def use_release(self, gpus=1):
"""Use the latest DeepForest model release from github and load model.
Optionally download if release doesn't exist.
Returns:
model (object): A trained keras model
gpus: number of gpus to parallelize, default to 1
"""
# Download latest model from github release
release_tag, self.weights = utilities.use_release()
# load saved model and tag release
self.__release_version__ = release_tag
print("Loading pre-built model: {}".format(release_tag))
if gpus == 1:
with warnings.catch_warnings():
# Suppress compilte warning, not relevant here
warnings.filterwarnings("ignore", category=UserWarning)
self.model = utilities.read_model(self.weights, self.config)
# Convert model
self.prediction_model = convert_model(self.model)
elif gpus > 1:
backbone = models.backbone(self.config["backbone"])
n_classes = len(self.labels.keys())
self.model, self.training_model, self.prediction_model = create_models(
backbone.retinanet,
num_classes=n_classes,
weights=self.weights,
multi_gpu=gpus)
# add to config
self.config["weights"] = self.weights
def test_lengths(config):
"""Assert that a csv generator and tfrecords create
the same number of images in a epoch"""
created_records = tfrecords.create_tfrecords(
annotations_file="tests/output/testfile_tfrecords.csv",
class_file="tests/output/classes.csv",
image_min_side=config["image-min-side"],
backbone_model=config["backbone"],
size=100,
savedir="tests/output/")
# tfdata
tf_filenames = find_tf_filenames(path="tests/output/*.tfrecord")
# keras generator
backbone = models.backbone(config["backbone"])
generator = csv_generator.CSVGenerator(
csv_data_file="tests/output/testfile_tfrecords.csv",
csv_class_file="tests/output/classes.csv",
image_min_side=config["image-min-side"],
preprocess_image=backbone.preprocess_image,
)
fit_genertor_length = generator.size()
assert len(tf_filenames) == fit_genertor_length
def __init__(self, weights=None, saved_model=None):
self.weights = weights
self.saved_model = saved_model
# Read config file - if a config file exists in local dir use it,
# if not use installed.
if os.path.exists("deepforest_config.yml"):
config_path = "deepforest_config.yml"
else:
try:
config_path = get_data("deepforest_config.yml")
except Exception as e:
raise ValueError(
"No deepforest_config.yml found either in local "
"directory or in installed package location. {}".format(e))
print("Reading config file: {}".format(config_path))
self.config = utilities.read_config(config_path)
# Create a label dict, defaults to "Tree"
self.read_classes()
# release version id to flag if release is being used
self.__release_version__ = None
# Load saved model if needed
if self.saved_model:
print("Loading saved model")
# Capture user warning, not relevant here
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
self.model = models.load_model(saved_model)
self.prediction_model = convert_model(self.model)
elif self.weights:
print("Creating model from weights")
backbone = models.backbone(self.config["backbone"])
self.model, self.training_model, self.prediction_model = create_models(
backbone.retinanet, num_classes=1, weights=self.weights)
else:
print(
"A blank deepforest object created. "
"To perform prediction, either train or load an existing model."
)
self.model = None
def test_equivalence(config, setup_create_tensors):
# unpack created tensors
tf_inputs, tf_targets = setup_create_tensors
# the image going in to tensorflow should be equivalent
# to the image from the fit_generator
backbone = models.backbone(config["backbone"])
# CSV generator
generator = csv_generator.CSVGenerator(
csv_data_file="tests/output/testfile_tfrecords.csv",
csv_class_file="tests/data/classes.csv",
image_min_side=config["image-min-side"],
preprocess_image=backbone.preprocess_image,
)
# find file in randomize generator group
first_file = generator.groups[0][0]
gen_filename = os.path.join(generator.base_dir,
generator.image_names[first_file])
original_image = generator.load_image(first_file)
inputs, targets = generator.__getitem__(0)
image = inputs[0, ...]
targets = targets[0][0, ...]
with tf.Session() as sess:
# seek the randomized image to match
tf_inputs, tf_targets = sess.run([tf_inputs, tf_targets])
# assert filename is the same as generator
# assert gen_filename == filename
# tf_image = tf_image[0,...]
tf_inputs = tf_inputs[0, ...]
tf_targets = tf_targets[0][0, ...]
# Same shape
# assert tf_image.shape == image.shape
assert tf_inputs.shape == image.shape
assert tf_targets.shape == targets.shape
def create_tfrecords(annotations_file,
class_file,
backbone_model="resnet50",
image_min_side=800,
size=1,
savedir="./"):
"""
Args:
annotations_file: path to 6 column data in form
image_path, xmin, ymin, xmax, ymax, label
backbone_model: A keras retinanet backbone
image_min_side: resized image object minimum size
size: Number of images per tfrecord
savedir: dir path to save tfrecords files
Returns:
written_files: A list of path names of written tfrecords
"""
memory_used = []
# Image preprocess function
backbone = models.backbone(backbone_model)
# 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()
# bad_files = df[~df['FileType'].isin(["jpeg","jpg","png","gif","bmp"])]
# if not bad_files.empty:
# raise ValueError("Check annotations file, only JPEG, PNG, GIF, or BMP are allowed,
# {} incorrect files found /n {}: ".format(bad_files.shape[0],bad_files.head()))
# 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,
class_file,
batch_size=1,
image_min_side=image_min_side,
preprocess_image=backbone.preprocess_image)
# chunk size
indices = np.arange(train_generator.size())
chunks = [
indices[i * size:(i * size) + size]
for i in range(ceil(len(indices) / size))
]
written_files = []
for chunk in chunks:
# Create tfrecord dataset and save it for output
fname = savedir + "{}_{}.tfrecord".format(image_basename, chunk[0])
written_files.append(fname)
writer = tf.io.TFRecordWriter(fname)
images = []
regression_targets = []
class_targets = []
filename = []
original_image = []
for i in chunk:
# Original image
original_image.append(train_generator.load_image(i))
batch = train_generator.__getitem__(i),
# split into images and tar gets
inputs, targets = batch[0]
# grab image, asssume batch size of 1, squeeze
images.append(inputs[0, ...])
# Grab anchor targets
regression_batch, labels_batch = targets
# grab regression anchors
# regression_batch: batch that contains bounding-box regression targets
# for an image & anchor states (np.array of shape (batch_size, N, 4 + 1),
# where N is the number of anchors for an image, the first 4 columns
# define regression targets for (x1, y1, x2, y2) and the
# last column defines anchor states (-1 for ignore, 0 for bg, 1 for fg).
regression_anchors = regression_batch[0, ...]
regression_targets.append(regression_anchors)
# grab class labels - squeeze out batch size
# From retinanet: labels_batch: batch that contains labels & anchor states
# (np.array of shape (batch_size, N, num_classes + 1),
# where N is the number of anchors for an image and the last column defines
# the anchor state (-1 for ignore, 0 for bg, 1 for fg).
labels = labels_batch[0, ...]
print("Label shape is: {}".format(labels.shape))
class_targets.append(labels)
# append filename by looking at group index
current_index = train_generator.groups[i][0]
# Grab filename and append to the full path
fname = train_generator.image_names[current_index]
fname = os.path.join(train_generator.base_dir, fname)
filename.append(fname)
for image, regression_target, class_target, fname, orig_image in zip(
images, regression_targets, class_targets, filename,
original_image):
tf_example = create_tf_example(image, regression_target,
class_target, fname, orig_image)
writer.write(tf_example.SerializeToString())
memory_used.append(psutil.virtual_memory().used / 2**30)
#plt.plot(memory_used)
#plt.title('Evolution of memory')
#plt.xlabel('iteration')
#plt.ylabel('memory used (GB)')
#plt.savefig(os.path.join(savedir, "memory.png"))
return written_files
def main(forest_object,
args=None,
input_type="fit_generator",
list_of_tfrecords=None,
comet_experiment=None):
"""
Main Training Loop
Args:
forest_object: a deepforest class object
args: Keras retinanet argparse
list_of_tfrecords: list of tfrecords to parse
input_type: "fit_generator" or "tfrecord" input type
"""
# parse arguments
if args is None:
args = sys.argv[1:]
args = parse_args(args)
# create object that stores backbone information
backbone = models.backbone(args.backbone)
# make sure keras is the minimum required version
check_keras_version()
# optionally choose specific GPU
if args.gpu:
setup_gpu(args.gpu)
# optionally load config parameters
if args.config:
args.config = read_config_file(args.config)
# data input
if input_type == "fit_generator":
# create the generators
train_generator, validation_generator = create_generators(
args, backbone.preprocess_image)
# placeholder target tensor for creating models
targets = None
elif input_type == "tfrecord":
# Create tensorflow iterators
iterator = tfrecords.create_dataset(list_of_tfrecords, args.batch_size)
next_element = iterator.get_next()
# Split into inputs and targets
inputs = next_element[0]
targets = [next_element[1], next_element[2]]
validation_generator = None
else:
raise ValueError(
"{} input type is invalid. Only 'tfrecord' or 'for_generator' "
"input types are accepted for model training".format(input_type))
# create the model
if args.snapshot is not None:
print('Loading model, this may take a second...')
model = models.load_model(args.snapshot, backbone_name=args.backbone)
training_model = model
anchor_params = None
if args.config and 'anchor_parameters' in args.config:
anchor_params = parse_anchor_parameters(args.config)
prediction_model = retinanet_bbox(model=model,
anchor_params=anchor_params)
else:
weights = args.weights
# default to imagenet if nothing else is specified
if weights is None and args.imagenet_weights:
weights = backbone.download_imagenet()
print('Creating model, this may take a second...')
if input_type == "fit_generator":
num_of_classes = train_generator.num_classes()
else:
# Add background class
num_of_classes = len(forest_object.labels.keys())
model, training_model, prediction_model = create_models(
backbone_retinanet=backbone.retinanet,
num_classes=num_of_classes,
weights=weights,
multi_gpu=args.multi_gpu,
freeze_backbone=args.freeze_backbone,
lr=args.lr,
config=args.config,
targets=targets,
freeze_layers=args.freeze_layers)
# print model summary
print(model.summary())
# this lets the generator compute backbone layer shapes using the actual backbone model
if 'vgg' in args.backbone or 'densenet' in args.backbone:
train_generator.compute_shapes = make_shapes_callback(model)
if validation_generator:
validation_generator.compute_shapes = train_generator.compute_shapes
# create the callbacks
callbacks = create_callbacks(model, training_model, prediction_model,
validation_generator, args, comet_experiment)
if not args.compute_val_loss:
validation_generator = None
# start training
if input_type == "fit_generator":
history = training_model.fit_generator(
generator=train_generator,
steps_per_epoch=args.steps,
epochs=args.epochs,
verbose=1,
callbacks=callbacks,
workers=args.workers,
use_multiprocessing=args.multiprocessing,
max_queue_size=args.max_queue_size,
validation_data=validation_generator)
elif input_type == "tfrecord":
# Fit model
history = training_model.fit(x=inputs,
steps_per_epoch=args.steps,
epochs=args.epochs,
callbacks=callbacks)
else:
raise ValueError(
"{} input type is invalid. Only 'tfrecord' or 'for_generator' "
"input types are accepted for model training".format(input_type))
# Assign history to deepforest model class
forest_object.history = history
# return trained model
return model, prediction_model, training_model