The repository contains PyTorch implementation of "BioFaceNet: Deep Biophysical Face Image Interpretation". The current status of this repo is work in progress.
Link to Article: https://arxiv.org/abs/1908.10578
Requirements: Python3, PyTorch, NumPy, Pandas, Matplotlib
Pre-defined values: The matrices required for computation are uploaded in the utils folder. Update the "folder_path" in the cell "Python Code for setup.m" to point to the matrices in util folder.
Dataset: CelebA dataset is used for training, validating and testing the pipeline. The dataset can be downloaded from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
The code is in the form of a Jupyter notebook. Please run all the cells sequentially. Assign the path to the input image folder to the variable "data_root".
Note:
- The entire pipeline is not functional. The loss computation and backward propagation need an update.
- The dimensions provided in the comments of the functions are with respect to Matlab implementation. The same can be used for doing a first pass on the implementation of the functions in Python. In general, Matlab uses the dimensions as (Height x Width x Number of channels x Batch Size) and PyTorch refers to (Batch Size x Number of channels x Height x Width). Also, while comparing the Python and Matlab implementation, do note that indices in Python start with 0 and in Matlab, they start with 1.