README.md

metadata

license: mit
pipeline_tag: image-classification
library_name: pytorch
base_model: microsoft/swin-small-patch4-window7-224
metrics:
  - accuracy
  - f1
  - auc
tags:
  - swin-transformer
  - timm
  - image-classification
  - plant-disease
  - tea-leaf
  - rgb-hsv
  - color-aware
datasets:
  - tea-leaf-disease
language:
  - en

Swin Transformer (RGB + HSV) for Tea Leaf Disease Classification 🌱🍃

This repository provides a Swin Transformer Small model fine-tuned for tea leaf disease classification using a color-aware RGB + HSV fusion strategy.
The model achieves strong generalization performance with high accuracy and AUC on the test set.

---

🧠 Model Overview

• Architecture: Swin Transformer Small (swin_small_patch4_window7_224)
• Pretrained: Yes (ImageNet)
• Input: RGB + HSV
• HSV Fusion: Raw HSV channels (no sin/cos encoding)
• Gating: Vector gate (disabled in this run)
• DropPath: 0.2
• EMA: Enabled
• AMP: Enabled
• Framework: PyTorch (timm-style training)

---

📐 Model Complexity

Metric	Value
Parameters	49.47M
GFLOPs	17.16
Weights size	~200 MB

---

📊 Final Test Performance

Evaluation performed using EMA weights from the best checkpoint (epoch 93).

Metric	Score
Top-1 Accuracy	96.01%
Macro-F1	95.51%
Macro-AUC	99.59%

Benchmark details

• Test images: 212
• Total inference time: 2.30s
• Throughput: 92.3 images/sec

---

🚀 Inference Speed

• Post-warmup forward-only
• 92.3 img/s on GPU

---

🗂️ Training Details

• Experiment name: swin_small_hsv_raw
• Device: CUDA
• Epochs: 100
• Best checkpoint: Epoch 93
• Gradient accumulation: 1
• HSV gate warmup: 5 epochs

---

📦 Model Files

• model.safetensors — final EMA weights (recommended)
• Config and training artifacts included in repository

---

🧪 Intended Use

This model is designed for:

• Tea leaf disease classification
• Agricultural decision-support systems
• Research on color-aware vision transformers

⚠️ Not intended as a medical or agronomic diagnostic tool.

---

⚠️ Limitations

• Dataset-specific bias may exist

---

🧑‍💻 How to Use (PyTorch + timm)

import timm
import torch
from PIL import Image
from torchvision import transforms

# Create model
model = timm.create_model(
    "swin_small_patch4_window7_224",
    pretrained=False,
    num_classes=NUM_CLASSES
)

# Load weights
state = torch.load("model.safetensors", map_location="cpu")
model.load_state_dict(state, strict=False)
model.eval()

# Preprocessing
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=(0.485, 0.456, 0.406),
        std=(0.229, 0.224, 0.225)
    )
])

img = Image.open("tea_leaf.jpg").convert("RGB")
x = transform(img).unsqueeze(0)

with torch.no_grad():
    logits = model(x)
pred = logits.argmax(dim=1).item()

print("Predicted class:", pred)