xmed-lab / allspark Goto Github PK
View Code? Open in Web Editor NEWCVPR 2024: AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation
License: MIT License
CVPR 2024: AllSpark: Reborn Labeled Features from Unlabeled in Transformer for Semi-Supervised Semantic Segmentation
License: MIT License
感谢出彩的工作!
代码AllSpark中有一个InstanceNorm层
self.psi = nn.InstanceNorm2d(self.num_heads)
请问为什么用InstanceNorm以及为什么参数设置为num_heads呢?
k_l2u = rearrange(k_l2u, 'b n c -> n (b c)')
v_l2u = rearrange(v_l2u, 'b n c -> n (b c)')
k_l2u = repeat(k_l2u, 'n bc -> r n bc', r=batch_size)
v_l2u = repeat(v_l2u, 'n bc -> r n bc', r=batch_size)
还有cross attention中的这一段代码是什么作用呢?
Hi, thanks for your inspiring work. I notice that there is no quality control on pseudo labels. How is that possible to guarantee the training quality. I am also surprised that no cutmix augmentation is used in your work. Can you shed some light on how allspark can achieve such good performance without label filtering and strong augmentation?
20240429_205311.log
您好,如log文件所示,在训练到21epoch时,突然出现损失函数变大,性能下降的问题,请问这是否属于正常现象?
thanks for your work ,it's actually details.I followed your readme and install necessary package by sequence.but enconter the error:
I try some solutions such as pip install pip upgrade pip setuptools contain uninstall and install ,and I confirm the config path of cuda in bashrc contain CUDA_Home and LD_LIBRARY ,which point to the folder that contain bin lib and include in my conda envs .Due to the fault in the base env is not exist , I guess that is not related to the cuda .But the gpt suggest that base env has not cuda so that don't affect the installation of mmcv.Besides,the version of mmcv I get it's match your readme called 1.6.2 from https://pypi.tuna.tsinghua.edu.cn/simple to faster.About the cuda, due to it's binding to the torch so that .so called libtorch_cuda.so.In my a test.py in Generic_ssl,I use the ctypes..CDLL(library_path) to load them, and in the mmcv setup.py its seems so complicated.
Looking forward to hearing from your reply!
非常感谢你们出色的工作!!
我将mit_b5换掉,用了一个新的backbone运行,参数量19.2M,在VOC 2012 Full(1464)上进行训练,发现miou达到了86%,但是我用我自己数据集训练的时候,发现了一个很奇怪的现象。我用1/2data训练时miou先达到了95.3%,到了64个epoch之后,逐步降低到了55%。这可能是什么原因呢?我在全监督网络segformer训练我自己的模型,miou能达到97%,且细节信息分割地更加完整,但是用你们的模型时细节信息却不是那么完整,我也尝试了别的半监督框架,如ST++,但细节信息都差强人意,请问这个可能是什么原因呢?期待您的回答,谢谢!
作者您好,请问我想尝试将编码器换成Resnet50/101,解码器换成deeplabv3+。这需要更改哪里呀!
Thanks for sharing this nice work.
However, when i try to reproduce the results using your provided checkpoints, it seems that the checkpoints does not match the model architecture.
RuntimeError: Error(s) in loading state_dict for DistributedDataParallel:
Missing key(s) in state_dict: "module.decoder.allspark.map_in.0.weight", "module.decoder.allspark.map_in.0.bias", "module.decoder.allspark.attn_norm.weight", "module.decoder.allspark.attn_norm.bias", "module.decoder.allspark.attn.queue_ptr", "module.decoder.allspark.attn.kv_queue", "module.decoder.allspark.attn.q_u.weight", "module.decoder.allspark.attn.k_u.weight", "module.decoder.allspark.attn.v_u.weight", "module.decoder.allspark.attn.q_l2u.weight", "module.decoder.allspark.attn.k_l2u.weight", "module.decoder.allspark.attn.v_l2u.weight", "module.decoder.allspark.attn.out_u.weight", "module.decoder.allspark.attn.out_l2u.weight", "module.decoder.allspark.encoder_norm.weight", "module.decoder.allspark.encoder_norm.bias", "module.decoder.allspark.map_out.0.weight", "module.decoder.allspark.map_out.0.bias".
Unexpected key(s) in state_dict: "module.decoder.allspark.embeddings.patch_embeddings.weight", "module.decoder.allspark.embeddings.patch_embeddings.bias", "module.decoder.allspark.encoder.layer.0.attn_norm1.weight", "module.decoder.allspark.encoder.layer.0.attn_norm1.bias", "module.decoder.allspark.encoder.layer.0.attn.queue_ptr", "module.decoder.allspark.encoder.layer.0.attn.kv_queue", "module.decoder.allspark.encoder.layer.0.attn.query_u_u.weight", "module.decoder.allspark.encoder.layer.0.attn.key_u_u.weight", "module.decoder.allspark.encoder.layer.0.attn.value_u_u.weight", "module.decoder.allspark.encoder.layer.0.attn.query_l_u.weight", "module.decoder.allspark.encoder.layer.0.attn.key_l_u.weight", "module.decoder.allspark.encoder.layer.0.attn.value_l_u.weight", "module.decoder.allspark.encoder.layer.0.attn.out_u.weight", "module.decoder.allspark.encoder.layer.0.attn.out_l.weight", "module.decoder.allspark.encoder.layer.0.ffn_norm1.weight", "module.decoder.allspark.encoder.layer.0.ffn_norm1.bias", "module.decoder.allspark.encoder.layer.0.ffn1.fc1.weight", "module.decoder.allspark.encoder.layer.0.ffn1.fc1.bias", "module.decoder.allspark.encoder.layer.0.ffn1.fc2.weight", "module.decoder.allspark.encoder.layer.0.ffn1.fc2.bias", "module.decoder.allspark.encoder.layer.1.attn_norm1.weight", "module.decoder.allspark.encoder.layer.1.attn_norm1.bias", "module.decoder.allspark.encoder.layer.1.attn.queue_ptr", "module.decoder.allspark.encoder.layer.1.attn.kv_queue", "module.decoder.allspark.encoder.layer.1.attn.query_u_u.weight", "module.decoder.allspark.encoder.layer.1.attn.key_u_u.weight", "module.decoder.allspark.encoder.layer.1.attn.value_u_u.weight", "module.decoder.allspark.encoder.layer.1.attn.query_l_u.weight", "module.decoder.allspark.encoder.layer.1.attn.key_l_u.weight", "module.decoder.allspark.encoder.layer.1.attn.value_l_u.weight", "module.decoder.allspark.encoder.layer.1.attn.out_u.weight", "module.decoder.allspark.encoder.layer.1.attn.out_l.weight", "module.decoder.allspark.encoder.layer.1.ffn_norm1.weight", "module.decoder.allspark.encoder.layer.1.ffn_norm1.bias", "module.decoder.allspark.encoder.layer.1.ffn1.fc1.weight", "module.decoder.allspark.encoder.layer.1.ffn1.fc1.bias", "module.decoder.allspark.encoder.layer.1.ffn1.fc2.weight", "module.decoder.allspark.encoder.layer.1.ffn1.fc2.bias", "module.decoder.allspark.encoder.encoder_norm1.weight", "module.decoder.allspark.encoder.encoder_norm1.bias", "module.decoder.allspark.recon.conv.weight", "module.decoder.allspark.recon.conv.bias".
Is this problem caused by the incorrect allspark checkpoints, or my misuse? If possible, could you provide a evaluation script to reproduce the results?
Great work guys! I would like to know how much memory its necessary to train?
when will the final version release? i am very instreast in your reasearch work because of the excellent result on the leadboad.but sadly the code is not ready for use, do you any plan about when the final version will be released?
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.