因

希望早日开源，感觉很有研究价值！

文章中所提出的Pixel-level fusion具体是指什么融合方法，文章中所有的实验都是基于IR+RGB模式检测的吗

pandas._config.config.OptionError: "No such keys(s): 'mode.use_inf_as_null'"

您好，我在运行的时候遇到了pandas._config.config.OptionError: "No such keys(s): 'mode.use_inf_as_null'"的问题，请问您使用的哪个版本的python？是否遇到过这个问题？

下面是我的Python和Pandas

test.py问题

在训练的时候glops计算量不显示是怎么回事

在训练的时候glops计算量不显示是怎么回事

solved

Download via Baidu

Thank you for releasing the code!

Results from the proposed method is quite impressive to me.

I tried to test the code with pretrained weights,

but the download link via Baidu is not possible for me(maybe because it's outside of china)

Is there any other way or link to download the pretrained weights?

Thank you,

没找到train_up.py文件

训练问题

你好，看来您的代码，请问图片是{a}_ir.png和{a}_co.png，而标签是{a}导致无法训练这种问题该如何解决呢？

SR part error : Input and output sizes should be greater than 0, but got input (H: 16, W: 16) output (H: 0, W: 0)

Here, factor = 1 and low_level_feat = [64, 64]
please help me. .

Error message :
Exception has occurred: RuntimeError
Input and output sizes should be greater than 0, but got input (H: 16, W: 16) output (H: 0, W: 0)

File "/home/server-3/desktop/yeona/SuperYOLO-main/models/sr_decoder_noBN_noD.py", line 38, in forward
x = F.interpolate(x, size=[i*(factor // 2) for i in low_level_feat.size()[2:]], mode='bilinear', align_corners=True)
File "/home/server-3/desktop/yeona/SuperYOLO-main/models/deeplabedsr.py", line 60, in forward
x_sr= self.sr_decoder(x, low_level_feat,self.factor)
File "/home/server-3/desktop/yeona/SuperYOLO-main/models/SRyolo.py", line 243, in forward_once
output_sr = self.model_up(y[self.l1],y[self.l2]) #在超分上加attention
File "/home/server-3/desktop/yeona/SuperYOLO-main/models/SRyolo.py", line 190, in forward
y,output_sr,features = self.forward_once(steam,'yolo', profile) #zjq
File "/home/server-3/desktop/yeona/SuperYOLO-main/models/SRyolo.py", line 117, in init
m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch_steam, s, s),torch.zeros(1, ch_steam, s, s),input_mode)[0]]) # forward
File "/home/server-3/desktop/yeona/SuperYOLO-main/train.py", line 102, in train
model = Model(opt.cfg, input_mode = opt.input_mode ,ch_steam=opt.ch_steam,ch=opt.ch, nc=nc, anchors=hyp.get('anchors'),config=None,sr=opt.super,factor=down_factor).to(device) # create
File "/home/server-3/desktop/yeona/SuperYOLO-main/train.py", line 673, in
train(hyp, opt, device, tb_writer)
RuntimeError: Input and output sizes should be greater than 0, but got input (H: 16, W: 16) output (H: 0, W: 0)

超分辨率结果

您好，请问如果想得到最后的超分辨率结果应该如何做呢，我们想得到通过SR分支后的1024的图像，并且计算其PSNR值和SSIM，您有什么建议吗

How to train my custom Dataset

i put my dataset like yolov5 usually but got wrong, thanks

请问yaml文件中的l1 l2 c1 c2的作用是什么

请问yaml文件中的l1 l2 c1 c2的作用是什么

is there a weight file based on the entire VEDAI dataset

test.py: ModuleNotFoundError: No module named 'models.deeplab'

Hi, I am trying to run the test code that is provided.

Only when using the model weights "SuperYOLO_fold1_best.pt" and using the VEDAI dataset, the test.py code isn't working.

Traceback (most recent call last):
File "C:\Users\gsvpk\VScode\Own_Projects\superOG\test.py", line 374, in
test(opt.data,
File "C:\Users\gsvpk\VScode\Own_Projects\superOG\test.py", line 65, in test
model = attempt_load(weights, map_location=device) # load FP32 model
File "C:\Users\gsvpk\VScode\Own_Projects\superOG\models\experimental.py", line 118, in attempt_load
ckpt = torch.load(w, map_location=map_location) # load
File "C:\Users\gsvpk\anaconda3\envs\deeplearning\lib\site-packages\torch\serialization.py", line 809, in load
return _load(opened_zipfile, map_location, pickle_module, **pickle_load_args)
File "C:\Users\gsvpk\anaconda3\envs\deeplearning\lib\site-packages\torch\serialization.py", line 1172, in _load
result = unpickler.load()
File "C:\Users\gsvpk\anaconda3\envs\deeplearning\lib\site-packages\torch\serialization.py", line 1165, in find_class
return super().find_class(mod_name, name)
ModuleNotFoundError: No module named 'models.deeplab'

I have put the weights in the "weights" folder and the VEDAI/VEDAI1024 folders in the "dataset" folder.

Do you have any idea to how this can be resolved? Thank you in advance!

请问作者有没有写detect这个函数

使用自己的数据集训练

使用自己的训练集训练模型应该怎样整理数据集结构？

用small_EDSR_fold1进行test报错：ModuleNotFoundError: No module named 'models.deeplabedsr'

Output images showing up with inverted colors

I've noticed that while evaluating the model after finetuning on my own data, the output images in the runs/test folder are sometimes negative (inverted colors of the image inputted). This is not consistent, as an inverted image in one run will not be the same as another. I've looked through the code, but couldn't find anything that would be causing this or be using it as a processing step. Any help would be appreciated!

train_up.py文件未找到

希望作者早日发布，谢谢

train on visDrone

hello
thanks for implentation sharing
it was interesting for me.

is it possible to train SUPER-YOLO on other datasets such as VisDrone?

thank you.

图像分辨率

super-yolo训练时，输入图像分辨率是1024的吗，在transform.py文件以及提供的数据集中都看到，训练时采用10241024图片作为输入，512512图片作为测试。文章中的结果也是按照这个设置展开的吗？

如何修改batch_size

张博士您好，代码中修改batch_size的地方应该在哪里，配置文件里面没有这一处
在运行代mf模块的train文件中报错ZeroDivisionError: float division by zero


这两个地方的代码怎么理解，ch_in传递变成了64-3-1

mAP comparison with YOLOv5

Hi, and thanks for this great work.
I trained SuperYOLO with train_image_size =1024 and test_image_size=512 on VisDrone dataset and I want to compare the mAP result with YOLOv5, but I was not sure if I have to compare SuperYOLO with YOLOv5 that was trained on 512 image-size or 1024 image-size?

有预训练权重文件吗？

有预训练权重文件吗？

10-Fold data

Thank you for the google link of pretrained data.

I tested the pretrained data and have another question.

I downloaded the VEDAI dataset from the web site of "https://downloads.greyc.fr/vedai/"

because of the same link problem.

There are 10 fold related text files(fold01.txt~fold10.txt) in the folder of "Annotations512" from the web site.

Are those files same to the files of [baiduyun](your Baidu link)?

Thank you,

Super Resolution Module

Hi, thanks for your amazing works. I have some question about SR module.
This paper achieves small object detection without utilizing the super resolution modul. So how this research paper achieves small object detection.
Thanks a lot

数据集载入问题

Scaled weight_decay = 0.0005
train: Scanning 'D:\xxx\SuperYOLO-main\dataset\VEDAI_1024\images.cache' for images and labels... 0 found, 1089 missing, 0 empty, 0 corrupted: 100%|██████████| 1089/1089 [00:00<?, ?it/s]
Traceback (most recent call last):
File "D:/xxx/SuperYOLO-main/train.py", line 673, in
train(hyp, opt, device, tb_writer)
File "D:/xxx/SuperYOLO-main/train.py", line 212, in train
dataloader, dataset = create_dataloader(train_path, imgsz, batch_size, gs, opt,
File "D:\xxx\SuperYOLO-main\utils\datasets.py", line 92, in create_dataloader_sr
dataset = LoadImagesAndLabels_sr(path, imgsz, batch_size,
File "D:\xxx\SuperYOLO-main\utils\datasets.py", line 701, in init
assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {help_url}'
AssertionError: train: No labels in D:\xxx\SuperYOLO-main\dataset\VEDAI_1024\images.cache. Can not train without labels. See https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data

Process finished with exit code 1

> 你好，我看了你的源码，但是我的代码能力很有限

          > 你好，我看了你的源码，但是我的代码能力很有限

没有看到在哪的代码中体现出来测试的时候512的图没有经过下采样

没有downsample

Originally posted by @icey-zhang in #26 (comment)

对于没有lr的数据集怎么训练呢？

我在DOTA数据集上进行训练时，参数设为RGB无法跑通，想问下有什么解决方法呢？谢谢

完全来源了吗

请问代码已经完全开源了吗

如何训练自己的数据集？

非常感谢您的工作，我能够复现您在VEDAI数据集上的训练过程，但我更想使用您的模型在自己的数据集上进行训练，可以麻烦您在readme里给出一些指导吗？
因为我把您给出训练命令中的数据集文件直接替换后并不能进行训练，替换后的命令如下：
python train.py --cfg models/SRyolo_noFocus_small.yaml --super --train_img_size 1024 --hr_input --data data/custom.yaml --ch 3 --input_mode RGB --epochs 1 --device 2
此外我对于train.py里面的--ch_steam和--ch意义也不太明了，这给我带来了很大的困扰。如您愿意给予更多的指导，我也非常愿意进行有偿咨询！

train

您好，我这边在训练时，按照您的步骤先下载数据集然后transform,运行只训练RGB,还是出现标签找不到：No labels in D:\experienment\SuperYOLO-main\SuperYOLO-main\dataset\VEDAI_1024\images.cache. Can not train without labels；您知道这该如何解决么