NSFW pytorch implementation

This repo contains code for running Not Suitable for Work (NSFW) classification deep neural network pytorch models.

According to yahoo-open-nsfw fine-tune with resnet or dpn

Data

You can get some data from nsfw_data_scrapper. Then put your datas at data/train and data/test.

Train

Get pretrained resnet model, and put it in models/.

training with pretrained model:

python src/train.py --saved_model 'models/resnet50-19c8e357.pth'

training from scratch:

python src/train.py --saved_model ''

check from tensorboardx:

/bin/tensorboard --logdir runs/Feb18_14-47-30_ubuntu16-NF5280M4/

Test

python src/test_img.py --batch_size 1 --image_size 299 --test_data_dir data/test_save/  --saved_model models/epoch_53.pkl --save_path data/result/

Result

confusion matrix:

          drawings    hentai    neutral    porn    sexy
drawings   1660          216       101        8      15
hentai       65         1845        16       70       4
neutral      24           12      1805      117      42
porn          2           61        22     1901      14
sexy          3           11        63      524    1399

compare with yahoo's implementation using the same test datas

      sfw   nsfw
sfw   5711  2289
nsfw  267   1733

• notes: some datas from nsfw_data_scrapper are not proper classified

Contributors