Hyperparameters for Training/evaluation

[sarathsrk]

Hello, I have opened new thread to ask details about paramters, I will close that old issue.
My full image size is 1024x1024, In "dataset_params" you configured it as "data_patch_size": 725, and "input_patch_size": 512, but the Inria dataset image resolution is 5000x5000.
How did you give it as 725 & 512? Is it just to reduce GPU memory allocation?

[Lydorn]

Yes, the 5000x5000 image tiles are split into 725x725 patches, on which image augmention operations are applied. And then they are cropped to 512x512 (so that the black pixels from random rotations are removed). Using 512x512 patches allows to increase the batch size (4 times compared to 1024x1024). But you can try different values, or if you have enough VRAM you can try not cropping at all.

[sarathsrk]

Thank you, now I clearly understood the functions of these params. I have another question regarding "save_individual_outputs"
My config is
save_individual_outputs": {"seg": true, "seg_mask": true, "poly_shapefile": true, "poly_viz": true, "image": false, "seg_gt": false, "seg_opencities_mask": false, "seg_luxcarta": false, "crossfield": false, "uv_angles": false, "poly_geojson": true}

After inference done, it stores seg_mask, seg, poly_viz, cross_field(npy) and seg_luxcarta but it is not storing poly_shapefile and poly_geojson . in the data directory.

my command for inference python main.py --run_name 'inria_dataset_polygonized.unet_resnet101_pretrained.aaa' --in_filepath "./data/aaa115.png" Any idea on how do I enable shp file and geojson saving method?