def main(args):
# Grid search
model_output = []
for batch_size in args.batch_size:
logger.info("Generating data with batch of %s images...", batch_size)
# Data generator building
prepro_folder = utils.prepare_preprocessed_folder(
args.datapath, args.dataset, args.image_size)
nb_labels, train_gen, valid_gen = get_data(
prepro_folder,
args.dataset,
args.model,
args.image_size,
batch_size,
)
for parameters in itertools.product(
args.dropout,
args.network,
args.learning_rate,
args.learning_rate_decay,
):
logger.info("Instance: %s", utils.list_to_str(parameters))
# Data path and repository management
dropout, network, learning_rate, learning_rate_decay = parameters
instance_args = [
args.name,
args.image_size,
network,
batch_size,
dropout,
learning_rate,
learning_rate_decay,
]
instance_name = utils.list_to_str(instance_args, "_")
output_folder = utils.prepare_output_folder(
args.datapath, args.dataset, args.model, instance_name)
# Model running
model_output.append(
run_model(train_gen, valid_gen, args.model, output_folder,
instance_name, args.image_size, nb_labels,
args.nb_epochs, args.nb_training_image,
args.nb_validation_image, batch_size, *parameters))
logger.info("Instance result: %s", model_output[-1])
# Recover best instance starting from validation accuracy
best_instance = max(model_output, key=lambda x: x["val_acc"])
# Save best model
output_folder = utils.prepare_output_folder(args.datapath, args.dataset,
args.model)
instance_name = os.path.join(
output_folder,
"best-{}-" + str(args.image_size) + ".{}",
)
best_instance["model"].save(instance_name.format("model", "h5"))
with open(instance_name.format("instance", "json"), "w") as fobj:
json.dump(
{
key: best_instance[key]
for key in best_instance if key != "model"
},
fobj,
)
backend.clear_session()
if __name__=='__main__':
# Parse command-line arguments
parser = argparse.ArgumentParser(description=("Convolutional Neural Netw"
"ork on street-scene images"))
parser = add_instance_arguments(parser)
parser = add_hyperparameters(parser)
parser = add_training_arguments(parser)
args = parser.parse_args()
# Data path and repository management
aggregate_value = "full" if not args.aggregate_label else "aggregated"
instance_args = [args.name, args.image_size, args.network, args.batch_size,
aggregate_value, args.dropout,
args.learning_rate, args.learning_rate_decay]
instance_name = utils.list_to_str(instance_args, "_")
prepro_folder = utils.prepare_preprocessed_folder(args.datapath, args.dataset,
args.image_size,
aggregate_value)
# Data gathering
if (os.path.isfile(prepro_folder["training_config"]) and os.path.isfile(prepro_folder["validation_config"])
and os.path.isfile(prepro_folder["testing_config"])):
train_config = utils.read_config(prepro_folder["training_config"])
label_ids = [x['id'] for x in train_config['labels'] if x['is_evaluate']]
train_generator = generator.create_generator(
args.dataset,
args.model,
prepro_folder["training"],
args.image_size,
args.batch_size,
def predict(
filenames,
dataset,
problem,
datapath="./data",
aggregate=False,
name=None,
network=None,
batch_size=None,
dropout=None,
learning_rate=None,
learning_rate_decay=None,
output_dir="/tmp/deeposlandia/predicted",
):
"""Make label prediction on image indicated by ̀filename`, according to
considered `problem`
Parameters
----------
filenames : str
Name of the image files on the file system
dataset : str
Name of the dataset
problem : str
Name of the considered model, either `feature_detection` or
`semantic_segmentation`
datapath : str
Relative path of dataset repository
aggregate : bool
Either or not the labels are aggregated
name : str
Name of the saved network
network : str
Name of the chosen architecture, either `simple`, `vgg` or `inception`
batch_size : integer
Batch size used for training the model
dropout : float
Dropout rate used for training the model
learning_rate : float
Learning rate used for training the model
learning_rate_decay : float
Learning rate decay used for training the model
output_dir : str
Path of the output directory, where labelled images will be stored
(useful only if `problem=semantic_segmentation`)
Returns
-------
dict
Double predictions (between 0 and 1, acts as percentages) regarding
each labels
"""
# `image_paths` is first got as
# [[image1, ..., image_i], [image_j, ..., image_n]]
image_paths = [glob.glob(f) for f in filenames]
# then it is flattened to get a simple list
flattened_image_paths = sum(image_paths, [])
images = extract_images(flattened_image_paths)
model_input_size = images.shape[1]
if dataset == "aerial":
tile_size = utils.get_tile_size_from_image(model_input_size)
else:
tile_size = model_input_size
aggregate_value = "full" if not aggregate else "aggregated"
instance_args = [
name,
tile_size,
network,
batch_size,
aggregate_value,
dropout,
learning_rate,
learning_rate_decay,
]
instance_name = utils.list_to_str(instance_args, "_")
prepro_folder = utils.prepare_preprocessed_folder(
datapath, dataset, tile_size, aggregate_value
)
if os.path.isfile(prepro_folder["training_config"]):
train_config = utils.read_config(prepro_folder["training_config"])
label_ids = [
x["id"] for x in train_config["labels"] if x["is_evaluate"]
]
nb_labels = len(label_ids)
else:
logger.error(
(
"There is no training data with the given "
"parameters. Please generate a valid dataset "
"before calling the program."
)
)
sys.exit(1)
if any([arg is None for arg in instance_args]):
logger.info(
("Some arguments are None, " "the best model is considered.")
)
output_folder = utils.prepare_output_folder(datapath, dataset, problem)
instance_filename = (
"best-instance-" + str(tile_size) + "-" + aggregate_value + ".json"
)
instance_path = os.path.join(output_folder, instance_filename)
dropout, network = utils.recover_instance(instance_path)
model = init_model(
problem,
instance_name,
model_input_size,
nb_labels,
dropout,
network,
)
checkpoint_filename = (
"best-model-" + str(tile_size) + "-" + aggregate_value + ".h5"
)
checkpoint_full_path = os.path.join(output_folder, checkpoint_filename)
if os.path.isfile(checkpoint_full_path):
logger.info("Checkpoint full path : %s", checkpoint_full_path)
model.load_weights(checkpoint_full_path)
logger.info(
"Model weights have been recovered from %s",
checkpoint_full_path,
)
else:
logger.info(
(
"No available trained model for this image size"
" with optimized hyperparameters. The "
"inference will be done on an untrained model"
)
)
else:
logger.info("All instance arguments are filled out.")
output_folder = utils.prepare_output_folder(
datapath, dataset, problem, instance_name
)
model = init_model(
problem,
instance_name,
model_input_size,
nb_labels,
dropout,
network,
)
checkpoints = [
item
for item in os.listdir(output_folder)
if "checkpoint-epoch" in item
]
if len(checkpoints) > 0:
model_checkpoint = max(checkpoints)
checkpoint_full_path = os.path.join(
output_folder, model_checkpoint
)
model.load_weights(checkpoint_full_path)
logger.info(
"Model weights have been recovered from %s",
checkpoint_full_path,
)
else:
logger.info(
(
"No available checkpoint for this configuration. "
"The model will be trained from scratch."
)
)
y_raw_pred = model.predict(images)
result = {}
if problem == "feature_detection":
label_info = [
(i["category"], utils.GetHTMLColor(i["color"]))
for i in train_config["labels"]
]
for filename, prediction in zip(flattened_image_paths, y_raw_pred):
result[filename] = [
(i[0], 100 * round(float(j), 2), i[1])
for i, j in zip(label_info, prediction)
]
return result
elif problem == "semantic_segmentation":
os.makedirs(output_dir, exist_ok=True)
predicted_labels = np.argmax(y_raw_pred, axis=3)
encountered_labels = np.unique(predicted_labels)
meaningful_labels = [
x
for i, x in enumerate(train_config["labels"])
if i in encountered_labels
]
labelled_images = np.zeros(
shape=np.append(predicted_labels.shape, 3), dtype=np.int8
)
for i in range(nb_labels):
labelled_images[predicted_labels == i] = train_config["labels"][i][
"color"
]
for predicted_labels, filename in zip(
labelled_images, flattened_image_paths
):
predicted_image = Image.fromarray(predicted_labels, "RGB")
filename = filename.replace(".jpg", ".png")
predicted_image_path = os.path.join(
output_dir, os.path.basename(filename)
)
predicted_image.save(predicted_image_path)
result[filename] = os.path.basename(filename)
return {
"labels": summarize_config(meaningful_labels),
"label_images": result,
}
else:
logger.error(
(
"Unknown model argument. Please use "
"'feature_detection' or 'semantic_segmentation'."
)
)
sys.exit(1)
def predict(
filenames,
dataset,
problem,
datapath="./data",
name=None,
network=None,
batch_size=None,
dropout=None,
learning_rate=None,
learning_rate_decay=None,
output_dir="/tmp/deeposlandia/predicted",
):
"""Make label prediction on image indicated by ̀filename`, according to
considered `problem`
Parameters
----------
filenames : str
Name of the image files on the file system
dataset : str
Name of the dataset
problem : str
Name of the considered model, either `featdet` or `semseg`
datapath : str
Relative path of dataset repository
name : str
Name of the saved network
network : str
Name of the chosen architecture, either `simple`, `vgg` or `inception`
batch_size : integer
Batch size used for training the model
dropout : float
Dropout rate used for training the model
learning_rate : float
Learning rate used for training the model
learning_rate_decay : float
Learning rate decay used for training the model
output_dir : str
Path of the output directory, where labelled images will be stored
(useful only if `problem=semantic_segmentation`)
Returns
-------
dict
Double predictions (between 0 and 1, acts as percentages) regarding
each labels
"""
# `image_paths` is first got as
# [[image1, ..., image_i], [image_j, ..., image_n]]
image_paths = [glob.glob(f) for f in filenames]
# then it is flattened to get a simple list
flattened_image_paths = sum(image_paths, [])
images = extract_images(flattened_image_paths)
model_input_size = images.shape[1]
instance_args = [
name,
model_input_size,
network,
batch_size,
dropout,
learning_rate,
learning_rate_decay,
]
instance_name = utils.list_to_str(instance_args, "_")
prepro_folder = utils.prepare_preprocessed_folder(datapath, dataset,
model_input_size)
if os.path.isfile(prepro_folder["training_config"]):
train_config = utils.read_config(prepro_folder["training_config"])
label_ids = [
x["id"] for x in train_config["labels"] if x["is_evaluate"]
]
nb_labels = len(label_ids)
else:
raise FileNotFoundError(
"There is no training data with the given parameters. "
"Please generate a valid dataset before calling the program.")
output_folder = utils.prepare_output_folder(datapath, dataset,
model_input_size, problem)
instance_path = os.path.join(output_folder, output_folder["best-instance"])
dropout, network = utils.recover_instance(instance_path)
model = init_model(
problem,
instance_name,
model_input_size,
nb_labels,
dropout,
network,
)
if os.path.isfile(output_folder["best-model"]):
model.load_weights(output_folder["best-model"])
logger.info(
"Model weights have been recovered from %s",
output_folder["best-model"],
)
else:
logger.info(
"No available trained model for this image size with optimized hyperparameters. "
"The inference will be done on an untrained model")
y_raw_pred = model.predict(images, batch_size=2, verbose=1)
result = {}
if problem == "featdet":
label_info = [(i["category"], utils.GetHTMLColor(i["color"]))
for i in train_config["labels"]]
for filename, prediction in zip(flattened_image_paths, y_raw_pred):
result[filename] = [(i[0], 100 * round(float(j), 2), i[1])
for i, j in zip(label_info, prediction)]
return result
elif problem == "semseg":
os.makedirs(output_dir, exist_ok=True)
predicted_labels = np.argmax(y_raw_pred, axis=3)
encountered_labels = np.unique(predicted_labels)
meaningful_labels = [
x for i, x in enumerate(train_config["labels"])
if i in encountered_labels
]
labelled_images = np.zeros(shape=np.append(predicted_labels.shape, 3),
dtype=np.int8)
for i in range(nb_labels):
labelled_images[predicted_labels ==
i] = train_config["labels"][i]["color"]
for predicted_labels, filename in zip(labelled_images,
flattened_image_paths):
predicted_image = Image.fromarray(predicted_labels, "RGB")
filename = filename.replace(".jpg", ".png")
predicted_image_path = os.path.join(output_dir,
os.path.basename(filename))
predicted_image.save(predicted_image_path)
result[filename] = os.path.basename(filename)
return {
"labels": summarize_config(meaningful_labels),
"label_images": result,
}
else:
raise ValueError(
"Unknown model argument. Please use 'featdet' or 'semseg'.")
# Grid search
model_output = []
for batch_size in args.batch_size:
logger.info("Generating data with batch of %s images...", batch_size)
# Data generator building
prepro_folder = utils.prepare_preprocessed_folder(
args.datapath, args.dataset, args.image_size, aggregate_value)
nb_labels, train_gen, valid_gen = get_data(prepro_folder, args.dataset,
args.model,
model_input_size,
batch_size)
for parameters in itertools.product(args.dropout, args.network,
args.learning_rate,
args.learning_rate_decay):
logger.info(utils.list_to_str(parameters))
# Data path and repository management
dropout, network, learning_rate, learning_rate_decay = parameters
instance_args = [
args.name, args.image_size, network, batch_size,
aggregate_value, dropout, learning_rate, learning_rate_decay
]
instance_name = utils.list_to_str(instance_args, "_")
output_folder = utils.prepare_output_folder(
args.datapath, args.dataset, args.model, instance_name)
# Model running
model_output.append(
run_model(train_gen, valid_gen, args.model, output_folder,
instance_name, model_input_size, aggregate_value,
nb_labels, args.nb_epochs, args.nb_training_image,
args.nb_validation_image, batch_size, *parameters))