Data Extension and Class Addition for VSR
Code for paper: Harnessing GANs for Addition of New Classes in VSR
Link to paper: https://drive.google.com/open?id=13xofvtwqOAXWgjOYu8EFaeMQBTKiB0T7
Link to the supplementary folder (Appendix pdf, paper pdf ,and example reconstructed video with embedded audio) : https://drive.google.com/open?id=1nMgRuyM9juztd_fOzTH62X1v1eTh4R5u
Folder containing all generated examples: https://drive.google.com/drive/folders/13oPBUgOG3itRcztUNI8ItyE71Zvhqe-a?usp=sharing
python 3.6
torch 0.4.1
CUDA 9.0
torchaudio
For running the A2V model, run the following command
CUDA_VISIBLE_DEVICES=1 python driver.py
For training and video generation change following lines:
-
Epochs-> line 78 | num_epochs (variable name)
-
View-> line 257 | videos, audios = get_data("v1", 1) --> change v1 to v2,v3,v4,v5 to train for different views, and value 1 can go upto the number of speakers you want to train for (30 or 60) in our case .
-
comment out line 277 -> For forward pass on the model for video generation
-
uncomment line 420 -> generate_video('s9_v3_u28.npy') --> namme of the folder containing reference image also the name of the folder containing reference audio file to generate video
-
uncomment line 421 -> prep_video_gen_data('v1', 30, 60) --> change v1 to v2,v3,v4,v5 to generate videos for different views and set range of speakers anywhere between 0-60 in our case to generate videos specifically for those speakers for that particular view
For running the Visual Speech Recognition model, run the following command to train the model (for one view model):
CUDA_VISIBLE_DEVICES=1 python 1stream.py --data_pickle_path "Path to the picke file(without quotes)" \
--num_epoch 40 \
--num_classes 10 \
--one_stream_view 1
Please check the python file for details about various parameters.
For generating the pickle file look at pre_process_oulu.py
@misc{kumar2019harnessing, title={Harnessing GANs for Zero-shot Learning of New Classes in Visual Speech Recognition}, author={Yaman Kumar and Dhruva Sahrawat and Shubham Maheshwari and Debanjan Mahata and Amanda Stent and Yifang Yin and Rajiv Ratn Shah and Roger Zimmermann}, year={2019}, eprint={1901.10139}, archivePrefix={arXiv}, primaryClass={cs.LG} }
You can contact us at [email protected], [email protected], [email protected]