def generate_pseudo(model: nn.Module, dataset: Dataset, transform: Augmentation,
device: torch.device = None, args: Arguments.parse.Namespace = None, **kwargs) \
-> dict:
result = {}
evaluator = Evaluator(num_classes=dataset.num_classes)
ground = Path(args.unlabeled_gt)
dest = Path(
args.dest).joinpath('pseudo').joinpath(f"{kwargs['iteration']:08}")
dest.mkdir(exist_ok=True, parents=True)
model.eval()
for index in tqdm(range(len(dataset))):
name = dataset.pull_name(index)
destination = dest.joinpath(f'{name}{dataset.DETECTION_EXT}')
groundtruth = ground.joinpath(f'{name}{dataset.DETECTION_EXT}')
image = dataset.pull_image(index)
scale = torch.Tensor([
image.shape[1],
image.shape[0],
image.shape[1],
image.shape[0],
]).to(device)
image = Variable(
torch.from_numpy(transform(image)[0]).permute(
2, 0, 1).unsqueeze(0)).to(device)
detection = np.empty((0, 6), dtype=np.float32)
detections, *_ = model(image).data
for klass, boxes in enumerate(detections):
if klass != dataset.class_id:
continue
candidates = boxes[boxes[:, 0] >= args.thresh]
# filter out of image
candidates = candidates[(torch.sum(
(candidates < -1) | (candidates > 2), axis=1) == 0
).nonzero().squeeze(0), :].reshape(-1, 5)
# filter nan and inf
candidates = candidates[(
torch.
sum(torch.isinf(candidates)
| torch.isnan(candidates), axis=1) == 0
).nonzero().squeeze(0), :].reshape(-1, 5)
if candidates.size(0) == 0:
continue
candidates[:, 1:] *= scale
detection = np.concatenate((
detection,
np.hstack((
np.full((np.size(candidates, 0), 1), klass,
dtype=np.uint8),
candidates.cpu().detach().numpy(),
)),
))
if args.unlabeled_gt and groundtruth.exists():
gt = pd.read_csv(str(groundtruth), header=None).values
if detection.size and gt.size:
evaluator.update((
detection[:, 0].astype(np.int),
detection[:, 1].astype(np.float32),
detection[:, 2:].astype(np.float32),
None,
), (
np.ones(np.size(gt, 0), dtype=np.int),
gt.astype(np.float32),
None,
))
pd.DataFrame(detection).to_csv(str(destination),
header=None,
index=None)
if args.unlabeled_gt:
result.update({'acc': evaluator.mAP[1:].mean()})
model.train()
dataset.detections = list(sorted(dest.glob(f'*{dataset.DETECTION_EXT}')))
return result
