agrihybridFM-small-dino-pretrained

AgriHybridFM Small - DINO Pretrained

AgriHybridFM: A hybrid CNN-Transformer foundation model for agricultural vision tasks. Pretrained using DINO self-supervised learning on agricultural images.

Architecture

• Type: Hybrid CNN-Transformer
• CNN Backbone: SE-enhanced depthwise separable convolutions
• ViT Head: Efficient transformer with LPU, LMHSA, IRFFN
• Parameters: ~25M
• Input Size: 224x224

Usage

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import torch

# Download weights
weights_path = hf_hub_download(
    repo_id="aksshaymathewp/agrihybridFM-small-dino-pretrained",
    filename="model.safetensors"
)

# Load state dict
state_dict = load_file(weights_path)

# Create model (you need the AgriViT model code)
from models import create_agrivit
model = create_agrivit("small")
model.load_state_dict(state_dict)
model.eval()

Training Details

• Method: DINO self-supervised learning
• Image Size: 224x224
• Architecture Variant: small

Citation

If you use this model, please cite:

@misc{agrihybridfm2024,
  title={AgriHybridFM: A Hybrid CNN-Transformer Foundation Model for Agricultural Vision},
  author={Aksshay Mathew P},
  year={2024},
}

License

Apache 2.0