gu-ma / featurescoop

Experimental repo

Experimental repo

You probably don't want to fork it 😜

Steps

1. Scrap videos using youtube-dl or instaloader. You can use tools/scrapping.py to automate this step.
2. videos_add.py: Copy and convert videos into a folder in (see config.py for list of options) Then extract for each video:
   • Scenes - using ffmpeg / ffprobe
   • Frames - as images
   • Features for each frame - using the classifier
   • Predictions for each frame (labels) - using the classifier
3. videos_process.py: Builds (or load) single files concatenating all:
   • Scenes
   • Frames
   • Features (Reducing dimensions using PCA)
   • Predictions It also create a file holding the graph connecting all features.
4. videos_test.py: Runs different test such as finding closest images and finding path between 2 frames
5. startend_images_encode.py is used to encode some images to start and end the animation. startend_images_getpath.py is used to test the start and end image (this is a quick last minute fix that needs to be removed)
6. build_annoy_index.py encode pregenerated lines of poetry using spacy and builds an index of those lines using annoy
7. api.py starts a local api which can be called to getpath between 2 images or GetSimilarVideos

Requirements

P 3.6. Need to add version for all packages

Python

• keras
• tensorflow
• igraph
• numpy
• scipy
• pickle
• matplotlib
• tqdm
• sklearn
• rasterfairy
• moviepy
• PIL ??
• imageio ??
• instaloader
• annoy
• spacy + en-core-web-sm / en_vectors_web_lg
• termcolor
• colorama

Nonpython

• youtube-dl
• ffmpeg / ffprobe

Vid description + text comparison

TODO

import yt video etc... db?