blendtorch is a Python framework to seamlessly integrate Blender into PyTorch datasets for deep learning from artificial visual data. We utilize Eevee, a new physically based real-time renderer, to synthesize images and annotations in real-time and thus avoid stalling model training in many cases.
Feature summary
- Data Streaming: Stream distributed Blender renderings directly into PyTorch data pipelines in real-time for supervised learning and domain randomization applications. Supports arbitrary pickle-able objects to be send alongside images/videos. Built-in recording capability to replay data without Blender. Bi-directional communication channels allow Blender simulations to adapt during network training.
More info [examples/datagen], [examples/compositor_normals_depth], [examples/densityopt] - OpenAI Gym Support: Create and run remotely controlled Blender gyms to train reinforcement agents. Blender serves as simulation, visualization, and interactive live manipulation environment.
More info [examples/control]
The figure below visualizes the basic concept of blendtorch used in the context of generating artificial training data for a real-world detection task.
- Read the installation instructions below
- To get started with blendtorch for training data training read [examples/datagen].
- To learn about using blendtorch for creating reinforcement training environments read [examples/control].
The code accompanies our academic work [1],[2] in the field of machine learning from artificial images. Please consider the following publications when citing blendtorch
@inproceedings{robotpose_etfa2019_cheind,
author={Christoph Heindl, Sebastian Zambal, Josef Scharinger},
title={Learning to Predict Robot Keypoints Using Artificially Generated Images},
booktitle={
24th IEEE International Conference on
Emerging Technologies and Factory Automation (ETFA)
},
year={2019}
}
@inproceedings{blendtorch_icpr2020_cheind,
author = {Christoph Heindl, Lukas Brunner, Sebastian Zambal and Josef Scharinger},
title = {BlendTorch: A Real-Time, Adaptive Domain Randomization Library},
booktitle = {
1st Workshop on Industrial Machine Learning
at International Conference on Pattern Recognition (ICPR2020)
},
year = {2020},
}
blendtorch is composed of two distinct sub-packages: bendtorch.btt
(in pkg_pytorch) and blendtorch.btb
(in pkg_blender), providing the PyTorch and Blender views on blendtorch.
This package has been tested with
Other versions might work as well, but have not been tested.
git clone https://github.com/cheind/pytorch-blender.git <DST>
Ensure Blender executable is in your environments lookup PATH
. On Windows this can be accomplished by
set PATH=c:\Program Files\Blender Foundation\Blender 2.91;%PATH%
Open Blender at least once, and complete the initial settings. If this step is missed, some of the tests (especially the tests relating RL) will fail (Blender 2.91).
blender --background --python <DST>/scripts/install_btb.py
installs blendtorch-btb
into the Python environment bundled with Blender.
pip install -e <DST>/pkg_pytorch
installs blendtorch-btt
into the Python environment that you intend to run PyTorch from. While not required, it is advised to install OpenAI gym if you intend to use blendtorch for reinforcement learning
pip install gym
This step is optional. If you plan to run the unit tests
pip install -r requirements_dev.txt
pytest tests/
Run
blender --version
and check if the correct Blender version (>=2.83) is written to console. Next, ensure that blendtorch-btb
installed correctly
blender --background --python-use-system-env --python-expr "import blendtorch.btb as btb; print(btb.__version__)"
which should print blendtorch version number on success. Next, ensure that blendtorch-btt
installed correctly
python -c "import blendtorch.btt as btt; print(btt.__version__)"
which should print blendtorch version number on success.
Please see [examples/datagen] and [examples/control] for an in-depth architectural discussion. Bi-directional communication is explained in [examples/densityopt].
The following tables show the mean runtimes per batch (8) and per image for a simple Cube scene (640x480xRGBA). See benchmarks/benchmark.py for details. The timings include rendering, transfer, decoding and batch collating. Reported timings are for Blender 2.8. Blender 2.9 performs equally well on this scene, but is usually faster for more complex renderings.
Blender Instances | Runtime sec/batch | Runtime sec/image | Arguments |
---|---|---|---|
1 | 0.236 | 0.030 | UI refresh |
2 | 0.14 | 0.018 | UI refresh |
4 | 0.099 | 0.012 | UI refresh |
5 | 0.085 | 0.011 | no UI refresh |
Note: If no image transfer is needed, i.e in reinforcement learning of physical simulations, 2000Hz are easily achieved.
- Despite offscreen rendering is supported in Blender 2.8x it requires a UI frontend and thus cannot run in
--background
mode. If your application does not require offscreen renderings you may enable background usage (see tests/ for examples). - The renderings produced by Blender are by default in linear color space and thus will appear darker than expected when displayed.