mobulan / MPSA

Source code of the paper Multi-Granularity Part Sampling Attention for Fine-Grained Visual Classification
21 stars 3 forks source link

I am facing issues with the BIRADS classification task (6 categories) using breast ultrasound imaging and urgently hope to receive a response. Thank you very much! #2

Open YangJay2004 opened 1 month ago

YangJay2004 commented 1 month ago

我参加了一个比赛,任务是对乳腺超声影像(2500张,各类别不平衡)的BIRADS进行分类任务,所分的类别分别是2类,3类,4A,4B,4C,5类,我尝试了许多模型,但是效果都很差,包括预训练的resnet(acc60%)vit(60%)fastvit(60%),inception_resnet(65%)等等,我在你的项目中使用了你默认的参数(参考了swin-cub.yaml)进行训练,使用swin base达到了70%的正确率!!这让我很激动,并且相信你们的成果的潜力!

但是我不知道接下来该如何继续调参数或换模型来优化,每次train上的acc达到70%模型似乎就开始过拟合了。希望您能给出建议!

还有请问如果需要使用其他backbone,这些.pth的文件应该如何获取呢?我尝试从huggingface下载.bin文件并转换,但尝试了许久做了一些修改还是会报错。希望您能给出建议或下载地址!

比赛的ddl是本周日,这场比赛对我来说很重要,期望能得到您专业的答复!十分感谢!

Here’s the translation of the provided content into English,Since my English is't that good, the content is translated by AI,sorry for that.


I participated in a competition where the task was to classify BIRADS categories for breast ultrasound images (2500 images, with imbalanced classes). The categories I am classifying are 2, 3, 4A, 4B, 4C, and 5. I tried many models, but the results were poor, including pre-trained models like ResNet (acc 60%), ViT (60%), FastViT (60%), Inception_ResNet (65%), etc. In your project, I used your default parameters (referring to swin-cub.yaml) for training, and using the Swin base model, I achieved a 70% accuracy! This excites me and makes me believe in the potential of your results!

However, I am unsure how to continue optimizing by adjusting parameters or changing models, as it seems that every time the training accuracy reaches 70%, the model begins to o

verfit. I hope you can provide some advice!

Additionally, if I need to use other backbones, how can I obtain the .pth files? I tried downloading the .bin files from Hugging Face and converting them, but despite making some modifications after trying for a long time, I still encountered errors. I hope you can provide some suggestions or download links!

The competition deadline is this Sunday, and this competition is very important to me. I look forward to your professional response! Thank you very much!


my email is zhenghuiyang0058@gmail.com ,thanks!

6 classes examples are as following:

1102 1275 BUSI_0488 BUSI_0639 hcjz_birads4A_0059_0001 hcjz_birads4B_0163_0024

mobulan commented 1 month ago

非常感谢你认可我们的工作。首先关于不同模型的预训练模型可以在github上对应的项目找到,考虑到时间问题,我直接上传到坚果云链接如下 resnet50, vit. 对应的加载可以直接使用build.py中进行加载,并且在config文件中创建对应的yaml文件。 关于过拟合问题,我建议尝试直接冻结预训练主干网络,只微调我们方法额外引入的可学习参数,通过这种方法可能一定程度解决你们的问题。

YangJay2004 commented 1 month ago

收到!万分感谢!