Pretraining with my own dataset about spark HOT 3 CLOSED

Yes, you can refer to /pretrain/README.md for the complete cmd, which is:

$ cd /path/to/SparK/pretrain
$ torchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr=localhost --master_port=<some_port> main.py \
  --data_path=/path/to/imagenet --exp_name=<your_exp_name> --exp_dir=/path/to/logdir \
  --model=resnet50 --bs=512

The first line is missing, e.g. main.py is the training_script.

from spark.

Thank you very much. In the article, it is saying that 'All models are pre-trained with 1.28 million unlabeled images
from ImageNet-1K (Deng et al., 2009) training set for 1600 epochs.', which means that I can use SparK pretraining with unlabeled dataset. However, when I tried to use unlabeled image dataset , the following error is happening:
File "/home/user/SparK/pretrain/utils/imagenet.py", line 39, in init
super(ImageNetDataset, self).init(
File "/home/user/.local/lib/python3.10/site-packages/torchvision/datasets/folder.py", line 144, in init
classes, class_to_idx = self.find_classes(self.root)
File "/home/user/.local/lib/python3.10/site-packages/torchvision/datasets/folder.py", line 218, in find_classes
return find_classes(directory)
File "/home/user/.local/lib/python3.10/site-packages/torchvision/datasets/folder.py", line 42, in find_classes
raise FileNotFoundError(f"Couldn't find any class folder in {directory}.")

from spark.

You need to define a new Python class for your dataset, to replace our ImageNetDataset in https://github.com/keyu-tian/SparK/blob/main/pretrain/utils/imagenet.py#L30. Just define a class with __len__(self) and __getitem__(self, index: int) implemented. The getitem should return the index-th image in your dataset, and be processed by a transformation like the trans_train in /pretrain/utils/imagenet.py.

PS: i recommend to try your pretraining with or without --init_weight=/path/to/res50_withdecoder_1kpretrained_spark_style.pth. If this arg is used, you will pretrain from our pretrained model, rather than from scratch, which could be better.

from spark.