def MiDaS_small(pretrained=True, **kwargs):
""" # This docstring shows up in hub.help()
MiDaS small model for monocular depth estimation on resource-constrained devices
pretrained (bool): load pretrained weights into model
"""
model = MidasNet_small(None,
features=64,
backbone="efficientnet_lite3",
exportable=True,
non_negative=True,
blocks={'expand': True})
if pretrained:
checkpoint = (
"https://github.com/intel-isl/MiDaS/releases/download/v2_1/model-small-70d6b9c8.pt"
)
state_dict = torch.hub.load_state_dict_from_url(
checkpoint,
map_location=torch.device('cpu'),
progress=True,
check_hash=True)
model.load_state_dict(state_dict)
return model
def __init__(self, model_type, model_path, optimize):
print("initialize")
# select device
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device: %s" % self.device)
# load network
if model_type == "large":
self.model = MidasNet(model_path, non_negative=True)
self.net_w, self.net_h = 384, 384
elif model_type == "small":
self.model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True,
non_negative=True, blocks={'expand': True})
self.net_w, self.net_h = 256, 256
else:
print(f"model_type '{model_type}' not implemented, use: --model_type large")
assert False
self.transform = Compose(
[
Resize(
self.net_w,
self.net_h,
resize_target=None,
keep_aspect_ratio=True,
ensure_multiple_of=32,
resize_method="upper_bound",
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
PrepareForNet(),
]
)
self.model.eval()
self.optimize = optimize
if self.optimize:
rand_example = torch.rand(1, 3, self.net_h, self.net_w)
self.model(rand_example)
traced_script_module = torch.jit.trace(self.model, rand_example)
self.model = traced_script_module
if self.device == torch.device("cuda"):
self.model = self.model.to(memory_format=torch.channels_last)
self.model = self.model.half()
self.model.to(self.device)
def init_model(transform):
parser = argparse.ArgumentParser()
parser.add_argument('-mw', '--model_weights',
default='model-f6b98070.pt',
help='path to the trained weights of model'
)
parser.add_argument('-mt', '--model_type',
default='large',
help='model type: large or small'
)
parser.add_argument('--optimize', dest='optimize', action='store_true')
parser.add_argument('--no-optimize', dest='optimize', action='store_false')
parser.set_defaults(optimize=True)
args, unknown = parser.parse_known_args()
# set torch options
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
print("initialize")
# select device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device: %s" % device)
# load network
if args.model_type == "large":
model_path = "../MiDaS/"+args.model_weights
model = MidasNet(model_path, non_negative=True)
net_w, net_h = 384, 384
elif args.model_type == "small":
if "small" not in args.model_weights:
args.model_weights = "model-small-70d6b9c8.pt"
model_path = "../MiDaS/"+args.model_weights
model = MidasNet_small(model_path, features=64, backbone="efficientnet_lite3", exportable=True, non_negative=True, blocks={'expand': True})
net_w, net_h = 256, 256
else:
print(f"model_type '{model_type}' not implemented, use: --model_type large")
assert False
transform = Compose(
[
Resize(
net_w,
net_h,
resize_target=None,
keep_aspect_ratio=True,
ensure_multiple_of=32,
resize_method="upper_bound",
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
PrepareForNet(),
]
)
model.eval()
if args.optimize==True:
rand_example = torch.rand(1, 3, net_h, net_w)
model(rand_example)
traced_script_module = torch.jit.trace(model, rand_example)
model = traced_script_module
if device == torch.device("cuda"):
model = model.to(memory_format=torch.channels_last)
model = model.half()
model.to(device)
return (model, transform, device, args.optimize), args
def run(input_path,
output_path,
model_path,
model_type="large",
optimize=True):
"""Run MonoDepthNN to compute depth maps.
Args:
input_path (str): path to input folder
output_path (str): path to output folder
model_path (str): path to saved model
"""
print("initialize")
# select device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device: %s" % device)
# load network
if model_type == "large":
model = MidasNet(model_path, non_negative=True)
net_w, net_h = 384, 384
elif model_type == "small":
model = MidasNet_small(model_path,
features=64,
backbone="efficientnet_lite3",
exportable=True,
non_negative=True,
blocks={'expand': True})
net_w, net_h = 256, 256
else:
print(
f"model_type '{model_type}' not implemented, use: --model_type large"
)
assert False
transform = Compose([
Resize(
net_w,
net_h,
resize_target=None,
keep_aspect_ratio=True,
ensure_multiple_of=32,
resize_method="upper_bound",
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
PrepareForNet(),
])
model.eval()
if optimize == True:
rand_example = torch.rand(1, 3, net_h, net_w)
model(rand_example)
traced_script_module = torch.jit.trace(model, rand_example)
model = traced_script_module
if device == torch.device("cuda"):
model = model.to(memory_format=torch.channels_last)
model = model.half()
model.to(device)
# get input
img_names = glob.glob(os.path.join(input_path, "*"))
num_images = len(img_names)
# create output folder
os.makedirs(output_path, exist_ok=True)
print("start processing")
for ind, img_name in enumerate(img_names):
print(" processing {} ({}/{})".format(img_name, ind + 1, num_images))
# input
img = utils.read_image(img_name)
img_input = transform({"image": img})["image"]
# compute
with torch.no_grad():
sample = torch.from_numpy(img_input).to(device).unsqueeze(0)
if optimize == True and device == torch.device("cuda"):
sample = sample.to(memory_format=torch.channels_last)
sample = sample.half()
prediction = model.forward(sample)
prediction = (torch.nn.functional.interpolate(
prediction.unsqueeze(1),
size=img.shape[:2],
mode="bicubic",
align_corners=False,
).squeeze().cpu().numpy())
prediction /= 1000
# output
filename = os.path.join(
output_path,
os.path.splitext(os.path.basename(img_name))[0])
utils.write_depth(filename, prediction, bits=2)
print(prediction)
print(prediction.shape)
print("finished")
def __init__(self):
self.device = "cuda" if torch.cuda.is_available() else "cpu"
# Setup AdaBins model
self.adabins_nyu_infer_helper = InferenceHelper(dataset='nyu',
device=self.device)
self.adabins_kitti_infer_helper = InferenceHelper(dataset='kitti',
device=self.device)
# Setup DiverseDepth model
class DiverseDepthArgs:
def __init__(self):
self.resume = False
self.cfg_file = "lib/configs/resnext50_32x4d_diversedepth_regression_vircam"
self.load_ckpt = "pretrained/DiverseDepth.pth"
diverse_depth_args = DiverseDepthArgs()
merge_cfg_from_file(diverse_depth_args)
self.diverse_depth_model = RelDepthModel()
self.diverse_depth_model.eval()
# load checkpoint
load_ckpt(diverse_depth_args, self.diverse_depth_model)
# TODO: update this - see how `device` argument should be processsed
if self.device == "cuda":
self.diverse_depth_model.cuda()
self.diverse_depth_model = torch.nn.DataParallel(
self.diverse_depth_model)
# Setup MiDaS model
self.midas_model_path = "./pretrained/MiDaS_f6b98070.pt"
midas_model_type = "large"
# load network
if midas_model_type == "large":
self.midas_model = MidasNet(self.midas_model_path,
non_negative=True)
self.midas_net_w, self.midas_net_h = 384, 384
elif midas_model_type == "small":
self.midas_model = MidasNet_small(self.midas_model_path,
features=64,
backbone="efficientnet_lite3",
exportable=True,
non_negative=True,
blocks={'expand': True})
self.midas_net_w, self.midas_net_h = 256, 256
self.midas_transform = Compose([
Resize(
self.midas_net_w,
self.midas_net_h,
resize_target=None,
keep_aspect_ratio=True,
ensure_multiple_of=32,
resize_method="upper_bound",
image_interpolation_method=cv2.INTER_CUBIC,
),
NormalizeImage(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
PrepareForNet(),
])
self.midas_model.eval()
self.midas_optimize = True
if self.midas_optimize == True:
rand_example = torch.rand(1, 3, self.midas_net_h, self.midas_net_w)
self.midas_model(rand_example)
traced_script_module = torch.jit.trace(self.midas_model,
rand_example)
self.midas_model = traced_script_module
if self.device == "cuda":
self.midas_model = self.midas_model.to(
memory_format=torch.channels_last)
self.midas_model = self.midas_model.half()
self.midas_model.to(torch.device(self.device))
# Setup SGDepth model
self.sgdepth_model = InferenceEngine.SgDepthInference()
# Setup monodepth2 model
self.monodepth2_model_path = "pretrained/monodepth2_mono+stereo_640x192"
monodepth2_device = torch.device(self.device)
encoder_path = os.path.join(self.monodepth2_model_path, "encoder.pth")
depth_decoder_path = os.path.join(self.monodepth2_model_path,
"depth.pth")
# LOADING PRETRAINED MODEL
print(" Loading Monodepth2 pretrained encoder")
self.monodepth2_encoder = networks.ResnetEncoder(18, False)
loaded_dict_enc = torch.load(encoder_path,
map_location=monodepth2_device)
# extract the height and width of image that this model was trained with
self.feed_height = loaded_dict_enc['height']
self.feed_width = loaded_dict_enc['width']
filtered_dict_enc = {
k: v
for k, v in loaded_dict_enc.items()
if k in self.monodepth2_encoder.state_dict()
}
self.monodepth2_encoder.load_state_dict(filtered_dict_enc)
self.monodepth2_encoder.to(monodepth2_device)
self.monodepth2_encoder.eval()
print(" Loading pretrained decoder")
self.monodepth2_depth_decoder = networks.DepthDecoder(
num_ch_enc=self.monodepth2_encoder.num_ch_enc, scales=range(4))
loaded_dict = torch.load(depth_decoder_path,
map_location=monodepth2_device)
self.monodepth2_depth_decoder.load_state_dict(loaded_dict)
self.monodepth2_depth_decoder.to(monodepth2_device)
self.monodepth2_depth_decoder.eval()