I'm curious about configuring datasets for training.

[a510721]

🤔 What’s your question?

I'm curious about how to structure my datasets for training. I'm following the coco format and have the images and annotation.json files in the same folder.

.train/images
├── annotation.json
├── image1.png
├── image2.png

The code is as follows:
import lightly_train

def main():

lightly_train.train_instance_segmentation(
    out="out/my_experiment",
    overwrite=True,
    model="dinov3/vits16-eomt-inst-coco",
    data={
        "path": "/home/kys/workspace/code/thrid_party/detector/datasets/light_train",
        "train": "train/images",
        "val": "valid/images",
        "names": {
            0: "person",
        },
    },
)

The error is as follows:
Train images: 0, Val images: 0
Detected dataset size 0 and batch size 16. Reducing batch size to 0.
<lightly_train._data.yolo_instance_segmentation_dataset.YOLOInstanceSegmentationDataset object at 0x703ef9a6eea0>
Traceback (most recent call last):
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/main.py", line 42, in
main()
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/main.py", line 11, in main
lightly_train.train_instance_segmentation(
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/venv/lib/python3.12/site-packages/lightly_train/_commands/train_task.py", line 189, in train_instance_segmentation
return _train_task(config_cls=InstanceSegmentationTrainTaskConfig, **locals())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/venv/lib/python3.12/site-packages/lightly_train/_commands/train_task.py", line 499, in _train_task
_train_task_from_config(config=config)
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/venv/lib/python3.12/site-packages/lightly_train/_commands/train_task.py", line 649, in _train_task_from_config
train_dataloader = helpers.get_train_dataloader(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/venv/lib/python3.12/site-packages/lightly_train/_commands/train_task_helpers.py", line 581, in get_train_dataloader
dataloader = DataLoader(**dataloader_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/venv/lib/python3.12/site-packages/torch/utils/data/dataloader.py", line 388, in init
sampler = RandomSampler(dataset, generator=generator) # type: ignore[arg-type]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/kys/workspace/code/thrid_party/detector/lightly-train/venv/lib/python3.12/site-packages/torch/utils/data/sampler.py", line 162, in init
raise ValueError(
ValueError: num_samples should be a positive integer value, but got num_samples=0

[stegmuel]

Hi,

For instance segmentation, we use the YOLO format.
The expected data format is detailed here.
This tool can be used to convert your dataset annotations from COCO to YOLO format.