TransferLearning

Flower Classification with Transfer Learning (MobileNetV2)

This project demonstrates how to apply Transfer Learning using a pretrained MobileNetV2 model in PyTorch to perform multi-class image classification on the Flowers102 dataset.

The goal is to fine-tune a lightweight CNN pretrained on ImageNet to recognize 102 different flower species with minimal training time and computational cost.

Dataset

Oxford Flowers102 (from torchvision.datasets)

• 102 flower classes
• ~2,040 training images
• ~1,020 validation images
• RGB images with varying resolutions

Loaded via:

train_dataset = datasets.Flowers102(root='./data', split="train", transform=transform_train, download=True)
test_dataset  = datasets.Flowers102(root='./data', split="val",   transform=transform_test,  download=True)

Preprocessing & Data Augmentation

Two different pipelines are used:

Training Transformations

transform_train = transforms.Compose([
    transforms.Resize((224,224)),
    transforms.RandomHorizontalFlip(),
    transforms.RandomRotation(10),
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
    transforms.ToTensor(),
    transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5))
])

Validation Transformations

transform_test = transforms.Compose([
    transforms.Resize((224,224)),
    transforms.ToTensor(),
    transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5))
])

Why?

• Data augmentation improves generalization.
• Normalization centers pixel values.
• Resizing ensures compatibility with MobileNet input.

Data Loaders

train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)
test_loader  = DataLoader(test_dataset,  batch_size=32, shuffle=False)

---

Model Architecture

A pretrained MobileNetV2 model is used as the backbone.

Only the final classification layer is replaced.

Original Classifier

Linear(1280 → 1000)

Modified Classifier

Linear(1280 → 102)

Implemented as:

model = models.mobilenet_v2(pretrained=True)
num_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(num_features, 102)

---

Transfer Learning Strategy

• All pretrained convolutional layers are reused.
• Only the final classifier layer is trained.
• This drastically reduces training time.

---

Training Setup

Parameter	Value
Backbone	MobileNetV2 (ImageNet)
Classes	102
Batch size	32
Optimizer	Adam
Learning rate	0.001
Loss function	CrossEntropyLoss
LR Scheduler	StepLR (γ=0.1 every 5)
Epochs	3
Device	CPU / CUDA

---

Training Loop

At each epoch:

• Forward pass
• Compute cross-entropy loss
• Backpropagation
• Update classifier weights
• Step learning rate scheduler

loss = criterion(outputs, labels)
loss.backward()
optimizer.step()

---

Model Saving

torch.save(model.state_dict(), "mobilenet_flowers102.pth")

---

Evaluation

The model is evaluated on the validation set using:

• Confusion Matrix
• Classification Report
• Overall Accuracy

---

Accuracy

Accuracy: 68%
Macro Avg F1-score: 0.66
Weighted Avg F1-score: 0.66

---

Confusion Matrix

Visualized using Seaborn:

cm = confusion_matrix(all_labels, all_preds)
sns.heatmap(cm, cmap="Blues")

This shows how well the model distinguishes between 102 flower categories.

---

Classification Report

Generated via:

print(classification_report(all_labels, all_preds))

Metrics include:

• Precision
• Recall
• F1-score
• Support for each class

Some classes are learned well, while others suffer due to:

• Limited samples
• High visual similarity between flowers

---

Key Concepts Demonstrated

• Transfer Learning
• Fine-tuning pretrained CNNs
• MobileNetV2
• Data augmentation
• Multi-class classification
• Learning rate scheduling
• Confusion matrix analysis
• PyTorch model customization

---

Limitations

• Only 3 training epochs
• No freezing/unfreezing experiments
• No test-time augmentation
• No hyperparameter tuning
• No class imbalance handling
• No top-k accuracy metrics

---

Possible Improvements

• Train for more epochs (10–30)
• Freeze backbone first, then unfreeze
• Use stronger models (ResNet50, EfficientNet)
• Add validation accuracy tracking
• Use early stopping
• Apply MixUp or CutMix
• Compute Top-5 accuracy
• Perform Grad-CAM visualization

How to Run

Install dependencies:

pip install torch torchvision matplotlib seaborn scikit-learn tqdm

Run the notebook:

TransferLearning.ipynb

---

Conclusion

This project shows how powerful transfer learning can be:

With only a few lines of code and minimal training:

• A model pretrained on ImageNet can be adapted
• To solve a 102-class fine-grained vision problem
• Achieving reasonable performance in minutes

This approach is widely used in:

• Medical imaging
• Face recognition
• Plant disease detection

Industrial inspection

• Low-data deep learning scenarios
• A strong foundation for real-world deep learning applications.