Context length too long called back by CLIP while training my own OmniControl

[Lucky1192405754]

Thanks for your amazing work！But I got a problem here...

When I try to train my own OmniControl model via

python -m train.train_mdm --save_dir save/my_omnicontrol --dataset humanml --num_steps 400000 --batch_size 64 --resume_checkpoint ./save/model000475000.pt --lr 1e-5

it always stop at

RuntimeError: Input a person is standing up facing forward, they rest their hands down to their sides in a relaxed motion. they then clamp their hands for a second, rest their arms again. they begin to walk forward, start off with their left leg.they are walking in a diagonally left way. after they take about 6 steps, then they sit down in mid air (like a relaxed squat) facing diagonally right. they place their left hand on their right side of their chest. is too long for context length 76

which traceback from

File "/data/miniconda3/envs/omnicontrol/lib/python3.7/site-packages/clip/clip.py", line 242, in tokenize
raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}")

I think it means the some of the texts from humanml datasets is too long for the CLIP limitation, which is defined as

context_length = default_context_length - 1in model/cmdm.py L149.

I've also tried to change the value of default_context_length, but it didn't work if more than 77. How can I solve this problem?

Sincerely thanks for your attention !