def parse(self, args=None, namespace=None):
if not self.initialized:
self.initialize()
self.params = self.parser.parse_args(args, namespace=namespace)
if self.params.configure == "kitti":
self.params.flow_checkpoint = "FlowNet2-KITTI"
self.params.model_type = "monodepth2"
self.params.overlap_ratio = 0.5
if 'matcher' in self.params:
self.params.matcher = 'sequential'
# Resolve unspecified parameters
model = get_depth_model(self.params.model_type)
if self.params.align <= 0:
self.params.align = model.align
if self.params.learning_rate <= 0:
self.params.learning_rate = model.learning_rate
if self.params.lambda_view_baseline < 0:
self.params.lambda_view_baseline = model.lambda_view_baseline
self.print_p()
return self.params
def __init__(self, range_dir, frames, params):
self.frames = frames
self.params = params
self.base_dir = params.path
self.range_dir = range_dir
self.out_dir = pjoin(self.range_dir, make_tag(params))
os.makedirs(self.out_dir, exist_ok=True)
print(f"Fine-tuning directory: '{self.out_dir}'")
# checkpoints文件路径
self.checkpoints_dir = pjoin(self.out_dir, "checkpoints")
os.makedirs(self.checkpoints_dir, exist_ok=True)
# 查询生成的模型
model = get_depth_model(params.model_type)
self.model = model()
# 查询存在GPU
num_gpus = torch.cuda.device_count()
print(f"Using {num_gpus} GPUs.")
if num_gpus > 1:
self.params.batch_size *= num_gpus
print(f"Adjusting batch size to {self.params.batch_size}.")
self.reference_disparity = {}
self.vis_depth_scale = None