alonlahav / meshwalker Goto Github PK

Hello, I contact you because of a problem when using the command with "tensorboard".
Indeed, after having installed all the necessary packages (in the right version), it is impossible for me to execute this command tensorboard <trained-model-folder>.
In my case, I use:

(the "0003-10.05.2022..13.58__cubes " folder is obtained after executing the command : python train_val.py cubes)

Because there is no tensorboard.main file and I get the following errors :

(I use windows 10)
Do you know how I could proceed to solve this problem ?

Thank you for sharing your work. You use the datasets the same as the dataset in MeshCNN, right? The dataset in MeshCNN is with edge_ label. How do you evaluate your method with face_label? Where is the label conversion code from the edge_ Label to the face_ label?

Hi,

When I am trying to execute the segmentation code (evaluate_segmentation) on ModelNet40 dataset, I get the following error

This is my script to execute the segmentation on pre-trained model
python evaluate_segmentation.py modelnet40 --- pretrained/0497-20.05.2020..21.17__modelnet__SeqLen800

Kindly help

Hello,
Hope you are well. I downloaded the code and your preprocessed data and ran the code as mentioned on the shrec data:
python train_val.py shrec11 16-04_A
However, after 4002 epochs the maximum test accuracy I am able to obtain is 0.82%:
26932) Epoch3985, iter 59760, TrnLoss: 2.15, test/accuracy_shrec16: 0.82, time: 4.5

Similarly, when I train:
python train_val.py shrec11 16-04_B,
the maximum test accuracy I achieve is 0.84%.
30624) Epoch3986, iter 59775, TrnLoss: 2.17, test/accuracy_shrec16: 0.84, time: 4.7

Your paper reported 98% accuracy on the three parts on average.
I am not sure why this could be happening and would appreciate your help.

Thank you.

Hi Alon,
I hope you are well. I downloaded the CUBES data set from the provided link and noticed the train/test split is different from the one mentioned in the original paper.
This is the line from your paper:
"This dataset contains 4600 objects, with 3910/690
training/testing split. Each object is a cube "engraved" with a shape
at a random face in a random location, as demonstrated in Fig. 4.
The engraved shape belongs to a dataset of 23 classes (e.g., car,
heart, apple, etc.), each contains roughly 20 shapes. "
However, the downloaded link provides a train/test split of 3722/659 with 22 classes. And the total meshes add up to 4381 instead of 4600.
I noticed the authors of MeshCNN had modified their dataset, and I have raised the same issue there, but I haven't received a reply from them.
ranahanocka/MeshCNN#120

I would really appreciate some clarification on this.

Hi Alon,
Hope you are well. Would it be possible for you to share the code used to visualize the t-SNE for the input and internal layers? I am referring to Fig. 8 of your paper?

Thanks,
Vinit

Hi, hope you are well.
It's an inspiring work, and I have some questions about the evaluation and test code.

1. The paper mentioned that several wallks were taken to predict the cls/seg result. But as the wallks may be different, the result u get after running evaluation_cls.py may be different. For example, I run the trainng script to train the model on cubes/human_seg dataset and run the evaluation code several times which give me different accuracy, some of them are higher than the reported num. So, how did you get the final reported accuracy. Run the evaluation code only once or run many times to get the average or take the minimum.
2. I find out that the code in this repo shows that the features for modelnet40 dataset is xyz rather than dxdydz, so which kind of features is used for the final modelnet40 cls result in the paper.
3. The pretrained folder website is not avaliable(403 Forbidden at 2021/12/12)(That's why I have the problem with Q2), could you upload it to google drive or dropbox or sth else?

Thanks for your work and the attention.

what means the moel_ftrs_ in the test_dataset?
And I want to know how to create model_ftrs_ in ply or obj files

I was trying out SGD, and found that params.learning_rate[0] is not defined!

I want to visualize results of the test data.
How can i do that?

Hello,
I'm confused about the params.n_classes in coseg_aliens. In the dataset, the number of classes is 4.
Thanks again.

Hi Alon,
I am running evaluate_segmentation on some pretrained models and would like to view the segmented 3D objects (colored).
I noticed there is a color class in the utils but no use for it.
Could you help me figure out what is the best way to get an quick visualization?
Thanks in advance,
Hadas

Hello,
I had a question about the number of epochs, compared to other methods (like MeshCNN) your method runs for far more epochs. Like in the case of ModelNet40, 500000 iterations. Is this due to a typo or does MeshWalker need to be trained so much?

Thanks again.

Hi,

Thank you for putting the code online. when i run it on modelnet40 or cubes, i got this error:

File "train_val.py", line 255, in run_one_job
train_val(params)
File "train_val.py", line 168, in train_val
gpu_tmpr = utils.get_gpu_temprature()
File "/home/abhishek/Scr/MeshWalker/utils.py", line 42, in get_gpu_temprature
temp = int(output)
ValueError: invalid literal for int() with base 10: ''

Thank you for your help

Hello,
Hope you are well. Would you please send me a copy of the raw data of shrec11, [email protected],thank you very much!

Hi, the paper mentioned that the walk number in classification task is 32, for segmentation task it's 32 * #SegClsNum.

But as I read the code, I don't know if I am misunderstanding the code, I found that for shrec, cubes, the default value is 16, see code. For modelnet40, the number is 64, see code

For the segmentation task, the walk number is 32 * 32 see code.

I wonder to know the number of walks for the results in the paper. Any help regarding this is appreciated! Thanks in advance.

Best wishes.