pokemon-transfer-resnet18

Pokemon ResNet18 Transfer Learning Classifier

Model Description

This model is a transfer-learning image classifier based on ResNet18, fine-tuned on a custom Pokemon image dataset with 6 classes:

• charizard
• charmander
• charmeleon
• ditto
• eevee
• ekans

The model was trained for the mandatory exercise "Computer Vision Classification & Model Comparison" and is intended to be compared against:

• an open-source zero-shot model (CLIP)
• a closed-source vision model (OpenAI)

Model Details

• Architecture: ResNet18 (torchvision.models.resnet18)
• Framework: PyTorch
• Input size: 224 x 224 RGB
• Output classes: 6
• Checkpoint format: .pth (state_dict + metadata)

Training Data

Dataset structure:

• data/pokemon/train/
• data/pokemon/test/

Classes:

• charizard, charmander, charmeleon, ditto, eevee, ekans

Preprocessing

Training transforms:

• RandomResizedCrop(224)
• RandomHorizontalFlip()
• ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2)
• Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

Evaluation transforms:

• Resize(256)
• CenterCrop(224)
• Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

Training Procedure

• Loss: CrossEntropyLoss
• Optimizer: Adam
• Learning rate: 1e-4
• Weight decay: 1e-4
• Batch size: 16
• Epochs: 4
• Device: CPU

Evaluation Results

Final performance from project reports:

• Best validation accuracy: 0.83784
• Test accuracy: 0.80

Class-level behavior (high-level):

• Strong performance on ditto, charmander, ekans
• Main confusion observed between charizard, charmeleon, and eevee on some samples

Intended Use

This model is intended for:

• educational experiments in transfer learning
• small-scale Pokemon image classification demos
• model-comparison workflows against CLIP/OpenAI vision systems

It is not intended for production or safety-critical applications.

Limitations

• Small custom dataset
• Limited class coverage (only 6 classes)
• Sensitivity to style/domain shift (sprites vs photos, color variants, edited images)

Ethical Considerations

This is a toy educational classifier trained on non-sensitive image categories. No personal or biometric data is used.

How to Use

Load checkpoint contents:

• state_dict: model weights
• labels: class names
• image_size: expected size
• architecture: model family

For complete inference logic and app integration, see the project source implementation.

Reproducibility Artifacts

• Model checkpoint: custom_resnet18.pth
• Project metrics files:
  • custom_model_metrics.json
  • model_comparison.json

Links

• Hugging Face Space: https://huggingface.co/spaces/kukalend/computer-Vision-classification
• Hugging Face model repo: https://huggingface.co/kukalend/pokemon-transfer-resnet18