search

neptune-ai / open-solution-mapping-challenge

Open solution to the Mapping Challenge :earth_americas:
https://www.crowdai.org/challenges/mapping-challenge
MIT License
380 stars 96 forks source link

hard coded target sizes #150

Closed taraspiotr closed 6 years ago

taraspiotr commented 6 years ago

Right now target_sizes are definded in pipeline_manager.py:

data = {'input': {'meta': meta_train,
                      'target_sizes': [(300, 300)] * len(meta_train)},
            'specs': {'train_mode': True},
            'callback_input': {'meta_valid': meta_valid}
            }

It can be problematic when images are all the same. I was thinking about storing target_sizes, which are also source sizes, in metadata.

jakubczakon commented 6 years ago

@taraspiotr that is how we usually do it. but this time, since it was all one size we left it like that. Feel free to refactor/add to meta when using other data sources.