# youtube-8m

**Repository Path**: garspace2/youtube-8m

## Basic Information

- **Project Name**: youtube-8m
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2022-05-05
- **Last Updated**: 2022-05-20

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

This is code of PhoenixLin(ranked 3rd/394) for the 2nd Youtube-8M Video Understanding Challenge

All the experiments are ran on 2 NVIDIA GTX 1080TI GPUs with a batch size of 160 (80 on each).
Local experiments could run at a speed of 400+ example/sec with SSD and finish in less than 10 hours.
The final submission will takes about 2 days to finish if you want to get the best results.
To reproduce the local experiments:
```bash
bash scripts/train_nextvlad_local.sh
```

To reproduce the final submission experiments:
```bash
bash scripts/train_mix_nextvlad_final_submission.sh
```

The parameters of the model can be adjusted by modifying the bash scripts. 

### Don't forget to change the data paths!

I have some modifications inside the train.py, eval.py and utils.py. If you plan to incorporate the solution 
into your own framework, don't forget those either.

The solution is based on the Youtube-8M Tensorflow Starter Code and the code from Willow, who is the winner of fist Youtube8M competition. 
Please refer to YOUTUBE8M_README.md or (https://github.com/google/youtube-8m) and (https://github.com/antoine77340/Youtube-8M-WILLOW)
for more details!

### extract audio feature : VGGish
## vggish weights from tensorflow model, refer to https://github.com/tensorflow/models/tree/master/research/audioset/vggish
refer to https://github.com/harritaylor/torchvggish

### generate video feature and audio feature into tfrecord
python feature_extract/feature_extract_main2.py

### extract text feature : RoBERT
refer to https://github.com/dbiir/UER-py/wiki/Modelzoo or https://huggingface.co/uer/chinese_roberta_L-2_H-128

blog: https://xw.qq.com/cmsid/20220504A08MCK00
blog: https://spaces.ac.cn/archives/8069