def run(plot,
df,
savedir,
raw_box_savedir,
rgb_pool=None,
saved_model=None,
deepforest_model=None):
"""wrapper function for dask, see main.py"""
from deepforest import deepforest
#create deepforest model
if deepforest_model is None:
if saved_model is None:
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
else:
deepforest_model = deepforest.deepforest(saved_model=saved_model)
#Filter data and process
plot_data = df[df.plotID == plot]
predicted_trees, raw_boxes = process_plot(plot_data, rgb_pool,
deepforest_model)
#Write merged boxes to file as an interim piece of data to inspect.
predicted_trees.to_file("{}/{}_boxes.shp".format(savedir, plot))
raw_boxes.to_file("{}/{}_boxes.shp".format(raw_box_savedir, plot))
def test_multi_train(multi_annotations, tmpdir):
test_model = deepforest.deepforest()
test_model.use_release()
test_model.config["epochs"] = 1
test_model.config["save-snapshot"] = False
test_model.config["steps"] = 5
test_model.train(annotations=multi_annotations, input_type="fit_generator")
boxes = test_model.predict_generator(annotations=multi_annotations)
# Test labels
labels = list(test_model.labels.values())
labels.sort()
target_labels = ["Dead", "Alive"]
target_labels.sort()
assert labels == target_labels
#Test reload models
test_model.model.save("{}/prediction_model.h5".format(tmpdir))
# new object
new_session = deepforest.deepforest(
saved_model="{}/prediction_model.h5".format(tmpdir))
new_boxes = new_session.predict_generator(annotations=multi_annotations)
assert labels == new_boxes.labels.unique()
pd.testing.assert_frame_equal(boxes, new_boxes)
def run(plot,
df,
rgb_pool=None,
hyperspectral_pool=None,
sensor="hyperspectral",
extend_box=0,
hyperspectral_savedir=".",
saved_model=None):
"""wrapper function for dask, see main.py"""
try:
from deepforest import deepforest
#create deepforest model
if saved_model is None:
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
else:
deepforest_model = deepforest.deepforest(saved_model=saved_model)
#Filter data and process
plot_data = df[df.plotID == plot]
predicted_trees = process_plot(plot_data, rgb_pool, deepforest_model)
plot_crops, plot_labels, plot_box_index = create_crops(
predicted_trees,
hyperspectral_pool=hyperspectral_pool,
rgb_pool=rgb_pool,
sensor=sensor,
expand=extend_box,
hyperspectral_savedir=hyperspectral_savedir)
except:
print("Plot {} failed".format(plot))
raise
return plot_crops, plot_labels, plot_box_index
def load_model(weights=None):
if weights:
model = deepforest.deepforest(weights=weights)
else:
model = deepforest.deepforest()
model.use_release()
return model
def load_model():
try:
model = deepforest.deepforest(saved_model="deepforest-model.h5")
print("Trained Model loaded")
except:
model = deepforest.deepforest()
model.use_release()
print("Prebuilt Model loaded")
return model
def run(plot,
df,
rgb_pool=None,
hyperspectral_pool=None,
extend_HSI_box=0,
extend_RGB_box=0,
hyperspectral_savedir=".",
saved_model=None,
deepforest_model=None):
"""wrapper function for dask, see main.py"""
from deepforest import deepforest
#create deepforest model
if deepforest_model is None:
if saved_model is None:
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
else:
deepforest_model = deepforest.deepforest(saved_model=saved_model)
#Filter data and process
plot_data = df[df.plotID == plot]
predicted_trees = process_plot(plot_data,
rgb_pool,
deepforest_model,
debug=True)
#Write merged boxes to file as an interim piece of data to inspect.
interim_dir = os.path.abspath(ROOT)
predicted_trees.to_file("{}/data/interim/{}_boxes.shp".format(
interim_dir, plot))
#Crop HSI
plot_HSI_crops, plot_labels, plot_domains, plot_sites, plot_heights, plot_elevations, plot_box_index = create_crops(
predicted_trees,
hyperspectral_pool=hyperspectral_pool,
rgb_pool=rgb_pool,
sensor="hyperspectral",
expand=extend_HSI_box,
hyperspectral_savedir=hyperspectral_savedir)
#Crop RGB, drop repeated elements, leave one for testing
plot_rgb_crops, plot_rgb_labels, _, _, _, _, _ = create_crops(
predicted_trees,
hyperspectral_pool=hyperspectral_pool,
rgb_pool=rgb_pool,
sensor="rgb",
expand=extend_RGB_box,
hyperspectral_savedir=hyperspectral_savedir)
#Assert they are the same
assert len(plot_rgb_crops) == len(plot_HSI_crops)
assert plot_labels == plot_rgb_labels
return plot_HSI_crops, plot_rgb_crops, plot_labels, plot_domains, plot_sites, plot_heights, plot_elevations, plot_box_index
def test_random_transform(annotations):
test_model = deepforest.deepforest()
test_model.config["random_transform"] = True
classes_file = utilities.create_classes(annotations)
arg_list = utilities.format_args(annotations, classes_file,
test_model.config)
assert "--random-transform" in arg_list
def from_retinanet(annotations_csv):
model = deepforest.deepforest()
model.use_release()
"""use the keras retinanet source to create detections"""
### Keras retinanet
# Format args for CSV generator
classes_file = utilities.create_classes(annotations_csv)
arg_list = utilities.format_args(annotations_csv, classes_file,
model.config)
args = parse_args(arg_list)
# create generator
validation_generator = csv_generator.CSVGenerator(
args.annotations,
args.classes,
image_min_side=args.image_min_side,
image_max_side=args.image_max_side,
config=args.config,
shuffle_groups=False,
)
all_detections = _get_detections(validation_generator,
model.prediction_model,
score_threshold=args.score_threshold,
max_detections=100)
all_annotations = _get_annotations(validation_generator)
return all_detections, all_annotations, validation_generator
def test_empty_plot():
#DeepForest prediction
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
plot_data = gpd.read_file(data_path)
rgb_sensor_path = prepare_field_data.find_sensor_path(
bounds=plot_data.total_bounds, lookup_pool=rgb_pool)
boxes = prepare_field_data.predict_trees(deepforest_model=deepforest_model,
rgb_path=rgb_sensor_path,
bounds=plot_data.total_bounds)
#fake offset boxes by adding a scalar to the geometry
boxes["geometry"] = boxes["geometry"].translate(100000)
#Merge results with field data, buffer on edge
merged_boxes = gpd.sjoin(boxes, plot_data)
#If no remaining boxes just take a box around center
if merged_boxes.empty:
merged_boxes = prepare_field_data.create_boxes(plot_data)
#If there are multiple boxes, take the center box
grouped = merged_boxes.groupby("individual")
cleaned_boxes = []
for value, group in grouped:
choosen_box = prepare_field_data.choose_box(group, plot_data)
cleaned_boxes.append(choosen_box)
merged_boxes = gpd.GeoDataFrame(pd.concat(cleaned_boxes),
crs=merged_boxes.crs)
merged_boxes = merged_boxes.drop(columns=["xmin", "xmax", "ymin", "ymax"])
def test_process_plot():
df = gpd.read_file(data_path)
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
merged_boxes = prepare_field_data.process_plot(plot_data=df, rgb_pool=rgb_pool, deepforest_model=deepforest_model)
assert df.shape[0] <= merged_boxes.shape[0]
def test_train(annotations):
test_model = deepforest.deepforest()
test_model.config["epochs"] = 1
test_model.config["save-snapshot"] = False
test_model.config["steps"] = 1
test_model.train(annotations=annotations, input_type="fit_generator")
return test_model
def run(plot, df, rgb_pool=None, hyperspectral_pool=None, extend_box=0, hyperspectral_savedir=".",saved_model=None, deepforest_model=None):
"""wrapper function for dask, see main.py"""
try:
from deepforest import deepforest
#create deepforest model
if deepforest_model is None:
if saved_model is None:
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
else:
deepforest_model = deepforest.deepforest(saved_model=saved_model)
#Filter data and process
plot_data = df[df.plotID == plot]
predicted_trees = process_plot(plot_data, rgb_pool, deepforest_model)
#Crop HSI
plot_HSI_crops, plot_labels, plot_sites, plot_elevations, plot_box_index = create_crops(
predicted_trees,
hyperspectral_pool=hyperspectral_pool,
rgb_pool=rgb_pool,
sensor="hyperspectral",
expand=extend_box,
hyperspectral_savedir=hyperspectral_savedir)
#Crop RGB, drop repeated elements, leave one for testing
plot_rgb_crops, plot_rgb_labels, _, _, _ = create_crops(
predicted_trees,
hyperspectral_pool=hyperspectral_pool,
rgb_pool=rgb_pool,
sensor="rgb",
expand=extend_box,
hyperspectral_savedir=hyperspectral_savedir)
#Assert they are the same
assert len(plot_rgb_crops) == len(plot_HSI_crops)
assert plot_labels==plot_rgb_labels
except Exception as e:
print("Plot {} failed {}".format(plot, e))
raise
return plot_HSI_crops, plot_rgb_crops, plot_labels, plot_sites, plot_elevations, plot_box_index
def test_predict_image(download_release):
test_model = deepforest.deepforest(
weights="tests/data/universal_model_july30.h5")
assert isinstance(test_model.model, keras.models.Model)
boxes = test_model.predict_image(image_path="tests/data/OSBS_029.tif",
show=False,
return_plot=False)
#Returns a 4 column numpy array
assert isinstance(boxes, np.ndarray)
assert boxes.shape[1] == 4
def release_model(download_release):
test_model = deepforest.deepforest()
test_model.use_release()
# Check for release tag
assert isinstance(test_model.__release_version__, str)
# Assert is model instance
assert isinstance(test_model.model, keras.models.Model)
return test_model
def test_use_release(download_release):
test_model = deepforest.deepforest()
test_model.use_release()
# Check for release tag
assert isinstance(test_model.__release_version__, str)
# Assert is model instance
assert isinstance(test_model.model, keras.models.Model)
assert test_model.config["weights"] == test_model.weights
assert test_model.config["weights"] is not "None"
def from_source(annotations_csv):
"""Use the code in this repo to get detections"""
model = deepforest.deepforest()
model.use_release()
predicted_boxes = model.predict_generator(annotations_csv)
true_boxes = pd.read_csv(
annotations_csv,
names=["image_path", "xmin", "ymin", "xmax", "ymax", "label"])
true_boxes["plot_name"] = true_boxes.image_path.apply(
lambda x: os.path.splitext(x)[0])
return predicted_boxes, true_boxes
def main(field_data,
rgb_dir,
savedir,
raw_box_savedir,
saved_model=None,
client=None,
shuffle=True):
"""Prepare NEON field data into tfrecords
Args:
field_data: shp file with location and class of each field collected point
height: height in meters of the resized training image
width: width in meters of the resized training image
savedir: direcory to save predicted bounding boxes
raw_box_savedir: directory save all bounding boxes in the image
client: dask client object to use
Returns:
None: .shp bounding boxes are written to savedir
"""
df = gpd.read_file(field_data)
plot_names = df.plotID.unique()
rgb_pool = glob.glob(rgb_dir, recursive=True)
if client is not None:
futures = []
for plot in plot_names:
future = client.submit(run,
plot=plot,
df=df,
rgb_pool=rgb_pool,
saved_model=saved_model,
savedir=savedir,
raw_box_savedir=raw_box_savedir)
futures.append(future)
wait(futures)
else:
from deepforest import deepforest
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
for plot in plot_names:
try:
run(plot=plot,
df=df,
rgb_pool=rgb_pool,
saved_model=saved_model,
deepforest_model=deepforest_model,
raw_box_savedir=raw_box_savedir)
except Exception as e:
print("Plot failed with {}".format(e))
traceback.print_exc()
continue
def test_multi_train(multi_annotations):
test_model = deepforest.deepforest()
test_model.config["epochs"] = 1
test_model.config["save-snapshot"] = False
test_model.config["steps"] = 1
test_model.train(annotations=multi_annotations, input_type="fit_generator")
# Test labels
labels = list(test_model.labels.values())
labels.sort()
target_labels = ["Dead", "Alive"]
target_labels.sort()
assert labels == target_labels
def test_reload_weights(release_model):
release_model.model.save_weights("tests/output/example_saved_weights.h5")
reloaded = deepforest.deepforest(
weights="tests/output/example_saved_weights.h5")
assert reloaded.prediction_model
# Predict test image and return boxes
boxes = reloaded.predict_image(image_path=get_data("OSBS_029.tif"),
show=False,
return_plot=False,
score_threshold=0.1)
# Returns a 6 column numpy array, xmin, ymin, xmax, ymax, score, label
assert boxes.shape[1] == 6
def test_predict_image(download_release):
# Load model
test_model = deepforest.deepforest(weights=get_data("NEON.h5"))
assert isinstance(test_model.model, keras.models.Model)
# Predict test image and return boxes
boxes = test_model.predict_image(image_path=get_data("OSBS_029.tif"),
show=False,
return_plot=False,
score_threshold=0.1)
# Returns a 6 column numpy array, xmin, ymin, xmax, ymax, score, label
assert boxes.shape[1] == 6
assert boxes.score.min() > 0.1
def submission_no_chm(tiles_to_predict):
#Predict
results = []
model = deepforest.deepforest()
model.use_release()
for path in tiles_to_predict:
try:
result = model.predict_image(path,return_plot=False)
result["plot_name"] = os.path.splitext(os.path.basename(path))[0]
results.append(result)
except Exception as e:
print(e)
continue
#Create plot name groups
boxes = pd.concat(results)
return boxes
def test_mAP_deepforest(annotations_csv):
model = deepforest.deepforest()
model.use_release()
#Original retinanet implementation
mAP_retinanet = model.evaluate_generator(annotations=annotations_csv)
#This repo implementation
predicted_boxes = model.predict_generator(annotations_csv)
true_boxes = pd.read_csv(
annotations_csv,
names=["image_path", "xmin", "ymin", "xmax", "ymax", "label"])
true_boxes["plot_name"] = true_boxes.image_path.apply(
lambda x: os.path.splitext(x)[0])
mAP = average_precision.calculate_mAP(true_boxes,
predicted_boxes,
iou_threshold=0.5)
assert mAP_retinanet == mAP
def main(
field_data,
RGB_size,
HSI_size,
rgb_dir,
hyperspectral_dir,
savedir=".",
chunk_size=400,
extend_box=0,
hyperspectral_savedir=".",
n_workers=20,
saved_model=None,
use_dask=False,
shuffle=True,
species_classes_file=None,
site_classes_file=None):
"""Prepare NEON field data into tfrecords
Args:
field_data: shp file with location and class of each field collected point
height: height in meters of the resized training image
width: width in meters of the resized training image
savedir: direcory to save completed tfrecords
extend_box: units in meters to add to the edge of a predicted box to give more context
hyperspectral_savedir: location to save converted .h5 to .tif
n_workers: number of dask workers
species_classes_file: optional path to a two column csv file with index and species labels
site_classes_file: optional path to a two column csv file with index and site labels
shuffle: shuffle lists before writing
Returns:
tfrecords: list of created tfrecords
"""
df = gpd.read_file(field_data)
plot_names = df.plotID.unique()
hyperspectral_pool = glob.glob(hyperspectral_dir, recursive=True)
rgb_pool = glob.glob(rgb_dir, recursive=True)
labels = []
HSI_crops = []
RGB_crops = []
sites = []
box_indexes = []
elevations = []
if use_dask:
client = start_cluster.start(cpus=n_workers, mem_size="10GB")
futures = []
for plot in plot_names:
future = client.submit(
run,
plot=plot,
df=df,
rgb_pool=rgb_pool,
hyperspectral_pool=hyperspectral_pool,
extend_box=extend_box,
hyperspectral_savedir=hyperspectral_savedir,
saved_model=saved_model
)
futures.append(future)
wait(futures)
for x in futures:
try:
plot_HSI_crops, plot_RGB_crops, plot_labels, plot_sites, plot_elevations, plot_box_index = x.result()
#Append to general plot list
HSI_crops.extend(plot_HSI_crops)
RGB_crops.extend(plot_RGB_crops)
labels.extend(plot_labels)
sites.extend(plot_sites)
elevations.extend(plot_elevations)
box_indexes.extend(plot_box_index)
except Exception as e:
print("Future failed with {}".format(e))
traceback.print_exc()
else:
from deepforest import deepforest
deepforest_model = deepforest.deepforest()
deepforest_model.use_release()
for plot in plot_names:
try:
plot_HSI_crops, plot_RGB_crops, plot_labels, plot_sites, plot_elevations, plot_box_index = run(
plot=plot,
df=df,
rgb_pool=rgb_pool,
hyperspectral_pool=hyperspectral_pool,
extend_box=extend_box,
hyperspectral_savedir=hyperspectral_savedir,
saved_model=saved_model,
deepforest_model=deepforest_model
)
except Exception as e:
print("Plot failed with {}".format(e))
traceback.print_exc()
continue
#Append to general plot list
HSI_crops.extend(plot_HSI_crops)
RGB_crops.extend(plot_RGB_crops)
labels.extend(plot_labels)
sites.extend(plot_sites)
elevations.extend(plot_elevations)
box_indexes.extend(plot_box_index)
if shuffle:
z = list(zip(HSI_crops, RGB_crops, sites, elevations, box_indexes, labels))
random.shuffle(z)
HSI_crops, RGB_crops, sites, elevations, box_indexes, labels = zip(*z)
#If passes a species label dict
if species_classes_file is not None:
species_classdf = pd.read_csv(species_classes_file)
species_label_dict = species_classdf.set_index("taxonID").label.to_dict()
else:
#Create and save a new species and site label dict
unique_species_labels = np.unique(labels)
species_label_dict = {}
for index, label in enumerate(unique_species_labels):
species_label_dict[label] = index
pd.DataFrame(species_label_dict.items(), columns=["taxonID","label"]).to_csv("{}/species_class_labels.csv".format(savedir))
#If passes a site label dict
if site_classes_file is not None:
site_classdf = pd.read_csv(site_classes_file)
site_label_dict = site_classdf.set_index("siteID").label.to_dict()
else:
#Create and save a new site and site label dict
unique_site_labels = np.unique(sites)
site_label_dict = {}
for index, label in enumerate(unique_site_labels):
site_label_dict[label] = index
pd.DataFrame(site_label_dict.items(), columns=["siteID","label"]).to_csv("{}/site_class_labels.csv".format(savedir))
#Convert labels to numeric
numeric_labels = [species_label_dict[x] for x in labels]
numeric_sites = [site_label_dict[x] for x in sites]
#Write tfrecords
tfrecords = create_records(
HSI_crops=HSI_crops,
RGB_crops=RGB_crops,
labels=numeric_labels,
sites=numeric_sites,
elevations=elevations,
box_index=box_indexes,
savedir=savedir,
RGB_size=RGB_size,
HSI_size=HSI_size,
chunk_size=chunk_size)
return tfrecords
def test_use_release(download_release):
test_model = deepforest.deepforest()
test_model.use_release()
#Assert is model instance
assert isinstance(test_model.model, keras.models.Model)
def test_train(annotations):
test_model = deepforest.deepforest()
test_model.config["epochs"] = 1
test_model.config["save-snapshot"] = False
test_model.train(annotations="tests/data/OSBS_029.csv")
def test_deepforest():
model = deepforest.deepforest(weights=None)
assert model.weights is None
tfrecords_path = tfrecords.create_tfrecords(annotations_file,
class_file,
size=1)
print("Created {} tfrecords: {}".format(len(tfrecords_path), tfrecords_path))
inputs, targets = tfrecords.create_tensors(tfrecords_path)
#### Fit generator ##
comet_experiment = Experiment(api_key="ypQZhYfs3nSyKzOfz13iuJpj2",
project_name="deepforest",
workspace="bw4sz")
comet_experiment.log_parameter("Type", "testing")
comet_experiment.log_parameter("input_type", "fit_generator")
#Create model
fitgen_model = deepforest.deepforest()
fitgen_model.config["epochs"] = 1
comet_experiment.log_parameters(fitgen_model.config)
#Train model
fitgen_model.train(annotations_file,
input_type="fit_generator",
comet_experiment=comet_experiment,
images_per_epoch=1000)
#Evaluate on original annotations
mAP = fitgen_model.evaluate_generator(annotations_file, comet_experiment)
boxes = fitgen_model.prediction_model.predict(inputs, steps=1)
comet_experiment.log_metric("mAP", mAP)
from typing import Any, Dict, List, Optional
from _create_xml import create_label_xml
from deepforest import deepforest
import pandas as pd
SCRIPT_DIR = os.path.abspath(os.getcwd())
IMAGE_DIR_PATH = "../data/image_data"
LABEL_DIR_PATH = "../data/predictions"
MODEL_DIR_PATH = "../data/model_data"
IMAGE_WIDTH = 400
IMAGE_HEIGHT = 400
MODEL = deepforest.deepforest(saved_model=f"{MODEL_DIR_PATH}/model.h5")
def _predict_image(image_path: str) -> Optional[pd.DataFrame]:
try:
return MODEL.predict_image(f"{image_path}",
show=False,
return_plot=False)
except:
return None
def _create_label_dir(dir_path: str) -> None:
if not os.path.exists(dir_path):
os.makedirs(dir_path)
for sub_dir in ["csv", "xml"]:
import glob
from comet_ml import Experiment
from deepforest import deepforest
#Local debug. If False, paths on UF hypergator supercomputing cluster
DEBUG = False
INPUT_TYPE = "tfrecord"
if DEBUG:
BASE_PATH = "/Users/ben/Documents/NeonTreeEvaluation_analysis/Weinstein_unpublished/"
else:
BASE_PATH = "/orange/ewhite/b.weinstein/NeonTreeEvaluation/"
##Pretrain deepforest on Silva annotations
deepforest_model = deepforest.deepforest()
# import comet_ml logger
comet_experiment = Experiment(api_key="ypQZhYfs3nSyKzOfz13iuJpj2",
project_name="deepforest",
workspace="bw4sz")
comet_experiment.log_parameters(deepforest_model.config)
comet_experiment.log_parameter("Type", "Pretraining")
if INPUT_TYPE == "fit_generator":
comet_experiment.log_parameter("Profiler", "fit_generator")
training_file = pd.read_csv(
BASE_PATH + "pretraining/crops/pretraining.csv",
names=["image_path", "xmin", "ymin", "xmax", "ymax", "label"])
unique_images = training_file.image_path.unique()
def test_predict_image_raise_error(download_release):
#Predict test image and return boxes
test_model = deepforest.deepforest(weights=get_data("NEON.h5"))
with pytest.raises(ValueError):
boxes = test_model.predict_image()
