Pytorch release

This a major version change transition from tensorflow backend to pytorch. The pytorch backend was previously in a different repo and pypi package deepforest-pytorch. Continued tensorflow updates means that the pinned 1.14.0 version for keras-retinanet (which itself is deprecated) is a real risk to package longevity. To avoid needing to patch upstream dependencies we are deprecating the tensorflow backend and moving to pytorch. The release

model score is very close, within 1% of the tensorflow model, and we do not expect significant performance changes. Please see the README for links to updating code and please submit an issues you may have as we transition to 1.0.

This release is the same as https://github.com/weecology/DeepForest-pytorch/releases/tag/v0.1.17

Conda Version Distribution

This release is inline with the first release of the conda build. The prebuilt model has not changed from previous release.

21SiteModel

Model training parameters: https://www.comet.ml/bw4sz/deepforest/fabe532d9e5f4edaa98edf0d2c080011

red is the new release, blue the previous release. Recall and precision for each of the sites in the NeonTreeEvaluation Benchmark

> summary_statistics(results,method="all")
# A tibble: 2 x 3
  Method                mean_precision mean_recall
  <chr>                          <dbl>       <dbl>
1 Weinstein_unpublished          0.617       0.726
2 Weinstein2019                  0.567       0.645

https://github.com/weecology/NeonTreeEvaluation

Python Package

This is the first model version - corresponding to Weinstein et al. 2019. Four site NEON model (NIWO, TEAK, SJER, MLBS sites) with pretraining and hand-annotations.