Great work!
Could you please make some PRs to MMSegmentation?

你好，请教一下单机多卡问题，
不管设置gpu数量为多少，实际只使用了一张卡
默认的bs=8，单卡显存可以吃满，改为16还是一张卡，也没有溢出
无任何报错

hi，i wanna train the model with 4 gpus in one machine
however, your code 'distrain.sh' and 'train.py' can only train with distributed mode in multi-machine
how can i modify the code ?

... Hello, thank you for your outstanding contribution to the open source code core, the performance of IDRNet is very impressive, I would like to know how to configure mask2former+Idrnet, as described in your paper

For example, PSPNet resnet50, cityscapes 1024 * 512

se_cf in the mobilenetV3, what does this bias mean, only keep an activation function on the line, otherwise it reports an error

作者您好，按照代码中的指令bash scripts/distest.sh 1 ssseg/configs/memorynetv2/memorynetv2_upernet_swinlarge_vspw.py memorynetv2_upernet_swinlarge_vspw/epoch_240.pth进行测试时，会遇到运行到一半中断的情况，报错原因看上去像是显存爆炸（观测了显存使用情况，直接冲到99%附近），请问可能是什么原因呢，您有遇到过类似情况吗？十分感谢~
显卡是GPU v100；报错bug：RuntimeError: DataLoader worker (pid 211462) is killed by signal: Killed.
Processing 7420/12196 in rank 1: : 7419it [1:48:14, 1.15it/s]WARNING:torch.distributed.elastic.multiprocessing.api:Sending process 211036 closing signal SIGTERM
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 211035) of binary:

How to convert pt to onnx.
ples.
thanks

您好，如果我想要在训练的时同时使用两种损失函数，那么config文件里面损失函数那块该怎么去写

您好，感谢你卓越的贡献，我想请问关于MCIBI++代码的context_within_image_cfg的is_on设置为false，跟论文里面的不一致，请问可以帮我解答一下嘛，谢谢

您好，感谢您开源的代码核突出贡献，我希望可以在自己的数据集上训练语义分割模型，为了保证一致性，请问您对数据集做了哪些预处理？我希望将我的数据集处理到和您在百度网盘中提供的数据集相同的程度。

作者您好，您在论文中提到在训练VSPW数据集时使用了2帧历史帧以获取时序信息，但是在代码中没有看到相关的模块或结构，vspw数据集的读取代码中使用了random.choice()只能随机读取一帧，请问相关部分的代码能麻烦您帮忙指出或提供吗，万分感谢！

Hi, thanks for your wonderful work, I want to ask where i can find ISNet code?

如题，感谢您提供的代码，但是否可以对在Windows环境下的单卡环境进行一定的优化呢，现在的整体代码都是多卡环境用起来很卡手

initmemory.py 请问这个文件的作用是啥？这个文件在训练的时候有不有用到？

Environment (使用环境)

• Installation method (安装方式):
• The version of sssegmentation (版本号):
• Operating system (操作系统):
• Python version (Python版本):

Question description (问题描述)

Screenshot (报错截图)

Advice (修复建议)

How to modify parameters to use single card training?

In addition to modifying the following in config:

SEGMENTOR_CFG.update(distributed{'is_on':False})

Hi @CharlesPikachu ！UNet 也是大名鼎鼎的分割模型啊，它在医学图像分割领域是 SOTA，个人认为 Supported Models 列表里应该有名字，而且应该在 FCN 之后。哈哈哈 🥇

虽然 PyTorch Hub 已经有预训练的 UNet 了，但我想要皮卡丘也有！ 🛩️

这里提供一些医学数据集给你参考：

• PyTorch Hub 里预训练 UNet 的数据集：LGG Segmentation Dataset
• DDTI，已经按照下一个数据集的格式预处理过，百度网盘提取码：qxds
• tn-scui2020，来自比赛 tn-scui2020
• 百度 AI Studio 医学影像数据集集锦
• List of Medical Datasets，来自 GitHub

Thanks for your work.
Will the pre-trained models available?

作者您好，我看您的代码中是不是没有训练时候的验证代码？

If there are more than one 'num_feats_per_cls' in the furture, 'break' will make this for loop only update the first memory_feature?

Environment (使用环境)

• Installation method (安装方式): https://sssegmentation.readthedocs.io/en/latest/GetStarted.html#installation
• The version of sssegmentation (版本号): latest (https://github.com/SegmentationBLWX/sssegmentation.git)
• Operating system (操作系统): Ubuntu 18.04.6 LTS
• Python version (Python版本): Python 3.8.16

Question description (问题描述)

Building wheel for chainercv (setup.py) ... error
error: subprocess-exited-with-error

× python setup.py bdist_wheel did not run successfully.
│ exit code: 1
╰─> [57 lines of output]

Hello .Thanks for your wonderful work.
What is your different compared to mmseg?Since both of ssseg and mmseg are based on mmcv and need a config to run .

非常精彩的论文，

有个问题，根据我的代码阅读，MemoryNetV2在推理的时候也会过self-attention block

是不是考虑在推理的时候换一种attention模块去做历史特征的聚合，这样会快比较多，也不会显著掉点？

note: This error originates from a subprocess, and is likely not a problem with pip.
ERROR: Failed building wheel for chainercv
Running setup.py clean for chainercv
Building wheel for chainer (setup.py) ... done
Created wheel for chainer: filename=chainer-7.8.1-py3-none-any.whl size=967716 sha256=8c04b26d59f9c9a7948bdaddf433a834e23959b9a1cc52bfc118919187807a64
Stored in directory: /home/mohammed/.cache/pip/wheels/c5/01/d5/3ef194fa83dbf1d97bab826ca2bf235071f412861189a7b9df
Building wheel for iopath (setup.py) ... done
Created wheel for iopath: filename=iopath-0.1.10-py3-none-any.whl size=31530 sha256=6bf140ecbf8c873308103a7c8ebd313cb75226c9d5673e94ac6333a0b1604b88
Stored in directory: /home/mohammed/.cache/pip/wheels/89/3e/24/0f349c0b2eeb6965903035f3b00dbb5c9bea437b4a2f18d82c
Building wheel for typing (setup.py) ... done
Created wheel for typing: filename=typing-3.7.4.3-py3-none-any.whl size=26305 sha256=3b4c4b6304afadcb8fc95d74ae6dd87c9bcd4b92f4c42dcdf90faad9a077f294
Stored in directory: /home/mohammed/.cache/pip/wheels/5e/5d/01/3083e091b57809dad979ea543def62d9d878950e3e74f0c930
Successfully built fvcore chainer iopath typing
Failed to build chainercv
ERROR: Could not build wheels for chainercv, which is required to install pyproject.toml-based projects

您好，我使用您的MemoryNet进行医学图像分割，当和OCRNet训练相同轮数的时候，MemoryNet的miou比OCRNet低很多，这是什么原因

In ISNet, the models are fine-tuned for 160K iterations on ADE20K and LIP datasets. So do you pretrained the models on other datasets?

train the datasets by mask and imgs . how to finish it

Hi, I am learning about Segmentation and want to try out the segmentation my custom data set. Could you please provide steps on how to use supported backbones with some particular architectures?

If I want to use SWIN-B as my backbone on DeepLabV3+ using a custom dataset, what should be the commands and all. I could not find anything on the docs and on the github page. Could you please help.

Screenshot (报错截图)
train.py line 173

Advice (修复建议)
cfg_file_path = self.cfg_file_path

Thanks for your work.
Will the pre-trained models available?

Environment (使用环境)

• Installation method (安装方式): Install SSSegmentation for Developing
• The version of sssegmentation (版本号): 1.5.3
• Operating system (操作系统): Linux。使用slurm任务调度工具。
• Python version (Python版本): 3.8.19

Question description (问题描述)
我将unet模型的某个训练配置文件稍加修改，得到了fcn_unets5os16_spider.py并将该文件作为训练我自己数据集的配置，bs=32，epochs设为20。按默认设置，每隔10个epoch在验证集val上验证效果。训练返回的结果文件中，第20个epoch结束后的验证结果如下【结果1】：

2024-04-08 12:06:49 INFO  Rank 0 finished
2024-04-08 12:06:49 INFO  All Finished, all_preds: 480, all_gts: 480
2024-04-08 12:06:51 INFO  {'iou': {'__background__': 0.9306402134161759, 'spider': 0.6665832034694705}, 'miou': 0.7986117084428233, 'accuracy': array([0.96680556, 0.78754992]), 'mpa': 0.8771777369668247}

(accuracy和mpa是我加进去的)

之后我用测试脚本dist_test.sh，加载了第20个epoch结束后保存的参数文件epoch_20.pth进行验证，调用测试脚本的命令如下：

bash scripts/dist_test.sh 1 ssseg/configs/unet/fcn_unets5os16_spider.py fcn_unets5os16_spider_ep20/epoch_20.pth --slurm

单独运行测试脚本得到的结果如下【结果2】：

2024-04-08 15:22:14 INFO  Rank 0 finished
2024-04-08 15:22:14 INFO  All Finished, all_preds: 480, all_gts: 480
2024-04-08 15:22:16 INFO  {'iou': {'__background__': 0.9324833796798122, 'spider': 0.6735323713466238}, 'miou': 0.803007875513218, 'accuracy': array([0.96818855, 0.79066773]), 'mpa': 0.879428142561482}

这和前面训练时的验证结果不一样。我之后又运行了一次单独运行测试脚本的命令，结果和【结果2】一样。

请问为什么会这样？

imagelevel.py :
47: feats_il = self.correlate_net(x, torch.cat([x_global, x], dim=1))

isanet.py：
47：context = super(SelfAttentionBlock, self).forward(x, x)

is there any problem？ bug？

您好，我有点疑惑，请问VOC, Cityscapes , Ade20k数据集的正则化都是：
('Normalize', {'mean': [123.675, 116.28, 103.53], 'std': [58.395, 57.12, 57.375]}), 嘛

您好，我最近看了您的mcibi论文，很精彩，我想请问一下论文中的表格3的模型的复杂度和FLOPS是如何获取的

Environment (使用环境)

• Installation method (安装方式): Install SSSegmentation for Developing
• The version of sssegmentation (版本号): 1.5.3
• Operating system (操作系统): Linux。使用slurm任务调度工具。
• Python version (Python版本): 3.8.19

Question description (问题描述)
请问SSSegmentation能否在单机单卡情况下使用，如果可以，请问应该修改哪些配置参数？

• For Custom Datasets
• For Fine-tuning

Hi,
I want to train my own dataset which has images in 4 channels - RGB images and IR(infrared) images. Could you help me out with that? How can i modify the codes of this repo to accommodate that extra channel?

Description
ssseg/modules/models/segmentors/base/base.py line128, it seems should use nn.SyncBatchNorm

Screenshot

Advice
change SyncBatchNorm to nn.SyncBatchNorm

(venv) mohammed@c24032:~/sssegmentation$ bash scripts/dist_train.sh 4 ssseg/configs/annnet/annnet_resnet50os16_ade20k.py

Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.

2024-01-10 08:33:59 WARNING ngpus_per_node is not equal to nproc_per_node, force ngpus_per_node = nproc_per_node by default
Traceback (most recent call last):
File "ssseg/train.py", line 252, in
main()
File "ssseg/train.py", line 247, in main
client.start()
File "ssseg/train.py", line 70, in start
torch.cuda.set_device(cmd_args.local_rank)
File "/home/mohammed/model/new ss/venv/lib/python3.8/site-packages/torch/cuda/init.py", line 261, in set_device
torch._C._cuda_setDevice(device)

the previous error has gone when i made the number of classes 2, now left with this.
File "/home/mohammed/sssegmentation/ssseg/modules/datasets/ade20k.py", line 83, in
class ADE20kDataset(BaseDataset):
File "/home/mohammed/sssegmentation/ssseg/modules/datasets/ade20k.py", line 108, in ADE20kDataset
assert num_classes == len(classnames) and num_classes == len(palette)
AssertionError
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 1615854) of binary: /home/mohammed/miniconda3/envs/myenv/bin/python
Traceback (most recent call last):

i set the number of classes as 2 which you can see below.
class ADE20kDataset(BaseDataset):
num_classes = 2
classnames = ['defect']
palette = [(255, 0, 255)]
clsid2label = {0: 0}

when i set it like this

num_classes = 2

        # classnames = ['defect', 'no_defect']
        # palette = [(255, 0, 255), (0, 255, 0)]
        # clsid2label = {0: 0, 1: 1}

it leads to the previous error were it shows t >= 0 && t < n_classes failed.

the below is how my scenecategory looks

NT_label_2 defect
NT_label_1 defect
NF_label_1 defect
NF_label_0 defect
N7_label_1 defect
N7_label_0 defect
N4_label_2 defect
N4_label_1 defect
N4_label_0 defect
N3_label_3 defect
N3_label_2 defect
N3_label_1 defect
N3_label_0 defect
N2_label_2 defect
N2_label_1 defect
N2_label_0 defect
N1_label_3 defect
N1_label_2 defect
N1_label_1 defect
N1_label_0 defect
m9614_label_9 defect
m9614_label_8 defect
M9287_label_6 defect
M9287_label_4 defect
M9287_label_3 defect
M9287_label_1 defect
M7436_label_8 defect
M7436_label_6 defect
M7436_label_2 defect

Hi @CharlesPikachu
Are the results of all the models in this repo single scale or multi-scale? best model or last model?

ModuleNotFoundError: No module named 'mmcv.ops'