Pokemon-classification-1stGen

Pokemon-classification-1stGen

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the Dusduo/1stGen-Pokemon-Images dataset. It has been trained to discriminate between the pokemons from the 1st Generation. It achieves the following results on the evaluation set:

• Loss: 0.4182
• F1: 0.9272

A demonstration of the model application is hosted on Spaces. Feel free to check it out!

Model description

Transformer-based vision model for pokemon image classification.

Intended uses & limitations

This model is intended to classify between pokemons from the 1st Generation. Therefore, when provided with images of pokemon from posterior generation, the model outputs won't be usable as such. Moreover, the model was not designed to handle non pokemon images as well as images presenting several entities. However, an additional layer can help mitigate the risk of wrongly classifying non pokemon images by analyzing the spread of the output (the confusion of the model), such a layer can be found in my implementation, available here.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6.56462271373806e-05
• train_batch_size: 4
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 7

Training results

Framework versions

• Transformers 4.35.2
• Pytorch 2.2.0.dev20231126+cu118
• Datasets 2.15.0
• Tokenizers 0.15.0