AttributeError: 'MDMModel' object has no attribute 'dec_depth'

[ccmCCMfk]

Thansk for your great work, I have two questions,
1.When I conduct post-training with the real robot data of Xinghaitu, the loss is relatively large at the beginning (about 0.5) and drops to around 0.02 after 5 to 20 epochs, possibly due to the addition of the waist degree of freedom. Is this normal?
2.I encountered the following problem when trying to save the depth map. Did I miss any files? Traceback (most recent call last):
[rank0]: File "/mnt/shared-nvme/eai/lingbot-vla/tasks/vla/train_lingbotvla.py", line 803, in
[rank0]: main()
[rank0]: File "/mnt/shared-nvme/eai/lingbot-vla/tasks/vla/train_lingbotvla.py", line 670, in main
[rank0]: log_depth(morgbd_model, depth_preds, depth_targets, steps=global_step, config=args.train.align_params, cls_token=cls_token)
[rank0]: File "/mnt/shared-nvme/eai/lingbot-vla/lingbotvla/models/vla/vision_models/module_utils.py", line 117, in log_depth
[rank0]: output_morgbd_preds = morgbd_model.dec_depth(depth_pred_feats, cls_token, num_tokens=256, resolution_level=3, img_h=224, img_w=224)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/usr/local/miniforge3/envs/lingbot/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 405, in getattr
[rank0]: return getattr(self._orig_mod, name)
[rank0]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]: File "/usr/local/miniforge3/envs/lingbot/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1962, in getattr
[rank0]: raise AttributeError(
[rank0]: AttributeError: 'MDMModel' object has no attribute 'dec_depth'
Epoch 1/69: 1%| | 2/292 [00:33<1:20:11, 16.59s/it]

[LuFan31]

Thank you all for your interest in LingBot-VLA!

1. I think this is normal. I suggest using the MSE flow-matching loss for real-robot post-training (just comment out train.loss_type: L1_fm in your config). It's generally lower than the L1 flow-matching loss provided in the README.

2. The dec_depth function is currently missing in the LingBot-Depth repository (which morgbd_model depends on), so depth reconstruction visualization is not yet supported. I have temporarily commented out the relevant calls. We will restore this functionality soon to allow for checking depth distillation effects.

[funguspooh]

How can I collect data from a real robot and convert it into the required format?
I have an UMI LERobot dataset, but I can't seem to use it.

KeyError("Column action not in the dataset. Current columns in the dataset:
['robot0_eef_pos', 'robot0_eef_pos_wrt_start', ..., 'actions', ...]")
KeyError: 'observation.state.arm.position'