search

chaosallen / IPNV2_pytorch

This is an pytorch implementation of "IPN-V2 and OCTA-500: Methodology and Database for Retinal Image Segmentation".
61 stars 10 forks source link

Preprocessing Images #3

Open zengmmm00 opened 3 years ago

zengmmm00 commented 3 years ago

Hi.

How should I prepocess the labels in the OCTA-500 dataset(/OCTA-500/OCTA-500_ground_truth/OCTA-500/OCTA_3M/GroundTruth) before putting them in to Label_RV? Without preprocessing the data in train_annotations train_annotations = train_annotations.to(device=device, dtype=torch.long) ranges from 0 to 255. However, the data in pred pred,_= net(train_images) ranges from -1 to 1. Thus, loss = criterion(pred, train_annotations) result in RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED .

Besides, do I need to run IPN before IPN_V2 and use the output of IPN as the input of IPN_V2?

chaosallen commented 3 years ago

If you download the database from https://ieee-dataport.org/open-access/octa-500, you need to preprocess the downloaded labels. Two operations are required： 1）Rotate 90 degrees：To align with 3D data 2）Change the gray value of the label image to：0-background 1-RV （2-FAZ，if you need）

WangHan-Chris commented 2 years ago

I rotated the label 90 degrees and changed the gray value of the label image to：0-background 1-RV，but when I ran the IPN_V2_train.py, the loss was always 0 after 60 iterations.

kaiyuxin commented 3 months ago

我将标签旋转了 90 度，并将标签图像的灰度值更改为：0-背景 1-RV，但是当我运行IPN_V2_train.py时，60 次迭代后损失始终为 0。

请问能不能帮助一下怎么预处理数据集呀

kaiyuxin commented 3 months ago

如果从 https://ieee-dataport.org/open-access/octa-500 下载数据库，则需要对下载的标签进行预处理。需要两个操作：1）旋转90度：与3D数据对齐 2）将标签图像的灰度值更改为：0-背景1-RV（2-FAZ，如果需要）

你好。

在将 OCTA-500 数据集（/OCTA-500/OCTA-500_ground_truth/OCTA-500/OCTA_3M/GroundTruth）放入Label_RV之前，我应该如何预置标签？如果不进行预处理，train_annotations 中的数据范围为 0 到 255。但是，pred 中的数据范围从 -1 到 1。因此，结果为 .train_annotations = train_annotations.to(device=device, dtype=torch.long)``pred,_= net(train_images)``loss = criterion(pred, train_annotations)``RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED

此外，我是否需要在IPN_V2之前运行 IPN 并使用 IPN 的输出作为 IPN_V2 的输入？

请问，这个标签数据您是用的Label里面的哪一个？