solution-6: Dev lgbm (#147) (#152)

• added lgbm based scoring model

Clean up

• cleaned up code
• added docstrings
• sped-up tta
• updated hyperparams

solution-4: Dev (#121)

* Dev fix generate_metadata (#106)

* fix generate_metadata

* temp fix in callbacks

* Dev loader refactor (#111)

* refactored loaders, fixed weighed loss calculation

* updated config

* reverted resnet 152 to 32 filters with no dropout

* removed print

* back to 1000 valid size

* back to old batch size

* back other params

* config

* fix bug in loaders (#112)

* fix validation bug (#115)

* Dev tta (#116)

* scheleton added

* transformation and inverse done

* tta working

* Dev validation using mAP precision (#117)

* mege TTA with mAP validation

* mAP validation

* fix mAP valid in eval (#120)

solution-3: Dev (#105)

* fix steps issues #48 and #49

* prepare sum of distances, not distances to 2 closest objects

* back to const erosion

* fixes

* fix

* [:-4] -> os.path.splitext()

* loss weighted by size of the object

* prepare masks, distances and sizes

* cleaning

* adapt models and loaders to handle size matrix and calculate size weights

* adapt models and loaders to handle size matrix and calculate size weights v2

* fix pipelines.py

* fix some issues with calculating size-weighted loss

* cleaning

* update mean and std

* fixes

* clean

* fix recall in evaluation

* fix bug in erosion (#91)

* Dev mosaic padding inference (#81)

* added mosaic seq, unet_mosaic pipe, mosaic loader

* added unet_weighted

* dropped input resize at inference

* dropped rescaling in loader, fixed postpro cropping

* local dev

* updated dilation/erosion, joined pipelines

* dropped unet mask saving

* added replication padding

* renamed mosaic->padded, moved params to configs

* padding->inference_padding

* config updates

* refactored padded unet

* refactored unet_padding

* Dev dice loss (#89)

* fix size weights

* add mixed dice + weighted ce loss

* fixes

* parametrize loss weights

* remove get_datagen function overriding

* dice loss per channel, some fixes

* fixes and smooth added to Dice loss instead of eps

* fixes and smooth added to Dice loss and eps, and parametrized

* sigmoid -> softmax in dice loss

* softmax2d

* move softmax to models.py

* parametrize softmax and sigmoid in dice loss

* Dev mask prep speed up (#94)

* distributed mask/distance/size generation added

* dropped deprecated

* dropped mask param

* Dev random crop (#97)

* local

* added random cropping, refactored augmentations

* Dev borders and dilation in preprocessing (#96)

* merge multithread

* preparing borders

* fix PR #96 and add update metadata generation

* Dev deeper archs (#102)

* dropped mask param

* added deeper resnets and spatial2d dropout

* updated config

* fixed casting

* updated index

* fix evaluate, add score builder in stream mode (#104)

* added initial version

* added simple evaluate on checkpoint script

* updated config

* added neptune file definition

* fixed conflicts

solution-2: Merge pull request #51 from minerva-ml/dev

Evaluation in chunks added, erosion pre - dilation post approach added, multiclass problem definition enabled

solution-1: Update README.md

Typo in readme