Test custom images with pre-trained model

[Brakebein]

I have downloaded the pre-trained models and the huge dataset. Running the vectorization & visualization command with the SU3 dataset works. Now, I would like to test some of my own images with the pre-trained model to see how well it works. However, I'm not sure how to do that. How do I have to organize custom images and what command I have to run to evaluate the pre-trained model with these images?

---

I also tried to run the prediction command with the SU3 dataset.

python train.py --eval -d 0 -i default --from logs/pretrained-wireframe/checkpoint_latest.pth.tar logs/pretrained-wireframe/config.yaml

However, I'm getting the error that it cannot use a CUDA device. My older Nvidia graphics card is not compatible with cudatoolkit 10.x. But it looks like that the pre-trained model has somewhere saying inside that it wants to run things on CUDA, and not fall back on CPU.

Error log

Traceback (most recent call last):
  File "train.py", line 204, in 
    main()
  File "train.py", line 124, in main
    checkpoint = torch.load(resume_from)
  File "/home/brakebein/anaconda3/envs/shapeunity/lib/python3.7/site-packages/torch/serialization.py", line 608, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/home/brakebein/anaconda3/envs/shapeunity/lib/python3.7/site-packages/torch/serialization.py", line 787, in _legacy_load
    result = unpickler.load()
  File "/home/brakebein/anaconda3/envs/shapeunity/lib/python3.7/site-packages/torch/serialization.py", line 743, in persistent_load
    deserialized_objects[root_key] = restore_location(obj, location)
  File "/home/brakebein/anaconda3/envs/shapeunity/lib/python3.7/site-packages/torch/serialization.py", line 175, in default_restore_location
    result = fn(storage, location)
  File "/home/brakebein/anaconda3/envs/shapeunity/lib/python3.7/site-packages/torch/serialization.py", line 151, in _cuda_deserialize
    device = validate_cuda_device(location)
  File "/home/brakebein/anaconda3/envs/shapeunity/lib/python3.7/site-packages/torch/serialization.py", line 135, in validate_cuda_device
    raise RuntimeError('Attempting to deserialize object on a CUDA '
RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

---

So, my next thought was that I have to train the model by myself on the CPU. So, I need to run the training command:

python ./train.py -d 0,1,2,3 --identifier baseline config/hourglass.yaml

But there is the next error I cannot resolve. I tried it several times, but it always errors at the same image or data chunk (in the folder that has been created in logs, 3 npz files has been generated, and in viz, 000002_img.jpg has been generated but not the other files (see screenshot)).

logs/.../viz/000000600/

Error log

progress | sum    | jmap   | jdir   | lmap   | joff   | ldir   | dpth   | jdep   | speed
Running validation...                                                                            Traceback (most recent call last):
  File "./train.py", line 204, in 
    main()
  File "./train.py", line 200, in main
    trainer.train()
  File "/media/brakebein/Daten4TB/projects/shapeunity/wireframe/trainer.py", line 260, in train
    self.train_epoch()
  File "/media/brakebein/Daten4TB/projects/shapeunity/wireframe/trainer.py", line 200, in train_epoch
    self.validate()
  File "/media/brakebein/Daten4TB/projects/shapeunity/wireframe/trainer.py", line 131, in validate
    self._plot_samples(i, index, H, target, f"{viz}/{index:06}")
  File "/media/brakebein/Daten4TB/projects/shapeunity/wireframe/trainer.py", line 225, in _plot_samples
    mask_target = target["jmap"][i].cpu()
IndexError: index 2 is out of bounds for dimension 0 with size 2

---

Of course, it would be better to avoid the training process. I just want to test some custom images on your pre-trained model, but it doesn't seem to be trivial. Any help is appreciated.

[Brakebein]

I still couldn't figure out how to test my own images with the pre-trained model. I would like to input a single image and see what wireframe it predicts.

Any hints or help would be appreciated.

[zhou13]

Sorry for missing the issue. Currently testing 3rd images are not supported. If you really want to test it, you need to modify the evaluation script to add a new dataset that includes your images. You need to zero-fill the ground truth entries if they are not available.

[Brakebein]

Thanks for your response. So I guess anytime soon, you won't implement it to test 3rd images.

So, I will try to modify the script. To zero-fill the ground truth entries: do I just need to use the format of the json files that are accompanied with each image? Or do I need to care about the npz files as well?

[zhou13]

I think you just need to fill the image. All other entries can be zero. You can either modifier the dataloader or provide a new json/npz. If the image intrinsic is know, you can also put it into https://github.com/zhou13/shapeunity/blob/f04615d6f688fa7eeec47fba344046ee24d67b99/eval_2d3d_metric.py#L168

[dioptrique]

Hi @Brakebein may I know if you managed to get the testing on custom images to work?

[Brakebein]

Not yet. But that's mainly because I was working on a different part in my project for the last 2 weeks. If this is done, I will continue with this issue (maybe in 2 weeks). If there is any progress in this matter until then, I would be interested as well.

[dioptrique]

@Brakebein Any updates on testing on custom images?

[Brakebein]

No, sorry. I finally tested their 2D approach (https://github.com/zhou13/lcnn) which was actually sufficient for me.

[moon6666]

您好，我使用自己的数据测试成功了，具体思路为：首先，自己创建一个data/SU1/001文件夹，并在其中存放512*512的自定义png图片，将自定义图片按照0000/0001/0002的格式重命名；其次，将作者提供的数据(SU3)中的其他对应名称的非png图片数据复制粘贴到该文件夹中；最后，直接使用命令“python train.py --eval -d 0 -i default --from logs/pretrained-wireframe/checkpoint_latest.pth.tar logs/pretrained-wireframe/config.yaml”测试结果即可。