#Get image name for indexing
image_name = os.path.basename(path_to_raster)
- # Load annotations file
+ # Load annotations file and coerce dtype
annotations = pd.read_csv(annotations_file)
#open annotations file
image_annotations = annotations[annotations.image_path == image_name].copy()
- if not annotations.shape[1]==6:
+ if not annotations.shape[1] == 6:
raise ValueError("Annotations file has {} columns, should have format image_path, xmin, ymin, xmax, ymax, label".format(annotations.shape[1]))
annotations_files = []
diff --git a/docs/_build/html/_modules/deepforest/retinanet_train.html b/docs/_build/html/_modules/deepforest/retinanet_train.html
index ce1cd5b0..0202ccf0 100644
--- a/docs/_build/html/_modules/deepforest/retinanet_train.html
+++ b/docs/_build/html/_modules/deepforest/retinanet_train.html
@@ -8,7 +8,7 @@
- deepforest.retinanet_train — DeepForest 0.0.3 documentation
+ deepforest.retinanet_train — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
diff --git a/docs/_build/html/_modules/deepforest/tfrecords.html b/docs/_build/html/_modules/deepforest/tfrecords.html
index 7255ab7b..87fbfa22 100644
--- a/docs/_build/html/_modules/deepforest/tfrecords.html
+++ b/docs/_build/html/_modules/deepforest/tfrecords.html
@@ -8,7 +8,7 @@
- deepforest.tfrecords — DeepForest 0.0.3 documentation
+ deepforest.tfrecords — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
@@ -166,12 +166,15 @@ Source code for deepforest.tfrecords
"""
import tensorflow as tf
import os
+import csv
import numpy as np
+
from math import ceil
import keras
import cv2
import pandas as pd
+
from keras_retinanet.preprocessing.csv_generator import CSVGenerator
from keras_retinanet import models
from keras_retinanet.models.retinanet import retinanet_bbox
@@ -222,6 +225,7 @@ Source code for deepforest.tfrecords
#filebase name
image_basename = os.path.splitext(os.path.basename(annotations_file))[0]
+ ## Syntax checks
#Check annotations file only JPEG, PNG, GIF, or BMP are allowed.
df = pd.read_csv(annotations_file, names=["image_path","xmin","ymin","xmax","ymax","label"])
df['FileType'] = df.image_path.str.split('.').str[-1].str.lower()
@@ -230,6 +234,13 @@ Source code for deepforest.tfrecords
if bad_file_types > 0:
raise ValueError("Check annotations file, only JPEG, PNG, GIF, or BMP are allowed, {} incorrect files found".format(bad_file_types))
+ #Check dtypes, cannot use pandas, or will coerce in the presence of NAs
+ with open(annotations_file, 'r') as f:
+ reader = csv.reader(f, delimiter=',')
+ row = next(reader)
+ if row[1].count(".") > 0:
+ raise ValueError("Annotation files should be headerless with integer box, {} is not a int".format(row[1]))
+
#Create generator - because of how retinanet yields data, this should always be 1. Shape problems in the future?
train_generator = CSVGenerator(
annotations_file,
diff --git a/docs/_build/html/_modules/deepforest/utilities.html b/docs/_build/html/_modules/deepforest/utilities.html
index b5fb2aa8..23cf8a2d 100644
--- a/docs/_build/html/_modules/deepforest/utilities.html
+++ b/docs/_build/html/_modules/deepforest/utilities.html
@@ -8,7 +8,7 @@
- deepforest.utilities — DeepForest 0.0.3 documentation
+ deepforest.utilities — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
@@ -199,14 +199,15 @@ Source code for deepforest.utilities
self.total = tsize
self.update(b * bsize - self.n)[docs]def use_release(save_dir = "data"):
+
[docs]def use_release(save_dir = "data", prebuilt_model="NEON"):
'''Check the existance of, or download the latest model release from github
Args:
save_dir (str): Directory to save filepath, default to "data" in toplevel repo
+
prebuilt_model: Currently only accepts "NEON", but could be expanded to include other prebuilt models. The local model will be called {prebuilt_model}.h5 on disk.
Returns:
-
output_path (str): path to downloaded model
+
release_tag, output_path (str): path to downloaded model
'''
#Find latest github tag release from the DeepLidar repo
@@ -215,19 +216,22 @@
Source code for deepforest.utilities
headers={'Accept': 'application/vnd.github.v3+json'},
)).read())
asset = _json['assets'][0]
- output_path = os.path.join(save_dir,asset['name'])
url = asset['browser_download_url']
+ #Naming based on pre-built model
+ output_path = os.path.join(save_dir,prebuilt_model + ".h5")
+
#Download if it doesn't exist
if not os.path.exists(output_path):
print("Downloading model from DeepForest release {}, see {} for details".format(_json["tag_name"],_json["html_url"]))
with DownloadProgressBar(unit='B', unit_scale=True,
miniters=1, desc=url.split('/')[-1]) as t:
urllib.request.urlretrieve(url, filename=output_path, reporthook=t.update_to)
+ print("Model was downloaded and saved to {}".format(output_path))
else:
print("Model from DeepForest release {} was already downloaded. Loading model from file.".format(_json["html_url"]))
- return output_path
+
return _json["html_url"], output_path
[docs]def xml_to_annotations(xml_path, rgb_dir):
"""Load annotations from xml format (e.g. RectLabel editor) and convert them into retinanet annotations format.
@@ -302,11 +306,11 @@
Source code for deepforest.utilities
classes_path = os.path.join(dirname,"classes.csv")
#get unique labels
- labels = annotations.label.unique()
+ labels = annotations.label.dropna().unique()
n_classes = labels.shape[0]
print("There are {} unique labels: {} ".format(n_classes,list(labels)))
- if n_classes > 1:
- raise ValueError("There are greater than 1 classes ({}), check annotation levels for file {}".format(list(labels), annotations_file))
+ #if n_classes > 1:
+ # raise ValueError("There are greater than 1 classes ({}), check annotation levels for file {}".format(list(labels), annotations_file))
#write label
with open(classes_path,'w') as csv_file:
diff --git a/docs/_build/html/_modules/index.html b/docs/_build/html/_modules/index.html
index 15b38fec..669573cf 100644
--- a/docs/_build/html/_modules/index.html
+++ b/docs/_build/html/_modules/index.html
@@ -8,7 +8,7 @@
- Overview: module code — DeepForest 0.0.3 documentation
+ Overview: module code — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
diff --git a/docs/_build/html/_modules/tests/test_deepforest.html b/docs/_build/html/_modules/tests/test_deepforest.html
index 41a7e463..7ce5d8e4 100644
--- a/docs/_build/html/_modules/tests/test_deepforest.html
+++ b/docs/_build/html/_modules/tests/test_deepforest.html
@@ -8,7 +8,7 @@
- tests.test_deepforest — DeepForest 0.0.3 documentation
+ tests.test_deepforest — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
@@ -198,27 +198,39 @@ Source code for tests.test_deepforest
model = deepforest.deepforest(weights=None)
assert model.weights is None[docs]@pytest.fixture()
-
def test_use_release(download_release):
+
[docs]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)
+
+
[docs]@pytest.fixture()
+
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
[docs]def test_predict_image(download_release):
-
#Load model
-
test_model = deepforest.deepforest(weights="tests/data/universal_model_july30.h5")
+
test_model = deepforest.deepforest(weights="tests/data/NEON.h5")
assert isinstance(test_model.model,keras.models.Model)
#Predict test image and return boxes
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
+
#Returns a 6 column numpy array, xmin, ymin, xmax, ymax, score, label
+
assert boxes.shape[1] == 6
[docs]@pytest.fixture()
def test_train(annotations):
@@ -230,7 +242,6 @@
Source code for tests.test_deepforest
-
#Check that weights have changed, following https://medium.com/@keeper6928/how-to-unit-test-machine-learning-code-57cf6fd81765
[docs]def test_train_no_freeze(annotations, test_train):
test_train.config["freeze_layers"] = 0
@@ -266,12 +277,12 @@
Source code for tests.test_deepforest
#Because the network uses batch norm layers, there is a slight drift in weights. It should be very small.
#assert np.array_equal(before, after)
-[docs]def test_predict_generator(test_use_release, annotations):
-
boxes = test_use_release.predict_generator(annotations=annotations)
-
assert boxes.shape[1] == 5
+[docs]def test_predict_generator(release_model, annotations):
+
boxes = release_model.predict_generator(annotations=annotations)
+
assert boxes.shape[1] == 7
-[docs]def test_evaluate(test_use_release, annotations):
-
mAP = test_use_release.evaluate_generator(annotations=annotations)
+
[docs]def test_evaluate(release_model, annotations):
+
mAP = release_model.evaluate_generator(annotations=annotations)
#Assert that function returns a float numpy value
assert mAP.dtype == float
@@ -291,20 +302,20 @@
Source code for tests.test_deepforest
arg_list = utilities.format_args(annotations, test_model.config)
assert "--random-transform" in arg_list
-
[docs]def test_predict_tile(test_use_release):
+
[docs]def test_predict_tile(release_model):
raster_path = "tests/data/OSBS_029.tif"
original_raster = Image.open(raster_path)
original_raster = np.array(original_raster)
#This should make the same prediction?
-
#predicted_raster = test_use_release.predict_tile(raster_path, return_plot = False, patch_size=400,patch_overlap=0)
-
#predicted_image = test_use_release.predict_image(raster_path, return_plot = False)
+
#predicted_raster = release_model.predict_tile(raster_path, return_plot = False, patch_size=400,patch_overlap=0)
+
#predicted_image = release_model.predict_image(raster_path, return_plot = False)
#assert predicted_raster.shape == predicted_image.shape
#Debug
-
predicted_raster = test_use_release.predict_tile(raster_path, return_plot = True, patch_size=300,patch_overlap=0.5)
-
predicted_image = test_use_release.predict_image(raster_path, return_plot = True)
+
predicted_raster = release_model.predict_tile(raster_path, return_plot = True, patch_size=300,patch_overlap=0.5)
+
predicted_image = release_model.predict_image(raster_path, return_plot = True)
assert original_raster.shape == predicted_raster.shape
diff --git a/docs/_build/html/_modules/tests/test_preprocess.html b/docs/_build/html/_modules/tests/test_preprocess.html
index c08c195c..48667727 100644
--- a/docs/_build/html/_modules/tests/test_preprocess.html
+++ b/docs/_build/html/_modules/tests/test_preprocess.html
@@ -8,7 +8,7 @@
-
tests.test_preprocess — DeepForest 0.0.3 documentation
+
tests.test_preprocess — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
diff --git a/docs/_build/html/_modules/tests/test_tfrecords.html b/docs/_build/html/_modules/tests/test_tfrecords.html
index dae670b4..496c923e 100644
--- a/docs/_build/html/_modules/tests/test_tfrecords.html
+++ b/docs/_build/html/_modules/tests/test_tfrecords.html
@@ -8,7 +8,7 @@
-
tests.test_tfrecords — DeepForest 0.0.3 documentation
+
tests.test_tfrecords — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
diff --git a/docs/_build/html/_modules/tests/test_utilities.html b/docs/_build/html/_modules/tests/test_utilities.html
index 2b64c46f..fb4ce38a 100644
--- a/docs/_build/html/_modules/tests/test_utilities.html
+++ b/docs/_build/html/_modules/tests/test_utilities.html
@@ -8,7 +8,7 @@
-
tests.test_utilities — DeepForest 0.0.3 documentation
+
tests.test_utilities — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
diff --git a/docs/_build/html/_sources/getting_started.md.txt b/docs/_build/html/_sources/getting_started.md.txt
index 27a29da3..9b89b71b 100644
--- a/docs/_build/html/_sources/getting_started.md.txt
+++ b/docs/_build/html/_sources/getting_started.md.txt
@@ -18,7 +18,7 @@ bioRxiv 790071; doi: https://doi.org/10.1101/790071
# Training
-While the prebuilt models outperform available tree crown detection tools, they can always be improved by adding data from the target area. In our work, we have found that even one hour worth of carefully chosen hand annotation can yield enormous improvements in accuracy and precision. We envision that for the majority of scientific applications, atleast some finetuning of the prebuilt model will be worthwhile.
+The prebuilt models can always be improved by adding data from the target area. In our work, we have found that even one hour worth of carefully chosen hand annotation can yield enormous improvements in accuracy and precision. We envision that for the majority of scientific applications, atleast some finetuning of the prebuilt model will be worthwhile.
```{python}
@@ -26,14 +26,22 @@ While the prebuilt models outperform available tree crown detection tools, they
```
-## Training Hardware
+# Evaluation
-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).
+Independent analysis of whether a model can generalize from training data to new areas is critical for creating a robust model. We stress that evaluation data must be different from training data, as neural networks have millions of parameters and can easily memorize thousands of samples. Therefore, while it would be rather easy to tune the model to get extremely high scores on the training data, it would fail when exposed to new images.
-For more details and recommendations for model training see [training overview]().
+DeepForest uses the keras-retinanet ```evaluate``` method to score images. This consists of an annotations.csv file in the following format
-# Evaluation
+```
+image_path, xmin, ymin, xmax, ymax, label
+```
+
+```{python}
+
+
+```
+For more on evaluation, see the [Evaluation Overview]()
## NEON Benchmark
diff --git a/docs/_build/html/_sources/training.md.txt b/docs/_build/html/_sources/training.md.txt
index b5f74074..9320e30b 100644
--- a/docs/_build/html/_sources/training.md.txt
+++ b/docs/_build/html/_sources/training.md.txt
@@ -22,5 +22,8 @@ Rectlabel screenshot
xml format explanation
+## Training Hardware
-### Guidelines for model Training
+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).
+
+For more details and recommendations for model training see [training overview]().
diff --git a/docs/_build/html/_static/documentation_options.js b/docs/_build/html/_static/documentation_options.js
index 99cd9e6a..23bb16ed 100644
--- a/docs/_build/html/_static/documentation_options.js
+++ b/docs/_build/html/_static/documentation_options.js
@@ -1,6 +1,6 @@
var DOCUMENTATION_OPTIONS = {
URL_ROOT: document.getElementById("documentation_options").getAttribute('data-url_root'),
- VERSION: '0.0.3',
+ VERSION: '0.1.0',
LANGUAGE: 'None',
COLLAPSE_INDEX: false,
FILE_SUFFIX: '.html',
diff --git a/docs/_build/html/authors.html b/docs/_build/html/authors.html
index d0818ae1..7f411595 100644
--- a/docs/_build/html/authors.html
+++ b/docs/_build/html/authors.html
@@ -8,7 +8,7 @@
-
<no title> — DeepForest 0.0.3 documentation
+
<no title> — DeepForest 0.1.0 documentation
@@ -59,7 +59,7 @@
- 0.0.3
+ 0.1.0
diff --git a/docs/_build/html/genindex.html b/docs/_build/html/genindex.html
index c3b80ff6..dfb342f7 100644
--- a/docs/_build/html/genindex.html
+++ b/docs/_build/html/genindex.html
@@ -9,7 +9,7 @@
-
Index — DeepForest 0.0.3 documentation
+
Index — DeepForest 0.1.0 documentation
@@ -60,7 +60,7 @@
- 0.0.3
+ 0.1.0
@@ -330,10 +330,12 @@
M
N
|
|
@@ -343,8 +345,6 @@ P
diff --git a/docs/_build/html/getting_started.html b/docs/_build/html/getting_started.html
index 184352ea..48976459 100644
--- a/docs/_build/html/getting_started.html
+++ b/docs/_build/html/getting_started.html
@@ -8,7 +8,7 @@
- What is DeepForest? — DeepForest 0.0.3 documentation
+ What is DeepForest? — DeepForest 0.1.0 documentation
@@ -61,7 +61,7 @@
- 0.0.3
+ 0.1.0
@@ -90,10 +90,7 @@
What is DeepForest?
How does it work?
Prebuilt models
-Training
-
+Training
Evaluation
@@ -185,19 +182,23 @@ Prebuilt models