yijinhuang / pytorch-classification Goto Github PK

您好，我是北航的一名研究生，非常感谢您在eyepacs上的调参训练结果。

根据您的环境要求，我配置了相同版本的环境。单卡训练时，我发现实验结果是可以复现的。但是我在多卡训练时遇到了一些问题，根据您的设置，random_seed=1，cudnn_deterministic=true，但是实验结果无法复现。可能我错误使用了您的代码，但我没能找到具体的原因，非常抱歉打扰您，能请您帮忙分析一下吗，十分感谢！

祝您安好！

Hi, I am wondering whether the weights of the DR model showcased in the paper is available in the repository?

%%%

1. The "AttributeError" is caused by the pickle mechanism, which means you need to copy the Python script containing the "ContrastiveModel" class to the same folder as "main.py".

%%%

Originally posted by @YijinHuang in #10 (comment)

I still got the same attribute error after doing so. Are there any further detailed instructions about how can I train the classification model to evaluate the weights from Lesion-based-Contrastive-Learning? Or are there any other parts I have to modify in the script?

Thanks for the help!

I have rectangular shape images of (100,50), (height, width). how do I set the input_size value ?

Hi, I'm really interested in your work, however, I came across some issue while using this repository for evaluation on pre-trained models in Lesion-based CL. In the instruction of how to reproduce the experimental results on diabetic retinopathy grading, it said that: "To fine-tune pretrained models, treat the pretrained weights as checkpoints by updating the item "checkpoint" in ~/configs/eyepacs.yaml."
My understanding is that I just have to modify the "checkpoint" item in ~/configs/eyepacs.yaml like this:

checkpoint: /root/Work/Retina_Seg/pytorch-classification/configs/final_model.pt # load weights from other pretrained model

After I trained the Lesion-based CL model, save the training weights and updated the "checkpoint" item in ~/configs/eyepacs.yaml , I run 'main.py', the error message occurred:

AttributeError: Can't get attribute 'ContrastiveModel' on <module 'modules' (<_frozen_importlib_external.NamespaceLoader object at 0x7f85f1d342d0>)>

Also, I changed the pretrained model to the one you provided (ie.resnet50_128_07.pt) and it shows:
RuntimeError: Error(s) in loading state_dict for ResNet:
Missing key(s) in state_dict: "fc.weight", "fc.bias".

I'm thinking how to resolve the error or are there any other step I have missed to evaluate the pre-trained models in Lesion-based CL?
Any help is appreciated!

您好！
又来向您请教问题了，最近又在您的代码上调参数，发现了一件很奇怪的现象。您默认使用的是SGD+nesterov方法，看您的代码也写了ADAM优化器，不知道为什么我把优化器换成ADAM之后，效果非常差，直接无法正常工作。 kappa指标一直为0，acc也保持不变，想来全预测为类别0了，如图所示，请问您碰到了这个情况吗？

微信图片_20220613154349
祝您安好！

您好，我是电子科大的一名研究生，最近也在做糖尿病视网膜病变分级的研究，很兴奋看到您在Resnet50上通过调参达到很好的效果，给后续研究者一个非常高的起点，非常感谢你。

可是在复现您代码的时候遇到了一些问题，结果如上图，那是训练的过程图。
我按照您的数据预处理方法,运行crop.py对图像进行裁切。
配置也是按照您的eyepacs配置的，仅仅改了路径。
唯一的不同点在于我的验证集是从训练集里面划分的，我的训练集，验证集，测试集 分别是 28090 7027 53570，但这个应该不影响。

疑问：为什么训练过程中的Accuracy会这么低呢？ 验证集的kappa也并不高。 我自己使用普通Resnet50，训练集Acc大概90% ，验证集Acc大概84% , kappa能有77%左右。 可能是我使用对您的代码使用的不对，但是我没找到具体原因，所以冒昧打扰您，请您帮忙分析一下，非常感谢您啦。
祝您安好！

I am currently trying to run the main.py file, however I am unable to do so because I get the following error: RuntimeError: 'Tensor' object has no attribute or method 'new_ones'.

This stems from the torcheval.metrics import. Importing torcheval works just fine, and I do have all packages installed according to the requirements file.

Please could you advise?