DeiT-S fine-tuned on Mini-ImageNet

Public checkpoint for inference testing. The weights start from the pretrained DeiT checkpoint deit_small_patch16_224.fb_in1k and are fine-tuned on Mini-ImageNet (100 classes, data revision bd8779f9d33c061ea6e75fdd3bce4e43dd679060).

Validation top-1 at the selected epoch: 0.9620. Validation top-5: 0.9942.

Use with timm

import timm
model = timm.create_model("hf-hub:m-ahmad-butt/deit-small-mini-imagenet", pretrained=True)
model = model.eval()

Input images are 224x224 and use the standard DeiT normalization. Class names are stored in labels.json. There are 100 classes.

Training hyperparameters

• Optimizer: AdamW
• Loss: CrossEntropyLoss
• Learning rate: 2.368863950364079e-05
• Weight decay: 0.08927180304353628
• Batch size: 32
• Max epochs: 10
• Scheduler: cosine_with_1_epoch_warmup
• Early stopping patience: 3
• Seed: 42
• Best epoch: 7