2. Visualize_data loading dataset error

[voyagedzero-glitch]

dataset = LeRobotDataset('omy_pnp', root='./demo_data_example') # if youu want to use the example data provided, root = './demo_data_example' instead!

---

TypeError Traceback (most recent call last)
Cell In[1], line 8
5 import torch
7 # dataset = LeRobotDataset('omy_pnp', root='./demo_data_example') # if youu want to use the example data provided, root = './demo_data_example' instead!
----> 8 dataset = LeRobotDataset('omy_pnp', root='./demo_data')

File ~/conda/envs/lerobotmujoco/lib/python3.10/site-packages/lerobot/common/datasets/lerobot_dataset.py:486, in LeRobotDataset.init(self, repo_id, root, episodes, image_transforms, delta_timestamps, tolerance_s, revision, force_cache_sync, download_videos, video_backend)
483 self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes)
485 # Check timestamps
--> 486 timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy()
487 episode_indices = torch.stack(self.hf_dataset["episode_index"]).numpy()
490 ep_data_index_np = {k: t.numpy() for k, t in self.episode_data_index.items()}

TypeError: stack(): argument 'tensors' (position 1) must be tuple of Tensors, not Column

It seems that there are some issues with the format or method of the dataset version. Is there any good solution?

[Xuan-waves]

It appears to be a widespread issue.
Make sure your conda environment is downloaded from the requirement.txt
1.try to update your datasets to 3.6.0,and you will get other error
2.add a monkey-patch to fix the new error:
try:
import datasets.features.features as features
_OLD_GENERATE_FROM_DICT = features.generate_from_dict
def _new_generate_from_dict(obj):
if isinstance(obj, dict) and obj.get("_type") == "List":
obj["_type"] = "Sequence"
return _OLD_GENERATE_FROM_DICT(obj)
features.generate_from_dict = _new_generate_from_dict
except (ImportError, AttributeError):
pass
you should add it in train_model.py before all code
It is better to add more： os.environ["TOKENIZERS_PARALLELISM"] = "false"
if you get touble about hf-connect:os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
3.the solution comes from the Internet
4.should be work now
5.The training effect of self-collected data is terrible :(

[bestgwl]

This issue is likely caused by the version of lerobot. You can try using the 69901b9b branch version, which should resolve the above problem。
best wishes!

[Gaxiaode]

This issue is likely caused by the version of lerobot. You can try using the 69901b9b branch version, which should resolve the above problem。 best wishes!
you mean as follow？
git+https://github.com/huggingface/lerobot.git@69901b9b#egg=lerobot

[sunfemat]

This issue is likely caused by the version of lerobot. You can try using the 69901b9b branch version, which should resolve the above problem。 best wishes!

i have the same issue after using the 69901b9b branch version.

[jeongeun980906]

I think you should downgrade your datasets to 3.xx if your version is 4.xx
I am planning to update the environment to enhance the performance. One thing that you can try is changing states (eef pose) to the current joint space, like in the env for pi_0 and smolvla.

[quadpixels]

This issue is likely caused by the version of lerobot. You can try using the 69901b9b branch version, which should resolve the above problem。 best wishes!

i have the same issue after using the 69901b9b branch version.

For me I also had to download datasets from 3.6.0 to 2.9.1 after downgrading lerobot to 699901b9b:
python -m pip install "datasets==2.19.1"

Also I'm using Windows so I had to set num_workers=0.

[quadpixels]

5.The training effect of self-collected data is terrible :(

I found during my training of the ACT model, the dataset I collected, grip's state changes 0 to 1 and never back to 0. I added 10 frames to the capture after success and captured 3 episodes; with as little as 100 epochs on a CPU, the deployed policy is able to put the mug onto the plate

[zy111x]

dataset = LeRobotDataset('omy_pnp', root='./demo_data_example') # if youu want to use the example data provided, root = './demo_data_example' instead!

TypeError Traceback (most recent call last) Cell In[1], line 8 5 import torch 7 # dataset = LeRobotDataset('omy_pnp', root='./demo_data_example') # if youu want to use the example data provided, root = './demo_data_example' instead! ----> 8 dataset = LeRobotDataset('omy_pnp', root='./demo_data')

File ~/conda/envs/lerobotmujoco/lib/python3.10/site-packages/lerobot/common/datasets/lerobot_dataset.py:486, in LeRobotDataset.init(self, repo_id, root, episodes, image_transforms, delta_timestamps, tolerance_s, revision, force_cache_sync, download_videos, video_backend) 483 self.episode_data_index = get_episode_data_index(self.meta.episodes, self.episodes) 485 # Check timestamps --> 486 timestamps = torch.stack(self.hf_dataset["timestamp"]).numpy() 487 episode_indices = torch.stack(self.hf_dataset["episode_index"]).numpy() 490 ep_data_index_np = {k: t.numpy() for k, t in self.episode_data_index.items()}

TypeError: stack(): argument 'tensors' (position 1) must be tuple of Tensors, not Column

It seems that there are some issues with the format or method of the dataset version. Is there any good solution?

i have the same issue!how do you solve finally？