Training Hardware

Training neural networks is computationally intensive. While small amounts of data, on the order of several hundred trees, can be trained on a laptop in a few hours, large amounts of data are best trained on dedicated graphical processing units (GPUs). Many university clusters have GPUs available, and they can be rented for short periods of time on cloud servers (AWS, Google Cloud, Azure).

Tracking experiments

To track experiments, we recommend using a comet_ml dashboard. DeepForest train and evaluate objects accept comet experiments.

from comet_ml import Experiment
from deepforest import deepforest

test_model = deepforest.deepforest()

comet_experiment = Experiment(api_key=<api_key>,
                                  project_name=<project>, workspace=<"username">)


test_model.train(annotations=annotations_file, input_type="fit_generator",comet_experiment=comet_experiment)

Fit_generator versus tfrecords

There are currently two ways to train a deepforest model, directly using the annotations file described above, or wrapping those data into a tfrecords files. The benefits of annotations file, which uses a keras fit_generator method, is its simplicity and transparency. The downside is training speed. For the vast majority of projects, using a single GPU will be sufficient for training data. However, if you using any pretraining or semi-supervised approach, and have millions or tens of millions of samples, the fit_generator does not scale well across multiple GPUs. To create a tfrecords file:

  1. Optional -> generate crops from training tiles
annotations_file = preprocess.split_raster(<path_to_raster>, config["annotations_file"], "tests/data/",config["patch_size"], config["patch_overlap"])
  1. Generate the anchors for training from the annotations file
created_records = tfrecords.create_tfrecords(annotations_file="tests/data/testfile_tfrecords.csv",
  1. Train the model by supplying a list of tfrecords and the original file
test_model.train(annotations="tests/data/testfile_tfrecords.csv",input_type="tfrecord", list_of_tfrecords=created_records)

This approach is >10X faster when used to scale across 8 GPUs on a single machine. Please note that tfrecords are very large on disk.