yt8m-feature-extractor

Extract features from video file as the format in Youtube-8M.

Note Google now share their own feature extractor with pre-trained model. That'd better to use theirs.

Description

• scripts/download.py: download videos from YouTube corresponding to the TFRecord file.
• scripts/decode.py: decode frames from video and save them to data folder.
• scripts/train_pca.py: load extracted inception_v3 features and fit PCA with them.
• scripts/test.py: test single video file and generate TFRecord.
• scripts/extract.py: extract inception_v3 features from decoded image folders.
• scripts/pack.py: transform and pack your downloaded videos into Youtube-8M-dataset-like TFRecord file.
• scripts/pipeline.py: download videos and extract inception_v3 features.
• scripts/label_converter.py: convert label numbers into names.
• scripts/checker.py: check if downloaded TFRecord is valid and complete.
• demo.sh: all-in-one shell script for testing single video file and get its tags.

Workflow

1. Run virtualenv -p python3 yt8m-env && source yt8m-env/bin/activate for virtual Python environment.
2. Run pip3 install -r requirements.txt for required Python packages.

• You may install pytorch by the instruction of its official website.

3. Modify misc/config.py for custom configuration.
4. Run python3 scripts/pipeline.py to download videos and extract inception_v3 features.
5. Once you've downloaded enough videos, you can run python3 scripts/train_pca.py to fit pca.
6. After fitting PCA, run python3 scripts/pack.py to transform and pack your downloaded videos into Youtube-8M-dataset-like TFRecord file.
7. Just run your training scripts for Youtube-8M and enjoy!