Code for iMaterialist 2019
This code only example how train model (fcn_resnet50 or deeplabv3_resnet50) for image segmentation with TPU and PyTorch
For train use scripts:
script/convert_csv - conver csv file to json for faster access
script/train.sh - run training loops with TPU
script/inference.sh - make submission file
Don't forget : conda activate torch-xla-0.5
For all scripts: TPU_IP_ADDRESS=10.128.17.2 - Replace it on YOUR TPU NODE IP
notebook/Fast Visualizing Model.ipynb - contained visualized prediction for trained model on 200 images. Normal training too long.