Default training configuration seems not to correspond with the paper.

mapillary / seamseg

Seamless Scene Segmentation

BSD 3-Clause "New" or "Revised" License

289 stars 53 forks source link

Default training configuration seems not to correspond with the paper. #10

Closed GengDavid closed 4 years ago

GengDavid commented 4 years ago

The default training configuration at config/defaults seems not to correspond with the paper such as anchor ratios, training schedule, input resolution for Vista Dataset.

ducksoup commented 4 years ago

@GengDavid the default configuration is not meant to reproduce the results from the paper, it's there just as a reference for "reasonable" parameters and to explain what the parameters are for.

If you are interested, the config.ini file included together with our Vistas pre-trained model contains the actual parameters that were used to train that model (using 8x V100 32GB gpus).