Skip to content

Latest commit

 

History

History
46 lines (37 loc) · 1.74 KB

README.md

File metadata and controls

46 lines (37 loc) · 1.74 KB
Raw

Real-time video style transfer with TensorStream

This example demonstrates how to use TensorStream with fast-neural-style model. Input frames from an RTMP stream are converted to Tensors and passed to a trained PyTorch model so an augmented video is produced. The whole pipeline works in real-time.

Requirements

Usage

  • Download saved models:
python download_saved_models.py
  • Run the example with default parameters:
python neural_style.py -w 808 -h 456

Note: on Windows you have to add path to ffmpeg executable file to PATH

  • Run the example with specific input/output video, different model weights and model input resolution:

Note: You can pass --help to get list of all available options, their description and default values:

python neural_style.py -m </path/to/saved-model.pth> -i </path/to/input-file-or-stream> -o </path/to/output-file-or-stream> -t <sec-to-record> -h <height> -w <width>
  • You can stream result video to a local address and open it in VLC or another player:
python neural_style.py -i tests/resources/bunny.mp4 -o udp://127.0.0.1:1234 -m ./saved_models/candy.pth -w 808 -h 456

for example, you can use MPlayer to view the results:

mplayer -demuxer +mpegts -framedrop -benchmark udp://127.0.0.1:1234?buffer_size=1000000

Example tested at 24 fps on:

  • RTMP stream - 1280x720, 24 fps
  • Model input resolution - 808x456
  • CPU - i7 7700k
  • GPU - Nvidia GTX 1080ti