llamass 0.0.1

A Light Loader for the AMASS dataset

llamass

A Light Loader for the AMASS dataset to make downloading and training on it easier.

To do:

• Note here about the dataset license
• Instructions on how to download the dataset
• Instructions on how to install the requirements for visualization
• Augmentations pulled from original AMASS repo
• Install nbqa and black, run on existing notebooks
• Example train/test splits by unpacking different datasets to different locations
• Add CC licensed picture of llamas to github preview
• add to pypi

Install

License Agreement

Before using the AMASS dataset I'm expected to sign up to the license agreeement here. This package doesn't require any other code from MPI but visualization of pose data does, see below.

Install with pip

Requirements are handled by pip during the install but in a new environment I would install pytorch first to install it with cuda.

pip install llamass

For Visualization

To do: provide script to install this and all its requirements (for curl into bash on colab would be nice)

• Human Body Prior, licensed under the SMPL-X project
• Body Visualizer, licensed under the SMPL-X project
• MAYBE mesh, does not require a sign up page

For MPI's mesh library, libboost-dev is required:

sudo apt-get install libboost-dev

How to use

Downloading the data

The AMASS website provides links to download the various parts of the AMASS dataset. Each is provided as a .tar.bz2 and I had to download them from the website by hand. Save all of these in a folder somehwere.

Unpacking the data

After installing llamass a console script is provided to unpack the tar.bz2 files downloaded from the AMASS website:

fast_amass_unpack -n 4 <.tar.bz2 directory> <directory to save unpacked data>

This will unpack the data in parallel in 4 jobs and provides a progress bar.

Alternatively, this can be access in the library using the llamass.core.unpack_body_models function:

import llamass.core

llamass.core.unpack_body_models("sample_data/", unpacked_directory, 4)

sample_data/sample.tar.bz2 extracting to /tmp/tmp2mpzo7r2

Using the data

Once the data is unpacked it can be loaded by a PyTorch DataLoader directly using the llamass.core.AMASS Dataset class.

• overlapping: whether the clips of frames taken from each file should be allowed to overlap
• clip_length: how long should clips from each file be?
• transform: a transformation function apply to all fields

import torch
from torch.utils.data import DataLoader

amass = llamass.core.AMASS(
    unpacked_directory, overlapping=False, clip_length=1, transform=torch.tensor
)

amass[0]

{'poses': tensor([[ 8.8278e-01,  9.1215e-01,  1.4123e+00, -7.6931e-01, -1.7432e-01,
           1.4412e-01, -8.8708e-01,  1.2578e-01, -1.1743e-01,  6.2168e-01,
          -3.0015e-02, -5.0349e-02,  1.2079e+00, -5.5397e-02, -2.9298e-01,
           1.3664e+00, -4.8525e-02,  9.4145e-02, -1.2289e-01, -1.7571e-03,
           2.2975e-02, -4.0428e-02,  1.7674e-02,  7.5107e-03,  2.2632e-01,
          -5.2651e-02,  1.2855e-01, -2.5180e-02, -2.4160e-02, -2.0444e-02,
           0.0000e+00,  0.0000e+00,  0.0000e+00,  0.0000e+00,  0.0000e+00,
           0.0000e+00, -2.9401e-02, -1.1034e-01,  1.4105e-02, -5.7415e-03,
          -2.5916e-01, -6.3831e-01, -8.2118e-02,  3.6791e-01,  4.3686e-01,
           3.3646e-01,  4.1084e-02,  5.4173e-02,  3.0905e-01, -2.4670e-01,
          -7.1902e-01,  1.2193e-01,  4.5621e-01,  3.2485e-01,  2.3960e-01,
          -1.6219e+00,  4.4240e-01,  2.7042e-01,  1.7819e+00, -5.1572e-01,
          -1.2912e-02, -1.2738e-01,  3.7089e-01,  1.1892e-01, -7.0561e-02,
           6.9949e-02,  2.9811e-01, -1.5940e-01,  7.2650e-02, -1.2700e-01,
          -4.6106e-02, -1.3061e-01, -4.6657e-02, -1.3755e-02, -1.9340e-01,
          -1.8488e-01,  2.1123e-01, -1.1649e-01, -1.5284e-01,  7.5247e-02,
          -1.0174e+00,  2.7336e-02, -1.3719e-01, -1.9953e-01, -6.0931e-01,
           1.7453e-01, -5.2527e-01,  1.9633e-01, -1.5435e-02, -5.0789e-01,
          -5.9722e-01, -3.0975e-02,  8.8964e-02, -2.4794e-01,  1.6460e-01,
          -5.9515e-01, -1.7807e-01,  4.5450e-02, -4.0581e-01, -1.6256e-01,
          -2.1566e-01, -2.6343e-01,  1.4715e+00,  3.3034e-01, -4.4416e-01,
          -7.1300e-01, -4.1882e-01,  3.7449e-01,  6.2814e-01,  7.9259e-02,
          -4.0158e-01,  1.3215e-01, -7.3160e-02,  7.5620e-01,  2.5732e-01,
           1.1187e-01,  6.6922e-01, -2.0742e-01,  1.5713e-01, -6.6628e-02,
          -2.0170e-01, -2.5842e-02,  8.0069e-01, -1.2071e-01, -1.2434e-01,
           9.0385e-01,  5.5542e-02,  2.0036e-01,  4.2557e-01, -6.8690e-01,
          -8.5928e-02,  3.8910e-01,  8.3299e-01, -2.6452e-01,  1.1903e-01,
          -6.2685e-01, -2.1657e-01, -3.2647e-01, -2.1732e-01,  6.6137e-02,
           5.8971e-01, -3.1604e-01,  3.0754e-02,  8.7229e-01, -1.8293e-01,
           1.7172e-01,  4.3822e-01,  1.1091e+00, -6.0599e-01,  1.7735e-01,
          -6.4371e-01,  1.1003e-01, -3.3680e-01,  3.1729e-01,  5.0325e-01,
           9.6051e-02]]),
 'dmpls': tensor([[ 0.5872, -0.5937,  0.0935, -0.1082,  1.5400, -0.3119, -2.0400,  0.9013]]),
 'trans': tensor([[-0.1988,  0.1371,  0.6797]]),
 'betas': tensor([[ 2.2140,  2.0062,  1.7169, -1.6117,  0.5180,  1.4124, -0.1580, -0.1450,
           0.0671,  1.9010,  0.2068,  0.5701, -0.0117, -0.1653,  0.6465,  0.2017]]),
 'gender': tensor([-1])}

amassloader = DataLoader(amass, batch_size=4, shuffle=True)

for data in amassloader:
    for k in data:
        print(k, data[k].size())
    break

poses torch.Size([4, 1, 156])
dmpls torch.Size([4, 1, 8])
trans torch.Size([4, 1, 3])
betas torch.Size([4, 1, 16])
gender torch.Size([4, 1])

Future Work

Caching the dataset may be easy to implement with joblib's Memory so I'm looking into this.