Open miscedence12 opened 1 year ago
I don't have the CamVid weights anymore. The Cityscapes weights are much more useful because Cityscapes is a larger dataset. You can train on the CamVid dataset yourself by preparing the dataset, downloading the Cityscapes trainval pretrained weights, and using the training config here (https://github.com/RolandGao/RegSeg/blob/main/configs/camvid_200epochs.yaml).
@RolandGao hello,Can you provide me with the directory structure of the Camvid dataset you use and show sample images? The data set I prepared by myself seems to be unable to complete the training,thanks!
CamVid does require a bit of preprocessing before it can be used. The labels are converted from rgb to class labels first using something like the rgb_mask2 function in camvid.py. The directory structure is -- train -- train_labels -- val -- val_labels -- test -- test_labels -- class_dict.csv
@RolandGao Thanks for your reply,I have made the following changes to camvid.py. Do you think it is right? My label is rgb image. line43: img = Image.open(self.images[index]).convert('RGB') line44: target = rgb_to_mask2(Image.open(self.masks[index]),color_to_class=self.color_to_class)
I think it should be correct, but you have to test it out yourself to confirm.
@RolandGao I've already verified it. It's right;Recently, I want to train the model on coco dataset. Have you ever trained it before? If so, is there any pretrained weights?
Best wishes! Jie
Hi,
Training on coco should be supported by the repository with the "coco_100epochs.yaml" config file, but I don't have pretrained weights.
Do I need to rewrite a training file to train the Camvid dataset?
No, you can use the same train.py. You only need to prepare the dataset.
@RolandGao Hello,I wonder to know whether you will publish the weights file trained on the Camvid dataset.