Skip to content

Latest commit

 

History

History
48 lines (34 loc) · 3.82 KB

PREPS.md

File metadata and controls

48 lines (34 loc) · 3.82 KB
Raw

Prepare Datasets and UV map Labels

Prerequisites

Windows 10 / Ubuntu 18.04
CUDA 9.0 / 9.1
Python 3.6
PyTorch 1.0.0
opencv-python
tqdm
chumpy (For converting SMPL model to basic numpy arrays)
h5py (For processing Human36m annotations)

Make sure you have an GPU card with at least 12 GB graphic memory. You can decrease batch size too, but the performance is not guaranteed.

(Optional) Linux users are encouraged to install torch-batched-svd for speedup.

Prepare dataset

We hereby datail the necessary steps to reproduce human36m experiments. Most people failed on this step, so I hope this instruction could make sure you don't become one of them.

  1. I tentatively suggest users to first try download human36m dataset from the official website. However manual authorization could take anywhere from 6 days to 6 months, so if you are keen to get things going, you can start by playing with this toy_dataset. Extraction code: 0o95 Alternative link Google Drive

  2. Unpack downloaded zip files into a single folder (which I suggest you name it human36m) and put it under path-to-your-datasets. Open data_utils/data_washing.py, change the root_dir variable in main function near line 190 to your h36m dataset path and run it to perform standard data augmentations, the washed dataset will be stored in path-to-your-datasets/human36m_washed.

Create UV maps

  1. To acquire SMPL's default UV map, follow these steps:
  • Register on SMPL's official website and download their maya fbx animation files, store them under data_utils.
  • Install Blender and open the IDE, import a fbx file inside Blender and then export it as .obj format.
  • Run data_utils/triangulation.py to convert Blender converted obj file into SMPL compatible one (with 13776 triangulated faces).

If you are interested, you can also try with radvani's UV map attached in the issues. Incidentally, you may want to read this issue carefully for more information in case you get stucked midway.

  1. After you acquired the converted .obj file, put it under data_utils and rename it as '{}_template.obj'.format(opt.uv_map) where options can be found in train.py (default is radvani). Run data_utils/uv_map_generator.py to calculate barycentric interpolation weights and other necessary caching informations. This could take 8~10 minutes.

  2. Processing SMPL model with data_utils/preprocess_smpl.py. You can also check the instructions here and here.

  3. Run data_utils/create_UV_maps to create UV labels. This takes 5 hours on my PC, you can monitor the progress through tqdm info. The processed UV maps will be stored at path-to-your-datasets/human36m_UV_map_{} where {} again is opt.uv_map.

Training

  1. Go back to the root folder, checkout training options in train.py and modify them in train.sh, specify your gpu id through os.environ['CUDA_VISIBLE_DEVICES'], run bash train.sh, and Bob's your uncle:)

The training script is designed to monitor losses in realtime through tqdm progress bar. However, if you are working on a remote server or cluster via ssh connections, it's better to submit your job with bash nohup_train.sh. This script replaced tqdm monitoring with a log.txt file under your checkpoint folder where loss info is printed every opt.save_result_freq batches.

Testing

We are still working on the training part. The testing scripts will be released later.