Open rtaori opened 2 years ago
@rtaori
Unfortunately, we don't have a plan to update our codebase now.
I can share the label map extraction code snippet, but I'm not sure whether this is the right and valid code.
I hope you can understand how the label maps were extracted through this code, but if you want fully-working codes, please wait for our next codebase update.
normalize = torchvision.transforms.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]
)
train_root = os.path.join(data_path, 'train')
train_dataset = ImageFolderWithPath(
root=train_root,
transform=torchvision.transforms.Compose([
torchvision.transforms.Resize((machine_annotator_image_size,
machine_annotator_image_size)),
torchvision.transforms.ToTensor(),
normalize,
]))
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_sizebatch_size,
shuffle=False,
num_workers=machine_annotator_num_workers,
pin_memory=True,
sampler=train_sampler
)
score_maps = {}
print('Train loader size: {}'.format(len(train_loader)))
for iteration, batch in enumerate(train_loader):
start_time = time.time()
input, target, img_names = batch
target = target.cuda()
with torch.no_grad():
# get score_map by removing GlobalAvergaPool. (See Appendix)
output_map = machine_annotator_model(input, score_map=True)
output_topk = output_map.topk(k=machine_annotator_topk, dim=1)
for idx, (t0, t1) in enumerate(zip(output_topk[0], output_topk[1])):
img_name = img_names[idx].split('/')[-1].split('.')[0]
t = torch.cat([t0.unsqueeze(0), t1.unsqueeze(0).float()],
dim=0)
score_maps[img_name] = t.cpu()
score_map_path = '{}_{}_sz{}_top{}_score_maps.pt'.format(
data.dataset,
machine_annotator_arch,
machine_annotator_image_size,
machine_annotator_topk)
Hi,
This is exciting work! Would you be able to please provide the code that you used to generate the relabel imagenet maps? I would like to apply this technique to my own dataset and the relabeling code would be super helpful. Even if it's not polished, it would be a great help :)
Thanks,