- Written by Jim Bohnslav, except where as noted
- [email protected]
DeepEthogram is an open-source package for automatically classifying each frame of a video into a set of pre-defined behaviors. Designed for neuroscience research, it could be used in any scenario where you need to detect actions from each frame of a video.
Example use cases:
- Measuring itching or scratching behaviors to assess the differences between wild-type and mutant animals
- Measuring the amount of time animals spend courting, and comparing between experimental conditions
- Counting licks from video for appetite measurement
- Measuring reach onset times for alignment with neural activity
DeepEthogram uses state-of-the-art algorithms for temporal action detection. We build on the following previous machine learning research into action detection:
- Hidden Two-Stream Convolutional Networks for Action Recognition
- Temporal Gaussian Mixture Layer for Videos
For full installation instructions, see this readme file.
In brief:
- Install PyTorch
pip install deepethogram
NEW! All datasets collected and annotated by the DeepEthogram authors are now available from this DropBox link: https://www.dropbox.com/sh/3lilfob0sz21och/AABv8o8KhhRQhYCMNu0ilR8wa?dl=0
If you have issues downloading the data, please raise an issue on Github.
I've written a Colab notebook that shows how to upload your data and train models. You can also use this if you don't have access to a decent GPU.
To use it, please click this link to the Colab notebook.
Then, click copy to Drive
at the top. You won't be able to save your changes to the notebook as-is.
DeepEthogram is now in Beta, version 0.1! There are major changes to the codebase and to model training and inference. Model performance, measured by F1, accuracy, etc. should be higher in version 0.1. Model training times and inference times should be dramatically reduced. For full information, see the DeepEthogram Beta docs
Rather than start from scratch, we will start with model weights pretrained on the Kinetics700 dataset. Go to
To download the pretrained weights, please use this Google Drive link.
Unzip the files in your project/models
directory. Make sure that you don't add an extra directory when unzipping! The path should be
your_project/models/pretrained_models/{models 1:6}
, not your_project/models/pretrained_models/pretrained_models/{models1:6}
.
Copyright (c) 2020 - President and Fellows of Harvard College. All rights reserved.
This software is free for academic use. For commercial use, please contact the Harvard Office of Technology Development ([email protected]) with cc to Dr. Chris Harvey. For details, see license.txt.
The major dependencies for DeepEthogram are as follows:
- pytorch, torchvision: all the neural networks, training, and inference pipelines were written in PyTorch
- pytorch-lightning: for nice model training base classes
- kornia: for GPU-based image augmentations
- pyside2: for the GUI
- opencv: for video and image reading and writing
- opencv_transforms: for fast image augmentation
- scikit-learn, scipy: for binary classification metrics
- matplotlib: plotting metrics and neural network outputs
- pandas: reading and writing CSVs
- h5py: saving inference outputs as HDF5 files
- omegaconf: for smoothly integrating configuration files and command line inputs
- tqdm: for nice progress bars
For GUI usage, we expect that the users will be working on a local workstation with a good NVIDIA graphics card. For training via a cluster, you can use the command line interface.
- CPU: 4 cores or more for parallel data loading
- Hard Drive: SSD at minimum, NVMe drive is better.
- GPU: DeepEthogram speed is directly related to GPU performance. An NVIDIA GPU is absolutely required, as PyTorch uses
CUDA, while AMD does not.
The more VRAM you have, the more data you can fit in one batch, which generally increases performance. a
I'd recommend 6GB VRAM at absolute minimum. 8GB is better, with 10+ GB preferred.
Recommended GPUs:
RTX 3090
,RTX 3080
,Titan RTX
,2080 Ti
,2080 super
,2080
,1080 Ti
,2070 super
,2070
Some older ones might also be fine, like a1080
or even1070 Ti
/1070
.
Test coverage is still low, but in the future we will be expanding our unit tests. For now, email me to get a copy
of testing_deepethogram_archive.zip
. Make a directory in tests called DATA
. Unzip this and move it to the deepethogram/tests/DATA
directory, so that the path is deepethogram/tests/DATA/testing_deepethogram_archive/{DATA,models,project_config.yaml}
. Then run pytest tests/
to run.
the zz_commandline
test module will take a few minutes, as it is an end-to-end test that performs model training
and inference. Its name reflects the fact that it should come last in testing.
- 0.1.1.post1/2: batch prediction
- 0.1.1.post0: flow generator metric bug fix
- 0.1.1: bug fixes
- 0.1: deepethogram beta! See above for details.
- 0.0.1.post1: bug fixes and video conversion scripts added
- 0.0.1: initial version