Hi Sun,
Great work. Currently, I am trying to repeat the evaluation script by running evaluation/run.py
I tried either

python evaluation/run.py --ground_truth --save_path tmp/result.xlsx --device 0

or

python3 evaluation/run.py --lgtm_ckpt checkpoints/lgtm/checkpoints/epoch=196-val_loss=0.3769.ckpt --save_path tmp/result.xlsx --device 0

Both of them give me the error "Intel MKL ERROR: Parameter 3 was incorrect on entry to DGEBAL.", I notice that FID.compute() becomes nan after running several batch of the test data. Could you tell me how can I smoothly running the evaluation script and repeat your result?

p.s. I am able to get correct result using the playground demo on test data.

Edit:
fid and diversity is super large when executing

python3 evaluation/run.py --lgtm_ckpt checkpoints/lgtm/checkpoints/epoch=196-val_loss=0.3769.ckpt --save_path tmp/result.xlsx --device 0

fid {'fid': 92.77582277963393}
retrieval_precision {'R_precision_1': 0.850781261920929, 'R_precision_2': 0.93359375, 'R_precision_3': 0.96875, 'Matching_Score': 33.174310302734376}
similarities right_leg {'similarity': 0.6510486602783203}
diversity {'diversity': 55.91614532470703}

Hi! Excelent work! I wonder where I can find the requirement.txt?

Hi! I'm confused about why is the performance of MDM, MLD, and MotionDiffuse in Table 2 different from that in the original paper. Could you tell me which paper the performance in Table 2 comes from?

When I run bash prepare_data_models.sh, I get the error.
`Get h3d features from Guo et al.
The processed motions will be stored in this folder:
datasets/motions/guoh3dfeats
0%| | 0/15136 [00:00<?, ?it/s]Error executing job with overrides: []
Traceback (most recent call last):
File "/data/newhome/weizixiong/Projects/LGTM/third_packages/TMR/prepare/compute_guoh3dfeats.py", line 79, in compute_guoh3dfeats
joints_m = swap_left_right(joints)
^^^^^^^^^^^^^^^^^^^^^^^
File "/data/newhome/weizixiong/Projects/LGTM/third_packages/TMR/prepare/compute_guoh3dfeats.py", line 21, in swap_left_right
assert len(data.shape) == 3 and data.shape[-1] == 3
^^^^^^^^^^^^^^^^^^^
AssertionError

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.
30%|█████████████████████████████████████████▋ | 4534/15136 [00:00<00:00, 49106.10it/s]`

Hi! Excelent work! I wonder where I can find the third_packages/TMR/prepare/body_part_annotations_augmentation.py?

from lgtm.dataset.HumanML3D import BodyPart_HumanML3D

This tool is for decomposition of full body motion description. You need to get an API key from OpenAI

tool = BodyPartAnnotationTool(
api_key="",
base_url="https://api.openai.com/v1",
)

Load LGTM

model = LGTM.load_from_checkpoint(Path("./checkpoints/lgtm/checkpoints/epoch=196-val_loss=0.3769.ckpt"))
model.freeze()

This part-level dataset is for de-normalized data. But you can use it for generation

dataset = BodyPart_HumanML3D(
HumanML3D(Path("third_packages/HumanML3D"), Path("data/glove"), "all"),
Path("./third_packages/TMR/datasets/annotations/humanml3d/annotations.json"),
)

where is the requirements.txt?