Reconstructing Facial Sketches from Voices using Dictionary Learning

This repository includes the code to reconstruct facial sketches from voices.

Repository Structure

• mlsp_project_code contains the old code from the course project
• voice2face contains the refactored code
• speaker_id_weights.pth is the speaker ID weights file

How to install

Use the requirements.txt file to install the dependencies. Preferably, use conda since PyTorch likes conda.

How to run

The dataloader expects a file containing the filepath of each sample along with labels. For example, take a look at vgg_voxceleb_edge_preserving.txt. Then train the model as follows:

$ python train.py vgg_voxceleb_edge_preserving.txt

How to contribute

1. Make sure you can understand and can run the code inside voice2face
2. Create a new directory in this repo, call it face_id or speaker_id and use the code inside voice2face as a template (makes it easier for everyone to read your code)
3. I used Weights&Biases to visualize the training metrics, feel free to use it, take it out or use anything else you like

Questions

Please reach out to me (Mahmoud) on Slack