Sorry to bother you, I have two question about this project. for training work, I use multi gpus, the speed about 1.79it/s in first two epoch, it takes one day. After that, the speed is so fast, I don't know why this happened. Also the result, the detections repeatability is more better than rpautrat/Superpoint's, but the descriptors: hpatches-i is 0.90. hpatches-v is 0.55. Compare with rpautrat/Superpoint's, the result is not good.

作者您好！我在训练完成后使用Superpoint推断时发现即使设置了with torch.no_grad():，当我连续推断两张图片时显存溢出，并且速度很慢，一张480p的图片放在gpu上需要1s `device = 'cuda:0' if torch.cuda.is_available() else 'cpu' # device='cpu' net = SuperPointBNNet(config['model'], device=device, using_bn=config['model']['using_bn']) net.load_state_dict(torch.load(config['model']['pretrained_model'],map_location=device)) net.to(device).eval()

example = cv2.imread("/home/xiefei/COCO_test2014_000000000001.jpg",0)
img_tensor = torch.as_tensor(example.copy(), dtype=torch.float, device=device)
img_tensor=img_tensor/255.
img_tensor=img_tensor.unsqueeze(0)
img_tensor=img_tensor.unsqueeze(0)

with torch.no_grad():
    t1=time.time()
    output=net(img_tensor)
    t2=time.time()
    print(t2-t1)`

首先感谢您的工作 我是一个深度学习新手 我在阅读 README.md 时，发现可能存在一些笔误，虽然不影响代码的运行，但是给我带来了一些困扰，比如 Usage 2.0 中表述的 in train.py, line 61，实际上由于代码改动，好像不在那一行了。 2.4 和 2.6 中的 python train.py ./config/magic_point_coco_train.py #with correct data dirs 和 python train.py ./config/superpoint_train.py #with correct data dirs 表述的 ./config/.py 是否应为 ./config/.yaml 。

Envs is installed as required.

0it [00:00, ?it/s]D:\Users\sichu\anaconda3\envs\point-torch\lib\site-packages\kornia\utils\helpers.py:96: UserWarning: torch.solve is deprecated in favor of torch.linalg.solveand will be removed in a future PyTorch release. torch.linalg.solve has its arguments reversed and does not return the LU factorization. To get the LU factorization see torch.lu, which can be used with torch.lu_solve or torch.lu_unpack. X = torch.solve(B, A).solution should be replaced with X = torch.linalg.solve(A, B) (Triggered internally at ..\aten\src\ATen\native\BatchLinearAlgebra.cpp:760.) out1, out2 = torch.solve(input.to(dtype), A.to(dtype)) D:\Users\sichu\anaconda3\envs\point-torch\lib\site-packages\torch\nn\functional.py:718: UserWarning: Named tensors and all their associated APIs are an experimental feature and subject to change. Please do not use them for anything important until they are released as stable. (Triggered internally at ..\c10/core/TensorImpl.h:1156.) return torch.max_pool2d(input, kernel_size, stride, padding, dilation, ceil_mode)

Traceback (most recent call last): File "train.py", line 139, in train_eval(model, data_loaders, config) File "train.py", line 30, in train_eval raw_outputs = model(data['raw']) File "D:\Users\sichu\anaconda3\envs\point-torch\lib\site-packages\torch\nn\modules\module.py", line 1051, in _call_impl return forward_call(*input, **kwargs) File "D:\experments\SuperPoint-Pytorch-master\SuperPoint-Pytorch-master\model\magic_point.py", line 42, in forward keep_top_k=self.topk).squeeze(dim=0) for p in prob] File "D:\experments\SuperPoint-Pytorch-master\SuperPoint-Pytorch-master\model\magic_point.py", line 42, in keep_top_k=self.topk).squeeze(dim=0) for p in prob] File "D:\experments\SuperPoint-Pytorch-master\SuperPoint-Pytorch-master\solver\nms.py", line 45, in box_nms indices = torchvision.ops.nms(boxes=boxes, scores=scores, iou_threshold=iou) File "D:\Users\sichu\anaconda3\envs\point-torch\lib\site-packages\torchvision\ops\boxes.py", line 35, in nms return torch.ops.torchvision.nms(boxes, scores, iou_threshold) RuntimeError: dets should have the same type as scores

RuntimeError: Expected 4-dimensional input for 4-dimensional weight [64, 1, 3, 3], but got 3-dimensional input of size [16, 120, 160] instead

my envs: cuda11.0 cudnn8.0.5 torch 1.7.1+cu110

Hi, fantastic work done! Applause for you!!

May I know if there is any instruction or codes for a quick demo using pre-trained models? Thank you in advance!

Also, may I know: between the weight released by magicleap and the weight released by rpautrat's superpoint, which one has better performance?

作者您好，我注意到您在训练superpoint时的descriptor loss部分的lambda_d设置为0.05。这个lambda_d不是添加在正样本上的权重吗？（正样本数量较少，理应设置大点的权重）Superpoint的作者在论文中陈述的设置为250，您为什么使用0.05呢？是您在实际训练中找到的更合适的新参数吗？ https://github.com/shaofengzeng/SuperPoint-Pytorch/blob/master/config/superpoint_train.yaml

我在训练时给正样本设置大的权重也没有获得好的效果，看到您这么设置，能不能问下这么设置的原因呢？

作者您好，近期在遵循您给出的训练步骤进行复现时，发现如果将输入图片尺寸修改为[480, 640]，则代码将无法正常工作，并提示显存不足，我目前所采用的设备是2080Ti显卡，8G显存，因为我个人觉得依照这个网络模型，[480,640]的图片尺寸不应该出现超出8G显存的问题，针对超出显存的报错，我不知道问题出在了哪里，我在训练的过程中即使将batch_size修改为1仍然会提示超出显存，所以想请教一下您，为什么网络面向[480,640]的图片尺寸时即使batch_size为1还会出现显存不够的问题呢？期待您的回复

I guess superpoint_v1 is the pytorch version model file converted by the TF version. Is the superpointbn.pth have same weights as the former one except the modified parts (I see that in the bn version, you made some modifications on the architecture)?

请问在dataset/utils/augmentation_legacy.py文件的elastic_transform函数中为什么会按照如下的方式定义min_row,max_row,min_col,max_col min_row = int(math.ceil(np.max(dy))) + padding
max_row = int(math.floor(np.min(dy))) + shape[0] - padding
min_col = int(math.ceil(np.max(dx))) + padding max_col = int(math.floor(np.min(dx))) + shape[1] - padding

Hi, thanks for your great job, if I trained magic point with rpautrat's project to get coco gt points, then use your version to train superpoint, will transforming the coco image to 240* 320 by function ratio_preserving_resize have a big impact on the results?

How should I modify the code to conduct multi GPU training? At the end of train.py file the following code is added before train_eval(model, data_loaders, config) function.

if torch.cuda.device_count() > 1: 
    model = torch.nn.DataParallel(model, device_ids=list(range(torch.cuda.device_count()))
model.to(device)

After that, When I train the code, I get the following messages.

Traceback (most recent call last):
  File "train.py", line 212, in <module>
    train_eval(model, data_loaders, config)
  File "train.py", line 84, in train_eval
    prob_warp, desc_warp, device)
  File "/root/Source/gtr/SuperPoint_pytorch/temp/SuperPoint-Pytorch-master/solver/loss.py", line 16, in loss_func
    device=device)
  File "/root/Source/gtr/SuperPoint_pytorch/temp/SuperPoint-Pytorch-master/solver/loss.py", line 65, in detector_loss
    ce_loss = F.cross_entropy(logits, labels, reduction='none',)
  File "/root/PersonalData/miniconda/envs/superpoint/lib/python3.7/site-packages/torch/nn/functional.py", line 3026, in cross_entropy
    return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
ValueError: Expected input batch_size (16) to match target batch_size (64).

It seems to be the problem of data set allocation under multi GPU training. But I don't know how to modify it.

Hi, thanks for your great job. I want to train magicpoint on kitti , so is it needed to use magic_point_coco_train.yaml ? I noticed there is a path of labels. but how to ensure its correctness so as to train magicpoint ? because if the label result generated in the previous step is average, what is the significance of this step of train magicpoint ?