AssertionError: Non-existent Key: train.dataset

[jinglan0829]

感谢作者提供的干净版本的pvnet! 当我运行了python run.py --type visualize --cfg_file configs/linemod.yaml train.datset CustomTrain test.dataset CustomTrain model mycat这个以后，出现了这种error, AssertionError: Non-existent Key: train.dataset，该如何解决呢？

[pengsida]

你字打错了 train.datset -> train.dataset

[jinglan0829]

我直接复制粘贴了，不好意思呀，没仔细看

[jinglan0829]

出来的结果不太对呀，自己数据集的train.json文件该如何获取呢？

[pengsida]

第一个问题：#1
第二个问题：https://github.com/zju3dv/clean-pvnet#training-on-the-custom-object

[ykim104]

I get this error when using custom dataset :(
/pytorch/aten/src/THCUNN/SpatialClassNLLCriterion.cu:103: void cunn_SpatialClassNLLCriterion_updateOutput_kernel(T *, T *, T *, long *, T *, int, int, int, int, int, long) [with T = float, AccumT = float]: block: [0,0,0], thread: [677,0,0] Assertiont >= 0 && t < n_classesfailed.
This happens when I call the dataloader - it might be that the computed annotations are outside [0,1) boundary? Potentially during interpolation?

It doesn't happen all the time for every image, but often enough.
I could't locate where you compute vector field annotation - can you give me a hint?

Thank you in advance and thank you for publishing this repo!

[pengsida]

I will check this problem.

[pengsida]

Vector field: https://github.com/zju3dv/clean-pvnet/blob/master/lib/datasets/linemod/pvnet.py#L53
This problem is probably caused by segmentation.

[ykim104]

yes i found out that I get error here seg_loss = self.seg_crit(output['seg'], mask)

thank you!

[pengsida]

The mask contains other values?

[ykim104]

Yess!! I had 255 instead 1's in my mask!
I also used ObjectDatasetTool but my masks were 2D
so I edited read_linemod_mask() in pvnet_data_utils.py :D

Thank you for your help!!!
if ann_type == 'real': try: return (np.asarray(Image.open(path))[:, :, 0] != 0).astype(np.uint8) # custom/custom mask return 3d matrix except: return (np.asarray(Image.open(path))).astype(np.uint8)/255.

[ykim104]

It works now!!! Thank you!

By the way, I also made few edits so I can have multiple object dataset inside custom folder. :D

[pengsida]

(np.asarray(Image.open(path)) != 0).astype(np.uint8) is a more safe way, since the background value is zero in most cases.
You could additionally define a custom dataloader for your dataset according to https://github.com/zju3dv/clean-pvnet/blob/master/project_structure.md
I feel happy if this project facilitates your work!