Arko007/saffron-verify-pretrained

SaffronVerify — ConvNeXt-Base (Pretrained)

A high-accuracy saffron quality classification model trained on the Arko007/saffron-verify dataset. The model classifies saffron images into three grades: Mogra, Lacha, and Adulterated.

Model Performance

Best checkpoint saved at Epoch 13 with early stopping triggered at Epoch 20.

Metric	Value
Macro F1	0.9888
Accuracy	98.96%
Val Loss	0.3562

Per-Class Results (Epoch 13 — Best Checkpoint)

Class	Precision	Recall	F1-Score	Support
mogra	0.98	0.98	0.98	56
lacha	0.98	0.98	0.98	64
adulterated	1.00	1.00	1.00	72
macro avg	0.99	0.99	0.99	192

Training Details

Parameter	Value
Base Model	convnext_base (ImageNet-21k pretrained via timm)
Image Size	512 × 512
Effective Batch Size	96 (16 per GPU × 2 GPUs × 3 grad accum)
Optimizer	AdamW (β₁=0.9, β₂=0.999)
Learning Rate	5e-6 (backbone) / 2.5e-5 (head)
Scheduler	Warmup (5 epochs) + Cosine Annealing
Regularization	Drop rate 0.3, Drop path 0.2, Label smoothing 0.1
Augmentation	Mixup (α=0.4) + CutMix (α=1.0)
AMP	float16
Hardware	2× NVIDIA Tesla T4 (DDP)
Best Epoch	13 / 50
Early Stopping	Patience 7 — triggered at Epoch 20

Training Progression

Epoch	Val Loss	Accuracy	Macro F1
1	1.0631	51.04%	0.5088
2	0.9541	71.88%	0.7154
3	0.8096	81.77%	0.8118
5	0.5122	90.62%	0.9033
7	0.4153	95.31%	0.9506
10	0.3676	97.92%	0.9777
13	0.3562	98.96%	0.9888
20	—	—	— (early stop)

Offline Data Augmentation

Training data was augmented offline from 167 real images to 3840 balanced training images (1280 per class) using a heavy Albumentations pipeline including random crops, flips, rotations, colour jitter, blur, noise, elastic transforms, perspective distortion, CoarseDropout, and CLAHE. Validation set was augmented from 41 real images to 192 balanced images (64 per class).

Usage

import torch
import timm
import torch.nn as nn
from torchvision import transforms
from PIL import Image

class SaffronVerifyModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.backbone = timm.create_model(
            "convnext_base", pretrained=False,
            num_classes=0, drop_rate=0.3, drop_path_rate=0.2
        )
        feat_dim = self.backbone.num_features
        self.head = nn.Sequential(
            nn.LayerNorm(feat_dim),
            nn.Dropout(p=0.3),
            nn.Linear(feat_dim, 512),
            nn.GELU(),
            nn.Dropout(p=0.15),
            nn.Linear(512, 3),
        )

    def forward(self, x):
        return self.head(self.backbone(x))

CLASSES = ["mogra", "lacha", "adulterated"]

# Load model
model = SaffronVerifyModel()
ckpt  = torch.load("best_model.pth", map_location="cpu")
model.load_state_dict(ckpt["model_state"])
model.eval()

# Preprocess
transform = transforms.Compose([
    transforms.Resize(512),
    transforms.CenterCrop(512),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406],
                         [0.229, 0.224, 0.225]),
])

img    = Image.open("saffron.jpg").convert("RGB")
tensor = transform(img).unsqueeze(0)

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

print(f"Predicted class: {CLASSES[pred]}")

Dataset

• Source: Arko007/saffron-verify
• Raw train: 64 mogra + 64 lacha + 39 adulterated = 167 images
• Raw val: 16 mogra + 16 lacha + 9 adulterated = 41 images
• Augmented train: 3840 (balanced, 1280/class)
• Augmented val: 192 (balanced, 64/class)

License

Apache 2.0