def setup(config):
mde = ex.get_and_configure(config['mde'])
# if mde == 'DORN':
# img_key = 'dorn_image'
# in_type = 'torch'
# in_order = 'nchw'
# out_type = 'torch'
# out_order = 'nchw'
# elif mde == 'DenseDepth':
# img_key = 'image'
# in_type = 'numpy'
# in_order = 'nhwc'
# out_type = 'numpy'
# out_order = 'nhwc'
# elif mde == 'MiDaS':
# img_key = 'midas_image'
# in_type = 'torch'
# in_order = 'nhwc'
# out_type = 'numpy'
# out_order = 'nhwc'
return MDE(mde,
key=config['img_key'],
in_type=config['in_type'],
in_order=config['in_order'],
out_type=config['out_type'],
out_order=config['out_order'])
def setup(config):
scene_data = np.load(IN_DATA_DIR/f'{config["scene"]}.npy', allow_pickle=True)[()]
model = ex.get_and_configure(config['method'])
transform = Compose([ex.transforms[t] for t in config['transform']])
if config['gpu'] is not None and torch.cuda.is_available():
os.environ['CUDA_VISIBLE_DEVICES'] = config['gpu']
device = torch.device('cuda')
print(f"Using gpu {config['gpu']} (CUDA_VISIBLE_DEVICES = {os.environ['CUDA_VISIBLE_DEVICES']}).")
else:
device = torch.device('cpu')
print("Using cpu.")
return CapturedEvaluation(model=model, scene_data=scene_data, transform=transform, device=device)
def setup(config):
model = ex.get_and_configure(config['method'])
if config['method'] in ['mde', 'median', 'gt_hist']:
dataset = NYUDepthv2(split=config['split'])
else:
dataset = NYUDepthv2Transient(split=config['split'], sbr=config['sbr'])
transform = Compose([ex.transforms[t] for t in config['transform']])
dataset.transform = transform
if config['gpu'] is not None and torch.cuda.is_available():
os.environ['CUDA_VISIBLE_DEVICES'] = config['gpu']
device = torch.device('cuda')
print(f"Using gpu {config['gpu']} (CUDA_VISIBLE_DEVICES = {os.environ['CUDA_VISIBLE_DEVICES']}).")
else:
device = torch.device('cpu')
print("Using cpu.")
return NYUv2Evaluation(model=model,
dataset=dataset,
device=device,
pre_cropped=config['pre_cropped'])
def setup(config):
mde_model = ex.get_and_configure('mde')
preproc = TransientPreprocessor(config['source_n_sid_bins'],
config['n_ambient_bins'],
config['edge_coeff'], config['n_std'])
if config['refl_est'] == 'gray':
refl_est = refl_est_gray
elif config['refl_est'] == 'red':
refl_est = refl_est_red
source_disc = SI(config['source_n_sid_bins'], config['source_alpha'],
config['source_beta'], config['source_offset'])
crop = config['crop'] if not config['crop'] == [0, 0, 0, 0] else None
return MDETransient(mde_model=mde_model,
preproc=preproc,
refl_est=refl_est,
source_disc=source_disc,
min_depth=config['min_depth'],
max_depth=config['max_depth'],
crop=crop,
radial=config['radial'],
fc=config['fc'],
image_key=config['image_key'],
transient_key=config['transient_key'])
self.device = device
def evaluate(self):
# Load preprocessed data:
data = self.transform(self.scene_data)
data = default_collate([data])
for k, v in data.items():
data[k] = v.to(self.device)
depth = self.model(data)
pred = {'depth': depth}
pred['metrics'] = self.compute_metrics(data, pred)
return pred
def compute_metrics(self, data, pred):
p = pred['depth'].cpu().squeeze()
d = data['depth'].cpu().squeeze()
m = data['mask'].cpu().squeeze()
metrics = get_depth_metrics(p, d, m)
return metrics
if __name__ == '__main__':
evaluator = ex.get_and_configure('CapturedEvaluation')
pred = evaluator.evaluate()
pred['depth'] = pred['depth'].numpy()
scene = ex.config['scene']
mde_name = ex.config['mde']
output_dir = Path(ex.config['output_dir'])/f'{scene}'/f'{mde_name}'
output_dir.mkdir(parents=True, exist_ok=True)
np.save(output_dir/f'{ex.config["method"]}.npy', pred)
def setup(config):
mde_model = ex.get_and_configure('mde')
return MDEMedian(mde_model, gt_key=config['median_gt_key'])
@ex.setup('median')
def setup(config):
mde_model = ex.get_and_configure('mde')
return MDEMedian(mde_model, gt_key=config['median_gt_key'])
@ex.entity
class MDEMedian:
def __init__(self, mde_model, gt_key, crop=NYUV2_CROP):
self.mde_model = mde_model
self.gt_key = gt_key
self.crop = crop
def __call__(self, data):
init = self.mde_model(data)
if self.crop is not None:
# Set median based on crop (makes a big difference)
init_median = torch.median(init[..., self.crop[0]:self.crop[1],
self.crop[2]:self.crop[3]])
else:
init_median = torch.median(init)
init_median = init_median.to(data[self.gt_key].device)
init = init.to(data[self.gt_key].device)
pred = init * (torch.median(data[self.gt_key]) / init_median)
return pred
if __name__ == '__main__':
from pdb import set_trace
model = ex.get_and_configure('MDEMedian')
self.in_order = in_order # Input channel location required by mde
self.out_type = out_type # Output type of mde
self.out_order = out_order # Output channel location of mde
def __call__(self, data):
"""
Wraps the mde so that it takes torch input and produces numpy output
image_tensor is a torch float tensor in [0, 1] in NCHW format from the torch dataloader
out is a torch tensor in NCHW format
"""
img = data[self.key]
if self.in_order == 'nhwc':
img = img.permute(0, 2, 3, 1)
if self.in_type == 'numpy':
img = img.cpu().numpy()
out = self.mde(img)
if self.out_type == 'numpy':
out = torch.from_numpy(out)
if self.out_order == 'nhwc':
out = out.permute(0, 3, 1, 2)
return out
if __name__ == '__main__':
from pdb import set_trace
mde_model = ex.get_and_configure('MDE')
test = {'image': torch.randn(1, 3, 480, 640)}
output = mde_model(test)
print(output.shape)
print(type(output))
def setup(config):
mde_model = ex.get_and_configure('mde')
return MDEGTHist(mde_model, config['gt_hist_gt_key'])
def summarize(all_metric_dicts):
"""Metrics is a list of dictionaries, each with <metric>: value entries"""
if len(all_metric_dicts) == 0:
return {}
summary = {}
metric_names = all_metric_dicts[0].keys()
for k in metric_names:
summary[k] = np.mean([metric_dict[k] for metric_dict in all_metric_dicts])
return summary
if __name__ == '__main__':
# parser = configargparse.get_arg_parser()
# ex.config = vars(parser.parse_args())
# set_trace()
evaluator = ex.get_and_configure('NYUv2Evaluation')
# set_trace()
print(f"Evaluating {ex.config['mde']} in " + \
f"{ex.config['method']} mode.")
preds = evaluator.evaluate()
summary = summarize([p['metrics'] for p in preds])
depth_preds_cropped = np.stack([p['depth_cropped'].cpu().squeeze() for p in preds], axis=0)
config = cfg()
method_name = ex.config['method']
mde_name = ex.config['mde']
output_dir = Path(ex.config['output_dir'])/f'{method_name}'/f'{mde_name}'
if ex.config['method'] == 'transient':
output_dir = output_dir/f'sbr_{ex.config["sbr"]}'
output_dir.mkdir(parents=True, exist_ok=True)
np.save(output_dir/'summary', summary)
np.save(output_dir/'preds_cropped', depth_preds_cropped)
