xiaofeiwang2018 / dr

Joint Learning of Multi-level Tasks for Diabetic Retinopathy Grading on Low-resolution Fundus Images

dr's Issues

This is not the right code for an indicated article. Can you share the updated code?

MultiScale_Intergrate should not remove the grad_fn?

Should we use F.interpolate instead of resizing it to prevent grad_fn disapeering?

Code Issue?

            extractor = ModelOutputs(cls_net)
            features, fc_output = extractor(cls_input)
            cam_all_sr = {'size256': [], 'size128': [], 'size64': [], 'size32': []}
            for batch in range(fc_output.shape[0]):
                device = fc_output[batch].device.index
                index = np.argmax(fc_output[batch].cpu().data.numpy())
                if index == 0:
                    one_hot = np.zeros((1, fc_output[batch].size()[-1]), dtype=np.float32)
                    one_hot[0][0] = 1
                if index > 0:
                    one_hot = np.ones((1, fc_output[batch].size()[-1]), dtype=np.float32)
                    one_hot[0][0] = 0
                if batch == 0:
                    one_hot_all = Variable(torch.from_numpy(one_hot).cuda(device), requires_grad=True)
                else:
                    one_hot_all = torch.cat((Variable(torch.from_numpy(one_hot).cuda(device), requires_grad=True), one_hot_all), 0)

if the index of batches is the same, there is no problem. But if they are different, it calculates the opposite loss between one_hot_all and fc_output.

Example:

fc_output one_hot_all
[[0.9, 0.1, 0.1, 0.1, 0.1] [[1, 0, 0, 0, 0]
[0.8, 0.2, 0.2, 0.3, 0.4]] [1, 0, 0, 0, 0]] There is no problem.

fc_output one_hot_all
[[0.2, 0.1, 0.1, 0.1, 0.8] [[0, 1, 1, 1, 1]
[0.3, 0.2, 0.7, 0.3, 0.4]] [0, 1, 1, 1, 1]] There is no problem.

fc_output one_hot_all
[[0.9, 0.1, 0.1, 0.1, 0.7] [[0, 1, 1, 1, 1]
[0.3, 0.2, 0.4, 0.9, 0.4]] [1, 0, 0, 0, 0]] There is problem here because of this line torch.cat((Variable(torch.from_numpy(one_hot).cuda(device), requires_grad=True), one_hot_all), 0).

should be one_hot_all = torch.cat((one_hot_all, Variable(torch.from_numpy(one_hot).cuda(device), requires_grad=True)), 0) ???

Thanks

GMSV algorithm?

Can you share the code of the GMSV algorithm in the article, not in the conference paper?

RandomCrop如何导入

您好，可以麻烦问一下这个怎么导入嘛？我搜不到这个库呀？
from transform.transforms_group import RandomCrop
我运行代码，一直在这个地方报错是怎么回事呢？
if self.transform is not None: # train
info = self.hr_transform(info)
报错如下
File "E:\MyUse\CVcode\DR-main\data_process.py", line 91, in getitem
info = self.hr_transform(info)
File "D:\Soft\Anaconda\envs\pytorchgpu_38\lib\site-packages\torchvision\transforms\transforms.py", line 60, in call
img = t(img)
File "D:\Soft\Anaconda\envs\pytorchgpu_38\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "D:\Soft\Anaconda\envs\pytorchgpu_38\lib\site-packages\torchvision\transforms\transforms.py", line 586, in forward
width, height = F._get_image_size(img)
File "D:\Soft\Anaconda\envs\pytorchgpu_38\lib\site-packages\torchvision\transforms\functional.py", line 67, in _get_image_size
return F_pil._get_image_size(img)
File "D:\Soft\Anaconda\envs\pytorchgpu_38\lib\site-packages\torchvision\transforms\functional_pil.py", line 26, in _get_image_size
raise TypeError("Unexpected type {}".format(type(img)))
TypeError: Unexpected type <class 'list'>

there is no U_Net_Cut model in network file?

Can you update the files?

about the number of epochs

您好，请问在第一阶段训练中，epoch数真的是30000吗？