nsfw-image-detector

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the imagefolder dataset. It achieves the following results on the evaluation set:

• Loss: 0.8138
• Accuracy: 0.9316
• Accuracy K: 0.9887

You can access 384 version on:

https://huggingface.co/LukeJacob2023/nsfw-image-detector-384

Model description

Labels: ['drawings', 'hentai', 'neutral', 'porn', 'sexy']

Intended uses & limitations

Training and evaluation data

A custom dataset about 28k images, if you need to improve your domain's accurate, you can contribute the dataset to me.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Accuracy K
0.7836	1.0	720	0.3188	0.9085	0.9891
0.2441	2.0	1440	0.2382	0.9257	0.9936
0.1412	3.0	2160	0.2334	0.9335	0.9932
0.0857	4.0	2880	0.2934	0.9347	0.9934
0.0569	5.0	3600	0.4500	0.9307	0.9927
0.0371	6.0	4320	0.5524	0.9357	0.9910
0.0232	7.0	5040	0.6691	0.9347	0.9913
0.02	8.0	5760	0.7408	0.9335	0.9917
0.0154	9.0	6480	0.8138	0.9316	0.9887

Framework versions

• Transformers 4.36.2
• Pytorch 2.0.0
• Datasets 2.15.0
• Tokenizers 0.15.0