Generative model-based compressive sensing for IMU data

1. Install packages

Dependencies:

• Python version: 3.8
• numpy
• pytorch
• pytorch-lightning
• PyYAML
• scipy
• scikit-learn
• trimesh
• pyrender
• smplx
• matplotlib

To install all packages: pip install -r requirements.txt

2. Download datasets

2.1. IMU dataset

Download DIP-IMU dataset (DIP_IMU_and_Others.zip) from DIP project website.

Unzip the file DIP_IMU_and_Others, modify the parameters file_path in eval_vae.py, and data_path in configs/*.yaml files.

2.2. SMPL dataset

Download the SMPL dataset from SMPL project website.

Select version 1.0.0 for Python 2.7 (female/male. 10 shape PCs).

Unzip the file, modify the parameters bm_fname in eval_vae.py.

3. Reproduce results

3.1. Pre-processing

Before training the models, we need to pre-processing the IMU dataset.

For this, let's change the parameter processing in the class IMUDataset in dataset.py to True at your first training.

The pre-processed datasets are then saved into npz files in directory DIP_IMU_and_Others.

Note that we need to pre-process the dataset only once. From the second training, processing parameter should be changed to False for faster training.

3.2. Training

CPU and GPU training: modify parameter gpus in the configs/*.yaml files. For CPU training, the setting is gpus: []. For single GPU training, the setting is gpus: [0].

Run train_mlp_vae.py to train CS-VAE model.

Run train_smpl_vae.py to train a VAE model for mapping IMU signals (204 features) to SMPL pose (72 features).

Run train_dip.py to train the DIP baseline.

All scenarios can be obtained by changing parameters h_in and eta in yaml files.

After training, the training results should be placed in directories like logs/VanillaVAE, logs/DIPVAE, logs/SMPLVAE, and logs/DecodingTime.

The running versions are automatically named by pytorch-lightning.

3.3. Plot results

We use eval_vae.py to evaluate and plot results.

Make sure that file_path and bm_fname parameters in this file are correct.

Change parameter EVALUATION to True and run the file to obtain all results.

If you only want to plot the results, set the EVALUATION parameter to False and rerun the file.