Smart-Bullet

A Cloud-Assisted Bullet Screen Filter based on Deep Learning. 3 parts were included:

• Data preprocessing
• CNN model train
• Smart Danmu Google Chrome extension

Data Preprocessing

Data crawl

We use the crawler to crawl data from Tensent video. The crawled bullet dataset can be found here.

Run

In the following we list some important arguments indata_preprocessing.py:

• --dataset: path to the input raw dataset.
• --upcount_num: set the positive data lower upcount level.
• --data_size: the number of data select to process.
• --num_negative_select: the num of selected negative data.

cd data_preprocessing
python data_preprocessing.py --dataset dataset/bulletData.csv --upcount_num 100 --num_negative_select 100000

The output will be two files (positive, negative) in data_preprocessing dir.

Model train

Use the processed data to train the cnn model:

cd cnn_train
python train.py

Model will be saved in cnn_train/runs/checkpoints dir.

Google Chrome extension and server build