Skip to content

Latest commit

 

History

History
56 lines (44 loc) · 2.9 KB

youtube8m.md

File metadata and controls

56 lines (44 loc) · 2.9 KB
Raw

English | 简体中文

YouTube-8M Data Preparation

Introduction

YouTube-8M is a large-scale video classification data set, containing more than 8 million video URLs. The tag system covers more than 3800 knowledge graph entities. One video corresponds to multiple tags (3-4 on average) and is labeled by machine.

**The length of each video is between 120s and 500s Due to the large amount of video data, the image classification model was used to extract frame-level features in advance, and PCA was used to reduce the dimensionality of the features to obtain multi-frame 1024-dimensional features. Similarly, the audio model was used to obtain multi-frame 128-dimensional features. Audio characteristics. **

The dataset used here is the updated YouTube-8M data set in 2018 (May 2018 version (current): 6.1M videos, 3862 classes, 3.0 labels/video, 2.6B audio-visual features).

Download

  1. Create a new directory for storing features (take the PaddleVideo directory as an example)

    cd data/yt8m
mkdir frame
cd frame

  2. Download the training and validation set to the frame folder

    curl data.yt8m.org/download.py | partition=2/frame/train mirror=asia python
curl data.yt8m.org/download.py | partition=2/frame/validate mirror=asia python

    The download process is shown in the figure image

    After the data download is complete, you will get 3844 training data files and 3844 verification data files (TFRecord format)

Conversion

  1. Install tensorflow to read tfrecord data 
    python3.7 -m pip install tensorflow-gpu==1.14.0
  2. Convert the downloaded TFRecord file into a pickle file for PaddlePaddle to use 
    cd .. # From the frame directory back to the yt8m directory
python3.7 tf2pkl.py ./frame ./pkl_frame/ # Convert train*.tfrecord and validate*.tfrecord in the frame folder to pkl format
  3. Generate a single pkl file path set, and split pkl into multiple small pkl files based on this file, and generate the final split pkl file path required 
    ls pkl_frame/train*.pkl> train.list # Write the path of train*.pkl to train.list
ls pkl_frame/validate*.pkl> val.list # Write the path of validate*.pkl into val.list

python3.7 split_yt8m.py train.list # Split each train*.pkl into multiple train*_split*.pkl
python3.7 split_yt8m.py val.list # Split each validate*.pkl into multiple validate*_split*.pkl

ls pkl_frame/train*_split*.pkl> train.list # Rewrite the path of train*_split*.pkl into train.list
ls pkl_frame/validate*_split*.pkl> val.list # Rewrite the path of validate*_split*.pkl into val.list