Lab - U-Net for RTS mapping

[carmengg]

U-Net for RTS mapping hands-on lab

Goal

To provide participants with practical experience in implementing and experimenting with the U-Net architecture for semantic segmentation on RTS data.

Breakdown

1. Data Loading
   • Task: Loading RTS data for the lab
2. Implementing U-Net Architecture
   • Task: Define the U-Net architecture using nn.Module in PyTorch
   • Guided step-by-step construction of the contracting path, bottleneck, and expansive path
   • Tips: Emphasize the importance of matching tensor dimensions
3. Defining the Loss Function and Optimizer
   • Task: Choose an appropriate loss function for segmentation (e.g., Dice loss, cross-entropy loss)
   • Set up an optimizer (e.g., Adam) for training
4. Mini Training Loop
   • Task: Implement a basic training loop to train the U-Net model on the RTS data subset
   • Monitor the loss and visualize some predictions after a few epochs
   • Tips: Discuss the importance of data augmentation and learning rate choices for segmentation tasks
5. Discussion and Troubleshooting
   • Share insights or observations from the training process
   • Encourage participants to discuss their experiences and any modifications they tried