vit-stanford-cars

This model is a fine-tuned version of google/vit-base-patch16-224-in21k on the Stanford Cars dataset.

It was produced as part of a course assignment to compare:

• Fine-tuned ViT (this model)
• Zero-shot CLIP baseline
• OpenAI vision baseline

Quick-run evaluation (CPU subset run) reported:

• Loss: 5.2723
• Accuracy: 0.0

Model description

The model predicts one of the official Stanford Cars fine-grained classes (196 total classes in the full dataset taxonomy).

Intended uses & limitations

Intended use:

• Educational benchmarking and demonstration of transfer learning on car classification.

Limitations:

• Quick subset run; accuracy is not representative of a full training run.
• May produce unstable predictions on unseen distributions.
• Not suitable for safety-critical or production decisions.

Training and evaluation data

Dataset:

• Source: tanganke/stanford_cars
• Full taxonomy: 196 classes

Preprocessing:

• RGB conversion
• ViT AutoImageProcessor normalization and resizing
• label -> labels integer mapping for supervised fine-tuning

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
5.2685	1.0	32	5.2723	0.0

Reproducibility notes

This checkpoint is intended as a minimum viable trained model for submission completeness. For better performance, retrain on the full Stanford Cars train split with more epochs.

Framework versions

• Transformers 5.5.4
• Pytorch 2.11.0+cpu
• Datasets 4.8.4
• Tokenizers 0.22.2