I fine-tuned a pre-trained EfficientNet-B5 model to classify images into three categories: pizza, steak, and sushi.
Model Details
Architecture:torchvision.models.efficientnet_b5
Weights:EfficientNet_B5_Weights.DEFAULT
Modifications: I froze all the base feature layers.
Training Procedure
I trained the model for 10 epochs using the Adam optimizer.
Batch Size: 32
Learning Rate: 0.001
Loss Function: CrossEntropyLoss
Transforms: I used the automatic transforms provided by the default EfficientNet-B5 weights.
Hardware: Trained using cuda (if available) with a set manual seed of 37 for reproducibility.
Dataset
I used a 20% subset of a pizza, steak, and sushi dataset. The data was split into train and test directories.
Evaluation Results
Accuracy and Loss Curves
Over the 10 epochs, the training and testing loss steadily decreased, with the testing loss ending impressively below 0.20. The testing accuracy consistently outperformed the training accuracy and finished highly stable at roughly 97.5%.
Loss and Accuracy Curves
Confusion Matrix
The model performs exceptionally well across all three classes on the test set, showing strong improvements over previous iterations:
Pizza: 45 correct, 0 misclassified as steak, 1 misclassified as sushi.
Steak: 55 correct, 0 misclassified as pizza, 3 misclassified as sushi.
Sushi: 46 correct, 0 misclassified as pizza, 0 misclassified as steak.
Confusion Matrix
Most Confident Wrong Predictions
I plotted the instances where the model was confident but incorrect. The model struggled specifically with predicting complex textures as sushi, though its highest confidence on these incorrect predictions hovered around only 0.51, indicating it was less confident in its errors compared to earlier models.
Wrong Predictions
How to use
python
1import torch
2import torchvision
34# I loaded the model architecture5weights = torchvision.models.EfficientNet_B5_Weights.DEFAULT
6model = torchvision.models.efficientnet_b5(weights=weights)78# I modified the classifier9model.classifier = torch.nn.Sequential(10 torch.nn.Dropout(p=0.2, inplace=True),11 torch.nn.Linear(in_features=2048, out_features=3, bias=True),12)1314# I loaded the saved weights15model.load_state_dict(torch.load("EfficientNet_B5_20percent.pth", map_location="cpu"))16model.eval()
