We run code in Google Colab
- Python 3.10.12
- PyTorch 2.0.1
- NVIDIA Tesla T4
Here we have use Durian Leaf Disease dataset. The dataset original consists of 420 healthy and disease leaf images divided into 4 class by disease.
Dataset can be found here
The classes uses in dataset are:
- Leaf Spot
- Algal Leaf Spot
- Leaf Blight
- No Disease
durian plant disease with augmentation by rotation and cutout methods in roboflow 90° Rotate: Clockwise, Counter-Clockwise, Upside Down Cutout: 7 boxes with 15% size each
Dataset Split: Train 70 %, Valid 15 %, Test 15 %
Dataset No Augmentation can be download here.
Dataset Augmentation can be download here.
Train VGG16, EfficientNet_b2, ResNet18 Pre-trained models with IMAGENET1K_V1
75 Epoch with Data_No_Augment and 25 Epoch with Data_Augment
Output of training is given below
Epoch 73/74 train Loss: 0.0003 Acc: 1.0000 valid Loss: 0.0464 Acc: 1.0000
Epoch 74/74 train Loss: 0.0004 Acc: 1.0000 valid Loss: 0.0424 Acc: 1.0000
Training complete in 2m 43s Best val Acc: 1.000000
Evaluation With Accuracy, weighted F1-Score
| Model | Augmentation | Accuracy | F1 Score |
|---|---|---|---|
| ResNet18 | No Augmentation | 94.12% | 0.9399 |
| Augmentation | 97.06% | 0.9706 | |
| VGG16 | No Augmentation | 89.71% | 0.8976 |
| Augmentation | 94.12% | 0.9421 | |
| EfficientNet B2 | No Augmentation | 90.20% | 0.9025 |
| Augmentation | 92.16% | 0.9217 |