Repository for ODS Summer of Code event with preprocessing and segmentation pipelines.

Preprocessing pipeline uses GDAL and OpenCV libraries. Segmentation pipeline uses PyTorch and Segmentation Models frameworks.

It is supposed to be used with Google Colaboratory (File -> Open notebook -> GitHub).

Kaggle Kernels – is a cloud-based development environment similar to Google Colab. Kaggle provides free access to the GPU for 37 hours per week.

1. Download and upload notebook

   wget https://raw.githubusercontent.com/MaritimeAI/ODS-SoC/master/segmentation.kaggle.ipynb

   Open Kaggle → Create → New Notebook → File → Upload Notebook → segmentation.kaggle.ipynb

2. Add dataset

   File → Add or upload data

3. Use Kaggle secrets to store your API key wandb for experiment tracking and models managment

   Add-ons → Secrets → Add a new secrets → Label ('wandb') → Value (Key)

4. Configuration settings

   config = {
       'classes': ['nodata', 'water', 'ice'],
       'batch_size_train': 1,
       'batch_size_valid': 1,
       'num_workers_train': 1,
       'num_workers_valid': 1,
       'model_encoder': 'ResNet32',
       'model_pretrain': 'ImageNet',
       'model_channels': 3,
       'data_split': 1,
       'expand': True,
       'debug': True,
       'flat': True,
   }

   • data_split - cross-validation for 5 folds or comment this line
   • debug - test run for one image
   • model_encoder - choose any from list encoders

5. Choose architecture model

   Notebook → section Model

   model = smp.MAnet(encoder_name=config['model_encoder'].lower(),
                     encoder_weights=config['model_pretrain'].lower(),
                     in_channels=config['model_channels'],
                     classes=len(CLASSES))

6. Save weights model

   Save version → Advenced settings → Always save output → Quick save → Save

7. Upload weights and trained model

   File → Add or upload data → Notebook Output Files → Your work → Chosse previos version

   In section Paths copy notebook name to save the weights to working directory

   NOTEBOOK_NAME = ''

   In section Train loop add last saved model

   NAME_PRELOAD = ''