CNN for multilabel classification Amazon Forest.pptx

CNN for multilabel classification

1. PA228: Project Amazon Forest: Multi-Label Image Classification

2. Methodology Task: To predict multiple labels (total 17) w.r.t. given satellite image of amazon forest. Major steps: Data analysis. Set of 40479 images: Split into 60% train, 20% validation & 20% test subsets. Each label converted to one-hot encoded vector. Images converted to numpy arrays & then to tensors. PyTorch neural network class: 3 convolutional layers with ReLU activation function. 2 fully connected layers (fc1 and fc2) with ReLU activation functions and a dropout layer. O utput layer using sigmoid activation function (17 classes). Training: L oss Function(BCE) & Optimizer(Adam), multiple configurations (10, 15, 20 epochs with different learning rates) tried and loss and results of testing analysed.

3. Results   Actual-Predicted label comparison (total 8096 x 17 = 137632): Correctly predicted labels: 129777 Wrongly predicted labels: 7855 F-beta Score: 0.943 Precision Weighted Score: 0.941 Reducing loss value during training phase. Example trained model output:

4. Conclusion   The project’s concept and working on multi-label classification has been a new learning. Using a GPU explicitly for training a network, has been very informative part of project. Difficult to understand how and why results varying so much with small configuration changes. Likely improvement can be achieved by making the training data more balanced (follow up task).

5. Thank You