KeyError: 'my_custom' when fintuning my own data

[zzksdu]

i also met this error, please give me some advices @yuzhms , thanks.
my work env has not docker and i configure the running env by following this steps: git clone this git, cd dexbotic, and pip install -e .
i follow the data.md to configure my own data, like this:
in example_exp.py
class CogActDataConfig(_CogACTDataConfig):

dataset_name: str = field(default='my_custom')

Action configuration - contains action-related parameter settings

action_config: CogActActionConfig = field(default_factory=CogActActionConfig)
creat a my_custom_dataset.py， content is like blows:
from dexbotic.data.data_source.register import register_dataset
import math

Define your dataset
MY_CUSTOM_DATASET = {
"my_robot_data": {
"data_path_prefix": "", # Image path prefix
"annotations": '/xxxxx', # Dataset path
"frequency": 1, # Data sampling frequency
},
}

meta_data = {
'non_delta_mask': [6], # Non-delta dim index, e.g. gripper
'periodic_mask': [3, 4, 5], # Indices of periodic action dimensions (e.g., rotation), used for handling wrapping
'periodic_range': 2 * math.pi # periodic range
}

register_dataset(MY_CUSTOM_DATASET, meta_data=meta_data, prefix='my_custom')
print("register_dataset done")

running command:torchrun --nproc_per_node=2 playground/example_exp.py, get error trace:
File "/xxx/dexbotic/dexbotic/exp/base_exp.py", line 473, in _get_dataset
robot_dataset = DexDataset(
File "/xxx/dexbotic/dexbotic/data/dataset/dex_dataset.py", line 38, in init
self._build_dataset_from_name(data_args.dataset_name)
File "/xxx/dexbotic/dexbotic/data/dataset/dex_dataset.py", line 104, in _build_dataset_from_name
dataset = CONVERSATION_DATA[name]
KeyError: 'my_custom'
so, how to solve this problem?
another question is: does dm0_arch support single image fintuning? @yuzhms

[yuzhms]

Hi! Two points to address your questions:

1. KeyError issue: Your dataset key should be my_custom_my_robot_data (prefix + dataset name). See the example in Data.md for reference.

2. Single image finetuning: Since dm0 was pretrained with 3 views, you need num_images=3 during finetuning. However, single-image training is supported — DexBotic will automatically pad two empty images to meet the 3-image input requirement.