data problem about pytorch-segmentation HOT 9 CLOSED

Hi,

Are you talking generally or for a specific dataset ?
If you want to add you dataset, just create a new loader in dataloader/, the new class inherits from base_loader and you only need to implement _set_files (setting the paths of the images and the labels) and _load_data (where you load a specific pair of image / label depending on the argument passes), the data directory is up to you, check the examples of the implemented loaders.

For one of the four datasets, in read me I've detailed the files you'll need, if you have any more questions about a specific dataset, I can detail its files.

from pytorch-segmentation.

from pytorch-segmentation.

It is up to you how you load you dataset, for creating a loader you have two ways of doing things, creating a loader that takes an argument split (val and train), and then you create two dataloaders, one for each split, this is the way the provided datasets are implemented. Or you can only create one dataloader and then pass the argument val_split and this way, you'll get a train loader and a also the val loader (calling get_val_loader). Take a look into base_loader.py and base_datasets.py and one of the example loaders to see how this is implemented.

All in all, first is to create your dataloader, by implementing two functions _set_files and _load_data

from pytorch-segmentation.

i see，but i dont know the role of voc.txt、voc_context_classes-400.txt、voc_context_classes-59.txt.

from pytorch-segmentation.

Oh, you can just ignore them, they're not used, they only map the class indices to the class name for each dataset, can be somewhat useful when debugging or you'd like to know the class predicted.

from pytorch-segmentation.

your mean is,if i use this model to test my data,it will show me the label represented by the color.

from pytorch-segmentation.

Not really no, those files are not used at all, for the color palettes, their defined in utils/palette.py.

from pytorch-segmentation.

i see,thanks a lot!

from pytorch-segmentation.

@yassouali

sorry for reopening the issue, but i searched the closed issues regarding the custom dataloader. I am creating custom dataloader for camvid dataset like your voc loader implementation. i have the raw image and segment maps in png. The train, train_val, val, test texts must be created? what is the difference between train_val and val?

edit: i get error for created dataloader as
return getattr(module, config[name]['type'])(*args, **config[name]['args'])
AttributeError: module 'dataloaders' has no attribute 'camvid'

thank you

from pytorch-segmentation.