class SuperGlueLightning(LightningModule):
def __init__(self, args: DictConfig):
super().__init__()
self.hparams = args # Will be logged to mlflow
# make sure the flags are properly used
assert not (args.exp.opencv_display and
not args.exp.viz), 'Must use --viz with --opencv_display'
assert not (args.exp.opencv_display and not args.exp.fast_viz
), 'Cannot use --opencv_display without --fast_viz'
assert not (args.exp.fast_viz
and not args.exp.viz), 'Must use --viz with --fast_viz'
assert not (args.exp.fast_viz and args.exp.viz_extension
== 'pdf'), 'Cannot use pdf extension with --fast_viz'
# store viz results
# eval_output_dir = Path(f'{ROOT_PATH}/' + args.data.eval_output_dir)
# eval_output_dir.mkdir(exist_ok=True, parents=True)
# print('Will write visualization images to directory \"{}\"'.format(eval_output_dir))
self.superglue = SuperGlue(args.model.superglue)
self.superpoint = SuperPoint(args.model.superpoint)
self.lr = None
def prepare_data(self) -> None:
self.train_set = SparseDataset(
f'{ROOT_PATH}/' + self.hparams.data.train_path,
self.hparams.model.superpoint.max_keypoints,
self.hparams.data.resize, self.hparams.data.resize_float,
self.hparams.model.superglue.min_keypoints, None, self.superpoint)
self.val_set = SparseDataset(
f'{ROOT_PATH}/' + self.hparams.data.val_path,
self.hparams.model.superpoint.max_keypoints,
self.hparams.data.resize, self.hparams.data.resize_float,
self.hparams.model.superglue.min_keypoints, None, self.superpoint)
self.train_set.files = self.train_set.files[:self.hparams.data.
train_size]
self.val_set.files = self.val_set.files[:self.hparams.data.val_size]
def train_dataloader(self) -> torch.utils.data.DataLoader:
return torch.utils.data.DataLoader(
dataset=self.train_set,
shuffle=True,
batch_size=self.hparams.data.batch_size.train,
drop_last=True)
def val_dataloader(self) -> torch.utils.data.DataLoader:
return torch.utils.data.DataLoader(
dataset=self.val_set,
shuffle=True,
batch_size=self.hparams.data.batch_size.val,
drop_last=True)
def configure_optimizers(self):
if self.lr is not None:
self.hparams.optimizer.learning_rate = self.lr
optimizer = torch.optim.Adam(self.superglue.parameters(),
lr=self.hparams.optimizer.learning_rate)
return optimizer
def training_step(self, batch, batch_ind):
for k in batch:
if k != 'file_name' and k != 'image0' and k != 'image1':
if type(batch[k]) == torch.Tensor:
batch[k] = Variable(batch[k])
else:
batch[k] = Variable(torch.stack(batch[k]))
# print(batch['keypoints0'].shape)
data = self.superglue(batch)
batch = {**batch, **data}
# print(batch['keypoints0'].shape)
if batch['skip_train']: # image has no keypoint
batch['loss'] = torch.zeros(1, requires_grad=True).type_as(
batch['scores0'])
return batch['loss']
# Loss & Metrics
self.log('train_hits_1',
hits_at_one(batch).mean(),
on_epoch=False,
on_step=True,
sync_dist=True)
self.log('train_loss_m',
batch['loss_m'],
on_epoch=False,
on_step=True,
sync_dist=True)
self.log('train_loss_um',
batch['loss_um'],
on_epoch=False,
on_step=True,
sync_dist=True)
self.log('train_loss',
batch['loss'],
on_epoch=False,
on_step=True,
sync_dist=True)
self.log('bin_score',
self.superglue.bin_score,
on_epoch=False,
on_step=True,
sync_dist=True)
return batch['loss']
def validation_step(self, batch, batch_ind):
for k in batch:
if k != 'file_name' and k != 'image0' and k != 'image1':
if type(batch[k]) != torch.Tensor:
batch[k] = torch.stack(batch[k])
data = self.superglue(batch)
batch = {**batch, **data}
if batch['skip_train']: # image has no keypoint
batch['loss'] = torch.zeros(1).type_as(batch['scores0'])
return batch['loss']
# Loss & Metrics
self.log('val_hits_1',
hits_at_one(batch).mean(),
on_epoch=True,
on_step=False,
sync_dist=True)
self.log('val_loss',
batch['loss'],
on_epoch=True,
on_step=False,
sync_dist=True)
self.log('val_loss_m',
batch['loss_m'],
on_epoch=True,
on_step=False,
sync_dist=True)
self.log('val_loss_um',
batch['loss_um'],
on_epoch=True,
on_step=False,
sync_dist=True)
return batch['loss']
@rank_zero_only
def on_train_epoch_end(self, outputs) -> None:
for eval_data in self.hparams.eval:
with open(f'{ROOT_PATH}/' + eval_data.pairs_list, 'r') as f:
pairs = [l.split() for l in f.readlines()]
if eval_data.max_length > -1:
pairs = pairs[0:np.min([len(pairs), eval_data.max_length])]
if eval_data.shuffle:
random.Random(0).shuffle(pairs)
if not all([len(p) == 38 for p in pairs]):
raise ValueError(
'All pairs should have ground truth info for evaluation.'
'File \"{}\" needs 38 valid entries per row'.format(
eval_data.pairs_list))
# Load the SuperPoint and SuperGlue models.
device = 'cuda' if torch.cuda.is_available() else 'cpu'
print('Running inference on device \"{}\"'.format(device))
config = {
'superpoint': {
'nms_radius': eval_data.nms_radius,
'keypoint_threshold': eval_data.keypoint_threshold,
'max_keypoints': eval_data.max_keypoints
},
'superglue': self.hparams.model.superglue,
}
matching = Matching(config).eval().to(device)
matching.superglue.load_state_dict(self.superglue.state_dict())
# Create the output directories if they do not exist already.
data_dir = Path(f'{ROOT_PATH}/' + eval_data.data_dir)
# moving_dir = Path(f'{ROOT_PATH}/' + 'data/ScanNet/test_subset')
print('Looking for data in directory \"{}\"'.format(data_dir))
results_dir = Path(os.getcwd() + '/' + eval_data.results_dir)
results_dir.mkdir(exist_ok=True, parents=True)
print('Will write matches to directory \"{}\"'.format(results_dir))
timer = AverageTimer(newline=True)
for i, pair in enumerate(pairs):
name0, name1 = pair[:2]
stem0, stem1 = Path(name0).stem, Path(name1).stem
matches_path = results_dir / '{}_{}_matches.npz'.format(
stem0, stem1)
eval_path = results_dir / '{}_{}_evaluation.npz'.format(
stem0, stem1)
viz_path = results_dir / '{}_{}_matches.{}'.format(
stem0, stem1, self.hparams.exp.viz_extension)
viz_eval_path = results_dir / \
'{}_{}_evaluation.{}'.format(stem0, stem1, self.hparams.exp.viz_extension)
# Handle --cache logic.
do_match = True
do_eval = True
do_viz = self.hparams.exp.viz
do_viz_eval = self.hparams.exp.viz
# if opt.cache:
# if matches_path.exists():
# try:
# results = np.load(matches_path)
# except:
# raise IOError('Cannot load matches .npz file: %s' %
# matches_path)
#
# kpts0, kpts1 = results['keypoints0'], results['keypoints1']
# matches, conf = results['matches'], results['match_confidence']
# do_match = False
# if opt.eval and eval_path.exists():
# try:
# results = np.load(eval_path)
# except:
# raise IOError('Cannot load eval .npz file: %s' % eval_path)
# err_R, err_t = results['error_R'], results['error_t']
# precision = results['precision']
# matching_score = results['matching_score']
# num_correct = results['num_correct']
# epi_errs = results['epipolar_errors']
# do_eval = False
# if opt.viz and viz_path.exists():
# do_viz = False
# if opt.viz and opt.eval and viz_eval_path.exists():
# do_viz_eval = False
# timer.update('load_cache')
if not (do_match or do_eval or do_viz or do_viz_eval):
timer.print('Finished pair {:5} of {:5}'.format(
i, len(pairs)))
continue
# If a rotation integer is provided (e.g. from EXIF data), use it:
if len(pair) >= 5:
rot0, rot1 = int(pair[2]), int(pair[3])
else:
rot0, rot1 = 0, 0
# Load the image pair.
image0, inp0, scales0 = read_image(data_dir / name0,
eval_data.resize, rot0,
eval_data.resize_float)
image1, inp1, scales1 = read_image(data_dir / name1,
eval_data.resize, rot1,
eval_data.resize_float)
# Moving
# os.makedirs(os.path.dirname(moving_dir / name0), exist_ok=True)
# os.makedirs(os.path.dirname(moving_dir / name1), exist_ok=True)
# shutil.copy(data_dir / name0, moving_dir / name0)
# shutil.copy(data_dir / name1, moving_dir / name1)
if image0 is None or image1 is None:
print('Problem reading image pair: {} {}'.format(
data_dir / name0, data_dir / name1))
exit(1)
timer.update('load_image')
if do_match:
# Perform the matching.
with torch.no_grad():
pred = matching({
'image0': inp0.cuda(),
'image1': inp1.cuda()
})
pred_np = {}
for (k, v) in pred.items():
if isinstance(v, list):
pred_np[k] = v[0].cpu().numpy()
elif isinstance(v, torch.Tensor):
pred_np[k] = v[0].cpu().numpy()
pred = pred_np
# pred = {k: v[0].cpu().numpy() for k, v in pred.items() if isinstance(v, torch.Tensor)}
kpts0, kpts1 = pred['keypoints0'], pred['keypoints1']
matches, conf = pred['matches0'], pred['matching_scores0']
timer.update('matcher')
# Write the matches to disk.
out_matches = {
'keypoints0': kpts0,
'keypoints1': kpts1,
'matches': matches,
'match_confidence': conf
}
np.savez(str(matches_path), **out_matches)
# Keep the matching keypoints.
valid = matches > -1
mkpts0 = kpts0[valid]
mkpts1 = kpts1[matches[valid]]
mconf = conf[valid]
if do_eval:
# Estimate the pose and compute the pose error.
assert len(
pair) == 38, 'Pair does not have ground truth info'
K0 = np.array(pair[4:13]).astype(float).reshape(3, 3)
K1 = np.array(pair[13:22]).astype(float).reshape(3, 3)
T_0to1 = np.array(pair[22:]).astype(float).reshape(4, 4)
# Scale the intrinsics to resized image.
K0 = scale_intrinsics(K0, scales0)
K1 = scale_intrinsics(K1, scales1)
# Update the intrinsics + extrinsics if EXIF rotation was found.
if rot0 != 0 or rot1 != 0:
cam0_T_w = np.eye(4)
cam1_T_w = T_0to1
if rot0 != 0:
K0 = rotate_intrinsics(K0, image0.shape, rot0)
cam0_T_w = rotate_pose_inplane(cam0_T_w, rot0)
if rot1 != 0:
K1 = rotate_intrinsics(K1, image1.shape, rot1)
cam1_T_w = rotate_pose_inplane(cam1_T_w, rot1)
cam1_T_cam0 = cam1_T_w @ np.linalg.inv(cam0_T_w)
T_0to1 = cam1_T_cam0
epi_errs = compute_epipolar_error(mkpts0, mkpts1, T_0to1,
K0, K1)
correct = epi_errs < 5e-4
num_correct = np.sum(correct)
precision = np.mean(correct) if len(correct) > 0 else 0
matching_score = num_correct / len(kpts0) if len(
kpts0) > 0 else 0
thresh = 1. # In pixels relative to resized image size.
ret = estimate_pose(mkpts0, mkpts1, K0, K1, thresh)
if ret is None:
err_t, err_R = np.inf, np.inf
else:
R, t, inliers = ret
err_t, err_R = compute_pose_error(T_0to1, R, t)
# Write the evaluation results to disk.
out_eval = {
'error_t': err_t,
'error_R': err_R,
'precision': precision,
'matching_score': matching_score,
'num_correct': num_correct,
'epipolar_errors': epi_errs
}
np.savez(str(eval_path), **out_eval)
timer.update('eval')
# if do_viz:
# # Visualize the matches.
# color = cm.jet(mconf)
# text = [
# 'SuperGlue',
# 'Keypoints: {}:{}'.format(len(kpts0), len(kpts1)),
# 'Matches: {}'.format(len(mkpts0)),
# ]
# if rot0 != 0 or rot1 != 0:
# text.append('Rotation: {}:{}'.format(rot0, rot1))
#
# make_matching_plot(
# image0, image1, kpts0, kpts1, mkpts0, mkpts1, color,
# text, viz_path, stem0, stem1, opt.show_keypoints,
# opt.fast_viz, opt.opencv_display, 'Matches')
#
# timer.update('viz_match')
#
# if do_viz_eval:
# # Visualize the evaluation results for the image pair.
# color = np.clip((epi_errs - 0) / (1e-3 - 0), 0, 1)
# color = error_colormap(1 - color)
# deg, delta = ' deg', 'Delta '
# if not opt.fast_viz:
# deg, delta = '°', '$\\Delta$'
# e_t = 'FAIL' if np.isinf(err_t) else '{:.1f}{}'.format(err_t, deg)
# e_R = 'FAIL' if np.isinf(err_R) else '{:.1f}{}'.format(err_R, deg)
# text = [
# 'SuperGlue',
# '{}R: {}'.format(delta, e_R), '{}t: {}'.format(delta, e_t),
# 'inliers: {}/{}'.format(num_correct, (matches > -1).sum()),
# ]
# if rot0 != 0 or rot1 != 0:
# text.append('Rotation: {}:{}'.format(rot0, rot1))
#
# make_matching_plot(
# image0, image1, kpts0, kpts1, mkpts0,
# mkpts1, color, text, viz_eval_path,
# stem0, stem1, opt.show_keypoints,
# opt.fast_viz, opt.opencv_display, 'Relative Pose')
#
# timer.update('viz_eval')
timer.print('Finished pair {:5} of {:5}'.format(i, len(pairs)))
# Collate the results into a final table and print to terminal.
pose_errors = []
precisions = []
matching_scores = []
for pair in pairs:
name0, name1 = pair[:2]
stem0, stem1 = Path(name0).stem, Path(name1).stem
eval_path = results_dir / \
'{}_{}_evaluation.npz'.format(stem0, stem1)
results = np.load(eval_path)
pose_error = np.maximum(results['error_t'], results['error_R'])
pose_errors.append(pose_error)
precisions.append(results['precision'])
matching_scores.append(results['matching_score'])
thresholds = [5, 10, 20]
aucs = pose_auc(pose_errors, thresholds)
aucs = [100. * yy for yy in aucs]
prec = 100. * np.mean(precisions)
ms = 100. * np.mean(matching_scores)
print('Evaluation Results (mean over {} pairs):'.format(
len(pairs)))
print('AUC@5\t AUC@10\t AUC@20\t Prec\t MScore\t')
print('{:.2f}\t {:.2f}\t {:.2f}\t {:.2f}\t {:.2f}\t'.format(
aucs[0], aucs[1], aucs[2], prec, ms))
self.log(f'{eval_data.name}/AUC_5',
aucs[0],
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/AUC_10',
aucs[1],
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/AUC_20',
aucs[2],
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/Prec',
prec,
on_epoch=True,
on_step=False)
self.log(f'{eval_data.name}/MScore',
ms,
on_epoch=True,
on_step=False)
}
# load training data
train_set = SparseDataset(opt.train_path, opt.max_keypoints)
train_loader = torch.utils.data.DataLoader(dataset=train_set,
shuffle=False,
batch_size=opt.batch_size,
drop_last=True)
superglue = SuperGlue(config.get('superglue', {}))
if torch.cuda.is_available():
superglue.cuda() # make sure it trains on GPU
else:
print("### CUDA not available ###")
optimizer = torch.optim.Adam(superglue.parameters(), lr=opt.learning_rate)
mean_loss = []
# start training
for epoch in range(1, opt.epoch + 1):
epoch_loss = 0
superglue.double().train()
for i, pred in enumerate(train_loader):
for k in pred:
if k != 'file_name' and k != 'image0' and k != 'image1':
if type(pred[k]) == torch.Tensor:
pred[k] = Variable(pred[k].cuda())
else:
pred[k] = Variable(torch.stack(pred[k]).cuda())
data = superglue(pred)