reconect-torch-copy

This repository is based on this repository

Installation

• conda env create --name reconet_torch --file environment.yaml
• conda activate reconet_torch
• Create a trainedl_models/ and put there all your trained models. See``model_creator.py` for more info

Example uses:

• To see your styled video stream: python video_real_time.py --mode video_stream

• For being able to style a (muted) local video: python video_real_time.py --mode existing_video --video-source daily

• Same as above, but now with the original audio of the local video: python video_real_time.py --mode existing_video --video-source daily --use-audio True

• In any case, press the key numbers (1, 2, 3, 4 ...) to see a different styled model. Press 0 to see the original video.

These example assumes that you have a video located in videos/daily.mp4. If you use audio, then this will be stored in àudios/daily.wav

Below is the other repository´s README

reconet-torch

This repository contains a PyTorch implementation of the ReCoNet paper. It use a lot of code from safwankdb

Contributors:

• Mohd Safwan
• Kushagra Juneja
• Saksham Khandelwal
• safwankdb

dataset

we use the MPI Sintel Dataset. The dataset contains flow fields, motion boundaries, unmatched regions, and image sequences. The image sequences are rendered with different levels of difficulty.

Train

python train.py --phase train --path mpidataset

video

python video_cv2.py --video-name 'avi/output_style.avi' --mode video_style --save-directory trained_models --model-name model.pth

nodes

You can use optim.Adam(), it results bubble artifacts in the following red boxes. I used optim.Adamax() in my code.