def do_debug(self, args=None):
if not args:
debug_check = spRun(['retinanet-debug', 'csv', self.PATH_TO_TRAIN_CSV ,self.PATH_TO_CLASSES_CSV])
try:
debug_check.check_returncode()
except:
print('From Tool: ATTENTION! \t retinanet model evaluating was failure! Next processes may not work properly. ')
else:
spRun((['retinanet-debug'] + args))
def do_build_dataset(self, args):
if not args:
spRun(['python', self.PATH_TO_BUILD_DATASET])
print('dataset succesfully builded')
self.do_debug()
else:
spRun((['python', self.PATH_TO_BUILD_DATASET] + args))
print('dataset succesfully builded')
self.do_debug()
def do_convert(self, args=None):
if not args:
convert_check = spRun(['retinanet-convert-model', self.PATH_TO_BEST_MODEL , self.PATH_TO_CONVERTED_MODEL])
try:
convert_check.check_returncode()
except:
print('From Tool: ATTENTION! \t retinanet model converting was failure! Next processes may not work properly. ')
else:
spRun((['retinanet-convert-model'] + args))
def do_train(self, args=None):
print('You can find the config file from mrcnn/own_dataset/own_config.py that includes default train arguments assigning.')
args = self.trainDefaultArgumentsSetter(args)
train_check = spRun((['python', self.PATH_TO_TRAIN_PY] + args))
try:
train_check.check_returncode()
except:
print('From Tool: ATTENTION! \t retinanet training was failure! Next processes may not work properly. ')
def do_train(self, args=None):
args = self.trainDefaultArgumentsSetter(args)
self.move_old_snapshots()
train_check = spRun((['retinanet-train'] + args))
try:
train_check.check_returncode()
except:
print('From Tool: ATTENTION! \t retinanet training was failure! Next processes may not work properly. ')
self.select_best_training_model()
self.do_convert()
self.do_evaluate()
def do_predict(self, args=None):
if not args:
spRun(['python', self.PATH_TO_PREDICT_PY])
else:
spRun((['python', self.PATH_TO_PREDICT_PY] + args))
def do_build_dataset(self, args):
if not args:
spRun(['python', self.PATH_TO_BUILD_DATASET])
else:
spRun((['python', self.PATH_TO_BUILD_DATASET] + args))
def do_test(self, args):
if not args:
spRun(['retinanet-test'])
else:
spRun((['retinanet-test'] + args))
def do_predict(self, args=None):
args = self.predictDefaultArgumentsSetter(args)
spRun((['python', 'retinanet/predict.py'] + args))
def main(args):
ap = argparse.ArgumentParser()
ap.add_argument(
"-m",
"--model",
default=0,
type=int,
choices=[0, 1, 2, 3, 4, 5],
help=("Model to be trained"
"\n0: ssd_mobilenet_v2_320x320_coco17_tpu-8 (DEFAULT)"
"\n1: efficientdet_d7_coco17_tpu-32"
"\n2: centernet_resnet50_v1_fpn_512x512_coco17_tpu-8"
"\n3: mask_rcnn_inception_resnet_v2_1024x1024_coco17_gpu-8"
"\n4: ssd_resnet101_v1_fpn_1024x1024_coco17_tpu-8"
"\n5: faster_rcnn_resnet152_v1_800x1333_coco17_gpu-8"))
ap.add_argument(
"-b",
"--batch_size",
type=int,
help=
'optional batch_size parameter for training. (Higher values require more memory and vice-versa)'
)
ap.add_argument(
"-d",
"--dont_use_checkpoint",
action='store_false',
help=
'If this parameter is passed. Training process does not use pre-trained weight, it starts from scratch'
)
args = vars(ap.parse_args())
model_index = args["model"]
batch_size = args["batch_size"]
dont_use_checkpoint = args["dont_use_checkpoint"]
models = [
'ssd_mobilenet_v2_320x320_coco17_tpu-8',
'efficientdet_d7_coco17_tpu-32',
'centernet_resnet50_v1_fpn_512x512_coco17_tpu-8',
'mask_rcnn_inception_resnet_v2_1024x1024_coco17_gpu-8',
'ssd_resnet101_v1_fpn_1024x1024_coco17_tpu-8',
'faster_rcnn_resnet152_v1_800x1333_coco17_gpu-8'
]
model_name = models[model_index]
pre_trained_models_dir = os.path.abspath(
os.path.join(ROOT_DIR, 'pre-trained-models'))
pre_trained_model = get_pre_trained_model(model_name,
pre_trained_models_dir)
config_file_new = edit_config(model_name, pre_trained_models_dir,
pre_trained_model, batch_size,
dont_use_checkpoint)
### call model_main.py with args for training
now = datetime.now()
dt_string = now.strftime("%d-%m-%Y_%H-%M-%S")
model_dir = os.path.abspath(os.path.join(ROOT_DIR, 'training', dt_string))
os.makedirs(os.path.abspath(os.path.join(model_dir, 'export', 'Servo')))
args_for_model_main = [
'--alsologtostderr', '--model_dir=' + model_dir,
('--pipeline_config_path=' + config_file_new)
]
spRun((['python', PATH_TO_MODEL_MAIN_PY] + args_for_model_main))
print('\n\nmodel_main_tf_2.py is Done...\n\n')
### call export_inference_graph.py with args to prepare model for prediction
model_ckpt = os.path.abspath(os.path.join(model_dir))
inference_graph_dir = os.path.abspath(
os.path.join(ROOT_DIR, 'trained-inference-graphs', dt_string))
os.makedirs(inference_graph_dir)
inference_graph_path = os.path.abspath(os.path.join(inference_graph_dir))
args_for_inference_graph = [
'--input_type=image_tensor',
('--pipeline_config_path=' + config_file_new),
('--trained_checkpoint_dir=' + model_ckpt),
('--output_directory=' + inference_graph_path)
]
print('args_for_inference_graph:', args_for_inference_graph)
spRun((['python', PATH_TO_EXPORT_INFERENCE_GRAPHPY] +
args_for_inference_graph))
label_map_src = os.path.abspath(
os.path.join(DATASET_DIR, 'train', 'ssd_label_map.pbtxt'))
label_map_dst = os.path.abspath(
os.path.join(ROOT_DIR, 'trained-inference-graphs', dt_string,
'ssd_label_map.pbtxt'))
copy2(label_map_src, label_map_dst)
num_classes_src = os.path.abspath(
os.path.join(DATASET_DIR, 'train', 'ssd_num_classes.txt'))
num_classes_dst = os.path.abspath(
os.path.join(ROOT_DIR, 'trained-inference-graphs', dt_string,
'ssd_num_classes.txt'))
copy2(num_classes_src, num_classes_dst)
def main(args):
ap = argparse.ArgumentParser()
ap.add_argument(
"-c",
"--configFile",
default=0,
type=int,
choices=[0, 1, 2, 3],
help=
("A .config file contains train-evaluate-prediction configurations based selected backbone"
"\n0: ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync.config (DEFAULT)"
"\n1: ssd_resnet101_v1_fpn_shared_box_predictor_oid_512x512_sync.config"
"\n2: ssd_mobilenet_v1_ppn_shared_box_predictor_300x300_coco14_sync.config"
))
ap.add_argument(
"-b",
"--batch_size",
type=int,
help=
'optional batch_size parameter for training. (Higher values require more memory and vice-versa)'
)
ap.add_argument(
"-d",
"--dont_use_checkpoint",
action='store_false',
help=
'If this parameter is passed. Training process does not use pre-trained weight, it starts from scratch'
)
args = vars(ap.parse_args())
config_file = args["configFile"]
batch_size = args["batch_size"]
dont_use_checkpoint = args["dont_use_checkpoint"]
config_file_new, num_steps_train = edit_config(config_file, batch_size,
dont_use_checkpoint)
### call model_main.py with args for training
now = datetime.now()
dt_string = now.strftime("%d-%m-%Y_%H-%M-%S")
model_dir = os.path.abspath(os.path.join(ROOT_DIR, 'training', dt_string))
os.makedirs(os.path.abspath(os.path.join(model_dir, 'export', 'Servo')))
model_dir_str = '\"' + str(model_dir) + '/\"'
config_file_new_str = '\"' + str(config_file_new) + '\"'
args_for_model_main = [
'--alsologtostderr', ' --model_dir=' + model_dir_str,
(' --pipeline_config_path=' + config_file_new_str)
]
spRun(['python', PATH_TO_MODEL_MAIN_PY, args_for_model_main])
### call export_inference_graph.py with args to prepare model for prediction
model_ckpt = os.path.abspath(
os.path.join(model_dir, ('model.ckpt-' + str(num_steps_train))))
inference_graph_dir = os.path.abspath(
os.path.join(ROOT_DIR, 'trained-inference-graphs', dt_string))
os.makedirs(inference_graph_dir)
inference_graph_path = os.path.abspath(
os.path.join(inference_graph_dir, 'output_inference_graph_v1.pb'))
inference_graph_path_str = '\"' + str(inference_graph_path) + '\"'
model_ckpt_str = '\"' + str(model_ckpt) + '\"'
args_for_inference_graph = [
'--input_type=image_tensor',
(' --pipeline_config_path=' + config_file_new_str),
(' --trained_checkpoint_prefix=' + model_ckpt_str),
(' --output_directory=' + inference_graph_path_str)
]
spRun(
['python', PATH_TO_EXPORT_INFERENCE_GRAPHPY, args_for_inference_graph])
label_map_src = os.path.abspath(
os.path.join(DATASET_DIR, 'train', 'ssd_label_map.pbtxt'))
label_map_dst = os.path.abspath(
os.path.join(ROOT_DIR, 'trained-inference-graphs', dt_string,
'ssd_label_map.pbtxt'))
copy2(label_map_src, label_map_dst)
num_classes_src = os.path.abspath(
os.path.join(DATASET_DIR, 'train', 'ssd_num_classes.txt'))
num_classes_dst = os.path.abspath(
os.path.join(ROOT_DIR, 'trained-inference-graphs', dt_string,
'ssd_num_classes.txt'))
copy2(num_classes_src, num_classes_dst)
