t2m checkpoint model size error

[yhyu13]

I've finished setup with py39 and torch 1.12.1 and ran

python demo.py --cfg ./configs/config_mld_humanml3d.yaml --cfg_assets ./configs/assets.yaml --example ./demo/example.txt

But there is an error in loading the pre-trained weight

Length-50: a man kicks with something or someone with his left leg.
Length-100: A person is skipping rope.
Length-100: a person walks backward slowly.
Pointer Pointing at 0
`text_config_dict` is provided which will be used to initialize `CLIPTextConfig`. The value `text_config["id2label"]` will be overriden.
/home/hangyu5/anaconda3/envs/t2m_mld/lib/python3.9/site-packages/torch/cuda/__init__.py:146: UserWarning: 
NVIDIA GeForce RTX 3090 with CUDA capability sm_86 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60 sm_70.
If you want to use the NVIDIA GeForce RTX 3090 GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

  warnings.warn(incompatible_device_warn.format(device_name, capability, " ".join(arch_list), device_name))
Traceback (most recent call last):
  File "/home/hangyu5/Documents/Git-repoMy/AIResearchVault/repo/AIGC/Text2Motion/motion-latent-diffusion/demo.py", line 337, in <module>
    main()
  File "/home/hangyu5/Documents/Git-repoMy/AIResearchVault/repo/AIGC/Text2Motion/motion-latent-diffusion/demo.py", line 105, in main
    model = get_model(cfg, dataset)
  File "/home/hangyu5/Documents/Git-repoMy/AIResearchVault/repo/AIGC/Text2Motion/motion-latent-diffusion/mld/models/get_model.py", line 7, in get_model
    return get_module(cfg, datamodule)
  File "/home/hangyu5/Documents/Git-repoMy/AIResearchVault/repo/AIGC/Text2Motion/motion-latent-diffusion/mld/models/get_model.py", line 17, in get_module
    return Model(cfg=cfg, datamodule=datamodule)
  File "/home/hangyu5/Documents/Git-repoMy/AIResearchVault/repo/AIGC/Text2Motion/motion-latent-diffusion/mld/models/modeltype/mld.py", line 86, in __init__
    self._get_t2m_evaluator(cfg)
  File "/home/hangyu5/Documents/Git-repoMy/AIResearchVault/repo/AIGC/Text2Motion/motion-latent-diffusion/mld/models/modeltype/mld.py", line 175, in _get_t2m_evaluator
    self.t2m_moveencoder.load_state_dict(
  File "/home/hangyu5/anaconda3/envs/t2m_mld/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1604, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for MovementConvEncoder:
    size mismatch for main.0.weight: copying a param with shape torch.Size([512, 259, 4]) from checkpoint, the shape in current model is torch.Size([512, 247, 4]).

Is it an known issue?

[billl-jiang]

Hi @yhyu13

From the error log, it seems that the input size of evaluators is not correct which indicates the feature size of the dataset is not correct. https://github.com/ChenFengYe/motion-latent-diffusion/blob/c28a06435077800fde4d76ef93eb2a4016a5120c/mld/models/modeltype/mld.py#L157-L161 Maybe the problem is caused by the incorrect dataset. The feature size of the HumanML3D data format is 263 and the feature size of the KIT-ML data format is 251. From the log it seems that your HumanML3d dataset feature size is 251 so maybe you are loading the KIT-ML data for the HumanML3D config.

1. First you can check whether the path of HumanML3D is correct. https://github.com/ChenFengYe/motion-latent-diffusion/blob/c28a06435077800fde4d76ef93eb2a4016a5120c/configs/assets.yaml#L13-L14
2. Then you can use NumPy to load the npy files in the HumanML3D folder manually and check whether the shape is [motion_length, 263]which 263 is the feature size of HumanML3D data format.

You can refer to HumanML3D for more details about the HumanML3D dataset.

[ehsanwwe]

i have the same error @yhyu13 how can you solve this ? are you remember ?

[FogSue]

Hi @yhyu13

From the error log, it seems that the input size of evaluators is not correct which indicates the feature size of the dataset is not correct.

https://github.com/ChenFengYe/motion-latent-diffusion/blob/c28a06435077800fde4d76ef93eb2a4016a5120c/mld/models/modeltype/mld.py#L157-L161

Maybe the problem is caused by the incorrect dataset. The feature size of the HumanML3D data format is 263 and the feature size of the KIT-ML data format is 251. From the log it seems that your HumanML3d dataset feature size is 251 so maybe you are loading the KIT-ML data for the HumanML3D config.

1. First you can check whether the path of HumanML3D is correct. https://github.com/ChenFengYe/motion-latent-diffusion/blob/c28a06435077800fde4d76ef93eb2a4016a5120c/configs/assets.yaml#L13-L14
2. Then you can use NumPy to load the npy files in the HumanML3D folder manually and check whether the shape is [motion_length, 263]which 263 is the feature size of HumanML3D data format.

You can refer to HumanML3D for more details about the HumanML3D dataset.

Hello, I believe I have correctly prepared the HumanML3D dataset, as I'm able to obtain the expected shape. However, I'm still encountering persistent shape mismatching errors.

[FogSue]

Hi @yhyu13 From the error log, it seems that the input size of evaluators is not correct which indicates the feature size of the dataset is not correct. https://github.com/ChenFengYe/motion-latent-diffusion/blob/c28a06435077800fde4d76ef93eb2a4016a5120c/mld/models/modeltype/mld.py#L157-L161

Maybe the problem is caused by the incorrect dataset. The feature size of the HumanML3D data format is 263 and the feature size of the KIT-ML data format is 251. From the log it seems that your HumanML3d dataset feature size is 251 so maybe you are loading the KIT-ML data for the HumanML3D config.

1. First you can check whether the path of HumanML3D is correct. https://github.com/ChenFengYe/motion-latent-diffusion/blob/c28a06435077800fde4d76ef93eb2a4016a5120c/configs/assets.yaml#L13-L14
2. Then you can use NumPy to load the npy files in the HumanML3D folder manually and check whether the shape is [motion_length, 263]which 263 is the feature size of HumanML3D data format.

You can refer to HumanML3D for more details about the HumanML3D dataset.

Hello, I believe I have correctly prepared the HumanML3D dataset, as I'm able to obtain the expected shape. However, I'm still encountering persistent shape mismatching errors.

This is caused by loading kit ckpt for humanml3d!