vision_based_transfomer_architectures

The project plans to analyze various transformer models using vision-based approaches with two distinct methodologies. The first methodology is training from scratch, the second method is fine-tuning the models with the chosen dataset. Chosen dataset is Butterfly & Months Image Classification 100 species from Kaggle.[1]

The dataset contains 12594 trains, 500 tests, and 500 validation images. Each image’s size is 224 x 224. There are 100 different classes in the dataset. Determined models to analyze:

1. ViT(Vision-based Transformer)
2. Swin Transformers
3. Data-efficient Image Transformers(DeIT)

3 different methods are applied to implement models:

• From the Scratch
• Fine-tuning
• Feature Extraction

Results:

Figure 1: Result of 3 different methods with ViT

Figure 2: Result of 3 different methods with Swin Transformer

Figure 3: Result of 2 different methods with Transformer

References

[1] Butterfly Image Classification 100 species, [Online].

Available: https://www.kaggle.com/datasets/gpiosenka/butterfly-images40-species