Adaptive Hilbert Diffusion Models for Controllable Smoothness in Continuous Function Generation

This repo is the official Pytorch implementation of Adaptive-HDM(Adaptive Hilbert Diffusion Models for Controllable Smoothness in Continuous Function Generation). This code is based on the MDM.

Generated Samples

Environment

This code was tested on Ubuntu 20.04.6 LTS and requires:

• Python 3.7
• conda3 or miniconda3
• CUDA capable GPU (one is enough)

Setup conda env:

conda env create -f environment.yml
conda activate ahdm
python -m spacy download en_core_web_sm
pip install git+https://github.com/openai/CLIP.git

Dataset

Text-to-Motion

Fllow the instructions in HumanML3D, then locate at ./dataset/HumanML3D.

cp -r ../HumanML3D/HumanML3D ./dataset/HumanML3D

Download dependencies for Text-to-Motion:

bash prepare/download_smpl_files.sh
bash prepare/download_glove.sh
bash prepare/download_t2m_evaluators.sh

Pre-Trained model

• Length Prediction Module
• Text-to-motion

Unzip and place them in ./save/.

Training

Text-to-Motion

For Text to motion, we don't compute kernels during training, but use precomputed kernel values. To calculate these kernels, use the command below.

python3 -m train.pre_calK

Subsequently, you can train with below command.

python3 -m train.train_GPmotion --save_dir 'save/train' --corr_noise --dataset humanml --eval_during_training --diffusion_steps 50 --corr_mode R_trs --param_lenK_path HumanML3D_K_param_data196_fps20_dim263_len10.pkl

Generation

Text-to-Motion

python3 -u sample.generate_GP --model_path path/your/model --num_samples 3 --num_repetitions 2 --dataset humanml --param_lenK_path HumanML3D_K_param_data196_fps20_dim263_len10.pkl --text_prompt "A man moves forward." --guidance_param 2.5 --corr_noise --corr_mode R_trs 

Evaluation

Text-to-Motion

python3 -m eval.eval_humanml -model_path path/your/model