Predicting windspeed of satellite images of hurricanes

This notebook was created for a Driven Data competition.

Tools Used

• Google Collab Notebook w/ TPUs
• Python, Keras, Tensorflow
• Convolutional Neural Network (CNN)

Technical Highlights

• Training set consisted of 70,000 366x366 JPG files (1 GB)
• To improve performance, Google Collab's TPUs were used
• JPG files were convered to TFRecord format and the TFRecord files were stored in a Google Bucket
• The TFRecords were streamed in batches to the model during training
• During training, model checkpoints (for the Best MSE) are saved to Google Drive
• At the end of training, the final model is also saved to Google Drive