Swin-Base Open-Source Image Classification Model - Accurately Identify Emoji, Achieve 84.78% Accuracy After 20 Rounds of Fine-Tuning

Swin Base Patch4 Window7 224 20epochs Finetuned Memes

Developed by jayanta

An image classification model based on the Swin Transformer architecture, fine-tuned for 20 epochs on the memes dataset with a validation accuracy of 84.78%

Image Classification

Transformers

Open Source License:Apache-2.0 #Image Classification #High Accuracy #Swin Transformer

Downloads 13

Release Time : 9/17/2022

Model Overview

This is a vision model based on Swin Transformer, specifically optimized for the task of classifying internet memes. The model excels at image classification tasks and is particularly well-suited for processing visual content from social media.

Model Features

High-Performance Image Classification

Achieves 84.78% accuracy and 85.04% F1 score on the memes dataset

Swin Transformer Architecture

Utilizes advanced hierarchical window attention mechanisms for efficient visual information processing

Lightweight Fine-tuning

Requires only 20 training epochs to achieve excellent performance with high training efficiency

Model Capabilities

Image Classification

Visual Content Understanding

Meme Recognition

Use Cases

Social Media Analysis

Meme Classification

Automatically identifies and classifies popular internet meme images

84.78% accuracy

Content Moderation

Inappropriate Content Detection

Identifies visual materials that may contain inappropriate content

🚀 swin-base-patch4-window7-224-20epochs-finetuned-memes

This model is a fine - tuned version of microsoft/swin-tiny-patch4-window7-224 on the imagefolder dataset. It offers significant value in image classification tasks by leveraging the pre - trained knowledge and adapting it to the specific imagefolder dataset. It achieves the following results on the evaluation set:

Loss: 0.7090
Accuracy: 0.8478

📚 Documentation

Model Details

Property	Details
Model Type	Fine - tuned version of microsoft/swin-tiny-patch4-window7-224
Training Data	imagefolder dataset
Metrics	On train split of imagefolder: Accuracy = 0.847758887171561 On custom test split: F1 = 0.8504084378729573, Precision = 0.8519647060733512, Recall = 0.8523956723338485

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.00012
train_batch_size: 32
eval_batch_size: 32
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e - 08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_epochs: 20

Training Results

Training Loss	Epoch	Step	Validation Loss	Accuracy
1.0238	0.99	40	0.9636	0.6445
0.777	1.99	80	0.6591	0.7666
0.4763	2.99	120	0.5381	0.8130
0.3215	3.99	160	0.5244	0.8253
0.2179	4.99	200	0.5123	0.8238
0.1868	5.99	240	0.5052	0.8308
0.154	6.99	280	0.5444	0.8338
0.1166	7.99	320	0.6318	0.8238
0.1099	8.99	360	0.5656	0.8338
0.0925	9.99	400	0.6057	0.8338
0.0779	10.99	440	0.5942	0.8393
0.0629	11.99	480	0.6112	0.8400
0.0742	12.99	520	0.6588	0.8331
0.0752	13.99	560	0.6143	0.8408
0.0577	14.99	600	0.6450	0.8516
0.0589	15.99	640	0.6787	0.8400
0.0555	16.99	680	0.6641	0.8454
0.052	17.99	720	0.7213	0.8524
0.0589	18.99	760	0.6917	0.8470
0.0506	19.99	800	0.7090	0.8478

Framework Versions

Transformers 4.22.1
Pytorch 1.12.1+cu113
Datasets 2.4.0
Tokenizers 0.12.1

📄 License

This model is licensed under the Apache - 2.0 license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご