Modify input image size for training/inference

neptune-ai / open-solution-mapping-challenge

Open solution to the Mapping Challenge :earth_americas:

https://www.crowdai.org/challenges/mapping-challenge

MIT License

377 stars 96 forks source link

Modify input image size for training/inference #215

Open cvKDean opened 4 years ago

cvKDean commented 4 years ago

Good day,

I would like to train on my own map images. However I noticed that the input image size is 300 300, following the crowdAI (AIcrowd) Mapping Challenge dataset. Is it possible to train on images with a different size/resolution (e.g. 900 900. 512*512)? Thank you.

jakubczakon commented 4 years ago

Sorry for late reply @cvKDean. I think you just need to change https://github.com/neptune-ml/open-solution-mapping-challenge/blob/master/neptune.yaml#L75-L76 to your size and you should be good to go.

cvKDean commented 4 years ago

Thank you for the response @jakubczakon. I was able to train on my own temporary dataset. However, the training seems to be finished at around epoch 30-31, when the number of epochs in my neptune.yaml file is set to 100. Is there something in the model that makes it finish training even before it reaches the defined number of epochs?