Newell et al. originally reported 0.881 validation accuracy on MPII. By using bigger batch size, more training and Adam, we achieved validation accuracies of 0.885 using a 2HG model and 0.901 using an 8HG model. In this implementation, validation accuracies of 0.889 and 0.904 are attained by adding a simple additional bottleneck to decoder network used as attention.
This repository provides everything necessary to train and evaluate a single-person pose estimation model on MPII. If you plan on training your own model from scratch, we highly recommend using multiple GPUs.
Requirements:
- Python 3 (code has been tested on Python 3.6)
- PyTorch (code tested with 1.0)
- CUDA and cuDNN
- Python packages (not exhaustive): opencv-python, tqdm, cffi, h5py, scipy (tested with 1.1.0)
Structure:
data/
: data loading and data augmentation codemodels/
: network architecture definitionstask/
: task-specific functions and training configurationutils/
: image processing code and miscellaneous helper functionstrain.py
: code for model trainingtest.py
: code for model evaluation
Download the full MPII Human Pose dataset, and place the images directory in data/MPII/
To train a network, call:
python train.py -e test_run_001
(-e,--exp
allows you to specify an experiment name)
To continue an experiment where it left off, you can call:
python train.py -c test_run_001
All training hyperparameters are defined in task/pose.py
, and you can modify __config__
to test different options. It is likely you will have to change the batchsize to accommodate the number of GPUs you have available.
Once a model has been trained, you can evaluate it with:
python test.py -c test_run_001
The option "-m n" will automatically stop training after n total iterations (if continuing, would look at total iterations)
An 8HG pretrained model is available here. It should yield validation accuracy of 0.904.
A 2HG pretrained model is available here. It should yield validation accuracy of 0.889.
Models should be formatted as exp/<exp_name>/checkpoint.pt
The train/val split is same as that found in authors' implementation
During training, occasionaly "ConnectionResetError" warning was displayed between epochs, but did not affect training.
PyTorch code extended from here. Implemented for a project under advisors Alejandro Newell and Prof. Jia Deng.