Skip to content

Latest commit

 

History

History
200 lines (139 loc) · 5.95 KB

README.md

File metadata and controls

200 lines (139 loc) · 5.95 KB
Raw

deepforestr

Lifecycle: experimental R-CMD-check-release CRAN status

R interface for DeepForest Python package, a deep learning package detecting individual organisms in airborne RGB images.

Installation

deepforestr is an R wrapper for the Python package, DeepForest. This means that Python and the DeepForest Python package need to be installed first.

Basic Installation

If you just want to use DeepForest from within R run the following commands in R. This will create a local Python installation that will only be used by R and install the needed Python package for you. If installing on Windows you need to install RTools before installing the R package.

devtools::install_github('weecology/deepforestr') # Install the R package from GitHub
deepforestr::install_deepforest() # Install Python & DeepForest; Takes ~3 minutes

After running these commands restart R.

Advanced Installation for Python Users

If you are using Python for other tasks you can use deepforestr with your existing Python installation (though the basic installation above will still work by creating a separate miniconda install and Python environment).

Install the DeepForest Python package

Install the DeepForest Python package into your prefered Python environment using pip:

pip install DeepForest

Select the Python environment to use in R

deepforestr will try to find Python environments with DeepForest (see the reticulate documentation on order of discovery for more details) installed. Alternatively you can select a Python environment to use when working with deepforestr (and other packages using reticulate).

The most robust way to do this is to set the RETICULATE_PYTHON environment variable to point to the preferred Python executable:

Sys.setenv(RETICULATE_PYTHON = "/path/to/python")

This command can be run interactively or placed in .Renviron in your home directory.

Alternatively you can select the Python environment through the reticulate package for either conda:

library(reticulate)
conda_create("name_of_environment")

# install Python packages e.g SciPy, deepforest
conda_install("name_of_environment", "scipy", "deepforest")
use_condaenv('name_of_environment')

or virtualenv:

library(reticulate)
# create a new environment
virtualenv_create("name_of_environment")

# install Python packages
virtualenv_install("name_of_environment", "scipy", "deepforest")
use_virtualenv("name_of_environment")

You can check to see which Python environment is being used with:

py_config()

Install the deepforestr R package

remotes::install_github("weecology/deepforestr") # development version from GitHub

Getting Started

Load the current release model

library(deepforestr)

model = df_model()
model$use_release()

Predict a single image

Return the bounding boxes in a data frame

image_path = get_data("OSBS_029.png") # Gets a path to an example image
bounding_boxes = model$predict_image(path=image_path, return_plot=FALSE)
head(bounding_boxes)

Return an image for visualization

image_path = get_data("OSBS_029.png") # Gets a path to an example image
predicted_image = model$predict_image(path=image_path, return_plot=TRUE)
plot(raster::as.raster(predicted_image[,,3:1]/255))

Predict a tile

Return the bounding boxes in a data frame

raster_path = get_data("OSBS_029.tif") # Gets a path to an example raster tile
bounding_boxes = model$predict_tile(raster_path, return_plot=FALSE)
head(bounding_boxes)

Return an image for visualization

raster_path = get_data("OSBS_029.tif") # Gets a path to an example raster tile
predicted_raster = model$predict_tile(raster_path, return_plot=TRUE, patch_size=300L, patch_overlap=0.25)
plot(raster::as.raster(predicted_raster[,,3:1]/255))

Note at patch_size is an integer value in Python and therefore needs to have the L at the end of the number in R to make it an integer.

Predict a set of annotations

csv_file = get_data("testfile_deepforest.csv")
root_dir = get_data(".")
boxes = model$predict_file(csv_file=csv_file, root_dir = root_dir, savedir=".")

Training

Set the training configuration

annotations_file = get_data("testfile_deepforest.csv")

model$config$epochs = 1
model$config["save-snapshot"] = FALSE
model$config$train$csv_file = annotations_file
model$config$train$root_dir = get_data(".")

Optionally turn on fast_dev_run for testing and debugging:

model$config$train$fast_dev_run = TRUE

Train the model

model$create_trainer()
model$trainer$fit(model)

Evaluation

csv_file = get_data("OSBS_029.csv")
root_dir = get_data(".")
results = model$evaluate(csv_file, root_dir, iou_threshold = 0.4)

Saving & Loading Models

Saving a model after training

model$trainer$save_checkpoint("checkpoint.pl")

Loading a saved model

new_model = df_model()
after = new_model$load_from_checkpoint("checkpoint.pl")
pred_after_reload = after$predict_image(path = img_path)

Note that when reloading models, you should carefully inspect the model parameters, such as the score_thresh and nms_thresh. These parameters are updated during model creation and the config file is not read when loading from checkpoint! It is best to be direct to specify after loading checkpoint.