I am trying to implement MeshCNN for Shapenet models. While It seems that I am able to get past the manifold issue using https://github.com/hjwdzh/Manifold but I am stuck with another problem. I have noticed that the number of pooled edges vary (maybe because feature vectors are different and therefore they appear differently in the queue each time) and often they are larger than the target number of pooled edges. Do you have any idea as to why this might be happening? Is there a way around this problem?

Thanks!

Hi, Rana,

I want to parse the human segmentation dataset, but I cannot find the description of it. Is there any online site or directed document about it? And, How can I understand/parse the .eseg and .seseg label files?

Thanks!

I would love to see a new feature added to the blender_process.py script allowing you to take a big mesh with an *.eseg file and then returning a simplified mesh with new corresponding annotations

When running the blender script to simplify the shapes, it always gives the following message: Error: File format is not supported in file 'C:\BatchAnalysis\MeshNetworkInput\0\test\filename.obj'

However, the .obj file is created and can be opened with FreeCAD for example.

Could you share other links?

This is more a question rather than an issue, so apologies if it is not the right place to ask it.

Have you also tried the network on the de facto ModelNet50 (classification) and ShapeNet Core (segmentation) datasets?

Thank you for your great work!
I run your code on my local machine.But I have a little question. I use a Cube Mesh and operate a MeshConv on it , I can get the output features, but how can I get the mesh result?

and output

Hey,

I have installed Pytorch without CUDA (since I don't have a NVIDIA gpu) and it raises an error that requires mending your code.
AttributeError: module 'torch._C' has no attribute '_cuda_setDevice'

It might be worth mentioning in the Readme file.

It fails at:

MeshCNN/options/base_options.py", line 54, in parse torch.cuda.set_device(self.opt.gpu_ids[0])

I have tried changing it locally to work but it seems like it leads to more errors relating that.
The issue is, since I can't use CUDA, which pytorch version should I install? (As I installed "CUDAless" version).

Thanks!

Hi I think I'm confused about how to test my own test case. run those commands(human segmentation related) works fine for me, but if I want to segment some other obj file, what should I do? I tried to put my obj file in test but that doesn't seem to work, it ask me to have seg files.
I don't need to test the result of the segmentation, just want to segment with trained data.
Thanks a lot and really appreciate your code.

Hi Rana,
I really appreciate you for sharing such a nice code and your mesh pooling idea is really great.
I used your segmentation code for my own dataset and I expected to get a vector that represents which edge belongs to which segment or final segmentation plot you provide in the Github.

Besides, what I get is in the checkpoints, human seg I get some obj files that shows how mesh-pool decimate faces (gray plots you provide in the Github which I can see through view.sh) I supposed these plots should be allocated to classification.
It would be so nice if you help me to find my answer.

Best regards,
Fatemeh

Hi,

I was wondering how hard it would be to adapt the code to work with regression (reconstructing input meshes, with e.g. an Autoencoder framework).

Have you tried this before?

I didn't delve too deep before being certain this could be done, but I suppose the main thing to alter would be the optimization error. Any hints on how I could tackle this? Should I start from a "segmentation" type dataset and compute .eseg files as the edge features instead of using a per vertex L2 error?

Thanks a lot

Hello @ranahanocka, I've noticed that the number of epochs when running the training scripts for segmentation and cubes classification is 2100, while the number of epochs for SHREC classification is 200. I am curious about this big difference, especially since the other parameters of the networks are so similar.

In my experiments, it seems like the testing accuracy is stable way before reaching 2100 epochs, so I was wondering if this is a small bug (--niter_decay is only specified in SHREC classification but not for the other experiments). If it isn't, I would appreciate if you could share your insights about why SHREC does not need that many training iterations compared to the other datasets.

Thanks!

Traceback (most recent call last):
File "train.py", line 23, in
for i, data in enumerate(dataset):
File "/home/ankit/MeshCNN/data/init.py", line 33, in iter
for i, data in enumerate(self.dataloader):
File "/home/ankit/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 637, in next
return self._process_next_batch(batch)
File "/home/ankit/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 658, in _process_next_batch
raise batch.exc_type(batch.exc_msg)
ValueError: Traceback (most recent call last):
File "/home/ankit/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 138, in _worker_loop
samples = collate_fn([dataset[i] for i in batch_indices])
File "/home/ankit/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 138, in
samples = collate_fn([dataset[i] for i in batch_indices])
File "/home/ankit/MeshCNN/data/classification_data.py", line 31, in getitem
edge_features = pad(edge_features, self.opt.ninput_edges)
File "/home/ankit/MeshCNN/util/util.py", line 22, in pad
return np.pad(input_arr, pad_width=npad, mode='constant', constant_values=val)
File "/home/ankit/anaconda3/envs/meshcnn/lib/python3.6/site-packages/numpy/lib/arraypad.py", line 1200, in pad
pad_width = _validate_lengths(narray, pad_width)
File "/home/ankit/anaconda3/envs/meshcnn/lib/python3.6/site-packages/numpy/lib/arraypad.py", line 985, in _validate_lengths
raise ValueError(fmt % (number_elements,))
ValueError: [(0, 0), (0, -150)] cannot contain negative values.

""
Do i need to change any base options to counter this error ?
""

How do we define these parameters for defining the MeshConvNet model?

thank you!

following this issue: #1

add check for numpy version and throw exception with useful message if version is too new

Hi Rana,
My task is to use MeshCNN generate some specific meshes.
I have same questions during using this.During the training time, the samples have the same adjacent-edges matrix? Meshes in my dataset have different , can I use MeshCNN?

Hi @ranahanocka,
I am currently trying to import the .obj files with their corresponding segmentation into Blender.
In my first attempt I imported the high resolution Objekt (.Off file) as well as the segmentation labels (.seg) into Blender. After matching the Face labels within the .seg files to their corresponding vertices and grouping them in vertice groups according to their labels, I received a nicely segmented human body.

Now I want to do the same thing with the low resolution object (.obj file) and the edge labels provided in the .eseg file. After matching the edge labels to their corresponding vertices and adding them to different vertex group depending on their label, I am not receiving any good results. The vertices of class head, look like this for example:

Do you have any Idea why it did not work with the low resolution object?
Seems like the order of the Vertex Indexes do not match with the labels anymore.

Could you please explain the function "__get_invalids" which is defined in file "mesh_pool.py"? I want to know what will happen after these operations?
"
MeshPool.__redirect_edges(mesh, edge_id, side, update_key_a, update_side_a)
MeshPool.__redirect_edges(mesh, edge_id, side + 1, update_key_b, update_side_b)
MeshPool.__redirect_edges(mesh, update_key_a, MeshPool.__get_other_side(update_side_a), update_key_b, MeshPool.__get_other_side(update_side_b))
MeshPool.__union_groups(mesh, edge_groups, key_a, edge_id)
MeshPool.__union_groups(mesh, edge_groups, key_b, edge_id)
MeshPool.__union_groups(mesh, edge_groups, key_a, update_key_a)
MeshPool.__union_groups(mesh, edge_groups, middle_edge, update_key_a)
MeshPool.__union_groups(mesh, edge_groups, key_b, update_key_b)
MeshPool.__union_groups(mesh, edge_groups, middle_edge, update_key_b)
"
Thank you very much!

Hi,
I want to change how MeshCNN collapses the edges in order to train it to reduce a mesh based on its normals. The smaller the difference between the input Triangle Normal and the Output, the more likely it is that the edge gets collapsed. The overall goal of me is to write a neural network that reduces the polygon count of any input mesh based on a set of criteria.

Where exactly can i find the code that decides whether or not a neuron should collapse an edge or not? i would guess that its in models/layers/mesh_pool.py and there somewhere in the "__build_queue" function. But since the code is not very well documented, i dont quite understand what its doing there.
Any idea where i should try to implement that feature?

Hi @ranahanocka!
I want to create new ground truth data for the human data set with slightly varying vertices to check whether your trained model will be able to segment those.
I managed to create the ground truth eseg file with the matlab code that you provided.
Unfortunately, the seseg-files are not created and I can't find a different function that I have to run. Is there a different way to create the seseg files?

Commit feaa6c1 breaks compatibility with pytorch=1.0.1, which is the one specified in the environment.yml file.

TypeError: can't convert np.ndarray of type numpy.bool_. The only supported types are: double, float, float16, int64, int32, and uint8.

Updating pytorch to 1.2 version fix the issue.

hi rana
We at printsyst are using mostly STL files. This is the standard output format of many cad tools and most common format for 3D printers.
why you used (and using) obj file instead?

Reuven

Hi Rana.

Does the training data include the augmented data generated for the classification task?I have a pretty small dataset to work on, and augmentated data would be really helpful.

I see that .npz files stored in ./train/cache folder. Are these files ever used?

Thanks.

Hi @ranahanocka
In __getitem__()
meta['edge_features'] = (edge_features - self.mean) / self.std
Which self.mean and self.std are read from 'mean_std_cache.p' file.

How to get such file? How to implement it in my own data without such file?

Thanks

I want to know where is the change of edge feature after edge pooling in the code? Is it in the file mesh_pool.py? I only find the function "def __pool_side(self, mesh, edge_id, mask, edge_groups, side)", but this function didn't involve the change of the edge feature.
Thank you please.

@ranahanocka First of all thanks for this interesting approach of CNN.

Unfortunately I am currently running into this error when trying to train the network on my own dataset. Faces were simplified to 600 with the blender script. Do you have an idea what this could possible causing this? Bad .obj files? Wrong parameters used for training configuration?

Console output:

saving the latest model (epoch 1, total_steps 16)
(epoch: 1, iters: 80, time: 0.045, data: 3.654) loss: 0.672 
Traceback (most recent call last):
  File "train.py", line 23, in <module>
    for i, data in enumerate(dataset):
  File "/kaggle/working/MeshCNN/data/__init__.py", line 33, in __iter__
    for i, data in enumerate(self.dataloader):
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 819, in __next__
    return self._process_data(data)
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 846, in _process_data
    data.reraise()
  File "/opt/conda/lib/python3.6/site-packages/torch/_utils.py", line 385, in reraise
    raise self.exc_type(msg)
ValueError: Caught ValueError in DataLoader worker process 2.
Original Traceback (most recent call last):
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/_utils/worker.py", line 178, in _worker_loop
    data = fetcher.fetch(index)
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/opt/conda/lib/python3.6/site-packages/torch/utils/data/_utils/fetch.py", line 44, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/kaggle/working/MeshCNN/data/classification_data.py", line 27, in __getitem__
    mesh = Mesh(file=path, opt=self.opt, hold_history=False, export_folder=self.opt.export_folder)
  File "/kaggle/working/MeshCNN/models/layers/mesh.py", line 16, in __init__
    fill_mesh(self, file, opt)
  File "/kaggle/working/MeshCNN/models/layers/mesh_prepare.py", line 11, in fill_mesh
    mesh_data = from_scratch(file, opt)
  File "/kaggle/working/MeshCNN/models/layers/mesh_prepare.py", line 61, in from_scratch
    post_augmentation(mesh_data, opt)
  File "/kaggle/working/MeshCNN/models/layers/mesh_prepare.py", line 185, in post_augmentation
    slide_verts(mesh, opt.slide_verts)
  File "/kaggle/working/MeshCNN/models/layers/mesh_prepare.py", line 198, in slide_verts
    if min(dihedral[edges]) > 2.65:
ValueError: min() arg is an empty sequence

Hi,

1. Would like to ask if this could work with a data set that has no defined category (example vases could be mixed in with mugs in the same dataset)
2. Are the number of segmented parts fixed? Or can there be a unlimited number of segments?
3. Can it be trained to segment finer details? Example i have a wooden plank, it will be segmented into 6 different surfaces. And if the wooden plank had holes drilled, the segments would increase by 1 for each hole. (i think this term is called surface segmentation? if it exists)

Regards,
Kessho

Hi, Rana,

Thank you for your excellent work. I found an error when running

bash ./scripts/shrec/train.sh

Below is the program output:

------------ Options -------------
arch: mconvnet
batch_size: 16
beta1: 0.9
checkpoints_dir: ./checkpoints
continue_train: False
dataroot: datasets/shrec_16
dataset_mode: classification
epoch_count: 1
export_folder: 
fc_n: 100
flip_edges: 0.2
gpu_ids: [0]
init_gain: 0.02
init_type: normal
is_train: True
lr: 0.0002
lr_decay_iters: 50
lr_policy: lambda
name: shrec16
ncf: [64, 128, 256, 256]
ninput_edges: 750
niter: 100
niter_decay: 100
no_vis: False
norm: group
num_aug: 20
num_groups: 16
num_threads: 3
phase: train
pool_res: [600, 450, 300, 180]
print_freq: 10
resblocks: 1
run_test_freq: 1
save_epoch_freq: 1
save_latest_freq: 250
scale_verts: False
seed: None
serial_batches: False
slide_verts: 0.2
verbose_plot: False
which_epoch: latest
-------------- End ----------------
loaded mean / std from cache
#training meshes = 480
---------- Network initialized -------------
[Network] Total number of parameters : 1.323 M
-----------------------------------------------
Traceback (most recent call last):
  File "train.py", line 23, in <module>
    for i, data in enumerate(dataset):
  File "/home/maiqi/yalong/project/more-personal/deep-3d/MeshCNN/data/__init__.py", line 33, in __iter__
    for i, data in enumerate(self.dataloader):
  File "/home/maiqi/yalong/software/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 637, in __next__
    return self._process_next_batch(batch)
  File "/home/maiqi/yalong/software/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 658, in _process_next_batch
    raise batch.exc_type(batch.exc_msg)
ValueError: Traceback (most recent call last):
  File "/home/maiqi/yalong/software/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 138, in _worker_loop
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "/home/maiqi/yalong/software/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 138, in <listcomp>
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "/home/maiqi/yalong/project/more-personal/deep-3d/MeshCNN/data/classification_data.py", line 27, in __getitem__
    mesh = Mesh(file=path, opt=self.opt, hold_history=False, export_folder=self.opt.export_folder)
  File "/home/maiqi/yalong/project/more-personal/deep-3d/MeshCNN/models/layers/mesh.py", line 16, in __init__
    fill_mesh(self, file, opt)
  File "/home/maiqi/yalong/project/more-personal/deep-3d/MeshCNN/models/layers/mesh_prepare.py", line 21, in fill_mesh
    mesh2fill.ve = mesh_data['ve']
  File "/home/maiqi/yalong/software/anaconda3/envs/meshcnn/lib/python3.6/site-packages/numpy/lib/npyio.py", line 262, in __getitem__
    pickle_kwargs=self.pickle_kwargs)
  File "/home/maiqi/yalong/software/anaconda3/envs/meshcnn/lib/python3.6/site-packages/numpy/lib/format.py", line 692, in read_array
    raise ValueError("Object arrays cannot be loaded when "
ValueError: Object arrays cannot be loaded when allow_pickle=False

Machine OS: ubuntu 16.04
conda environments:

#
# Name                    Version                   Build  Channel
astroid                   2.2.5                    py36_0  
blas                      1.0                         mkl  
ca-certificates           2019.1.23                     0  
certifi                   2019.3.9                 py36_0  
cffi                      1.12.3           py36h2e261b9_0  
cudatoolkit               9.2                           0  
cudnn                     7.3.1                 cuda9.2_0  
cycler                    0.10.0                   py36_0  
cython                    0.29.7           py36he6710b0_0  
dbus                      1.13.6               h746ee38_0  
expat                     2.2.6                he6710b0_0  
fontconfig                2.13.0               h9420a91_0  
freetype                  2.9.1                h8a8886c_1  
glib                      2.56.2               hd408876_0  
gst-plugins-base          1.14.0               hbbd80ab_1  
gstreamer                 1.14.0               hb453b48_1  
icu                       58.2                 h9c2bf20_1  
intel-openmp              2019.3                      199  
isort                     4.3.17                   py36_0  
jpeg                      9b                   h024ee3a_2  
kiwisolver                1.1.0            py36he6710b0_0  
lazy-object-proxy         1.3.1            py36h14c3975_2  
libedit                   3.1.20181209         hc058e9b_0  
libffi                    3.2.1                hd88cf55_4  
libgcc-ng                 8.2.0                hdf63c60_1  
libgfortran-ng            7.3.0                hdf63c60_0  
libpng                    1.6.37               hbc83047_0  
libstdcxx-ng              8.2.0                hdf63c60_1  
libtiff                   4.0.10               h2733197_2  
libuuid                   1.0.3                h1bed415_2  
libxcb                    1.13                 h1bed415_1  
libxml2                   2.9.9                he19cac6_0  
matplotlib                3.0.3            py36h5429711_0  
mccabe                    0.6.1                    py36_1  
mkl                       2019.3                      199  
mkl_fft                   1.0.12           py36ha843d7b_0  
mkl_random                1.0.2            py36hd81dba3_0  
ncurses                   6.1                  he6710b0_1  
ninja                     1.9.0            py36hfd86e86_0  
numpy                     1.16.3           py36h7e9f1db_0  
numpy-base                1.16.3           py36hde5b4d6_0  
olefile                   0.46                     py36_0  
openssl                   1.1.1b               h7b6447c_1  
pcre                      8.43                 he6710b0_0  
pillow                    6.0.0            py36h34e0f95_0  
pip                       19.1                     py36_0  
protobuf                  3.7.1                    pypi_0    pypi
pycparser                 2.19                     py36_0  
pylint                    2.3.1                    py36_0  
pyparsing                 2.4.0                      py_0  
pyqt                      5.9.2            py36h05f1152_2  
python                    3.6.8                h0371630_0  
python-dateutil           2.8.0                    py36_0  
pytorch                   1.0.1           cuda92py36h65efead_0  
pytz                      2019.1                     py_0  
qt                        5.9.7                h5867ecd_1  
readline                  7.0                  h7b6447c_5  
setuptools                41.0.1                   py36_0  
sip                       4.19.8           py36hf484d3e_0  
six                       1.12.0                   py36_0  
sqlite                    3.28.0               h7b6447c_0  
tensorboardx              1.6                      pypi_0    pypi
tk                        8.6.8                hbc83047_0  
torchvision               0.2.2                      py_3    pytorch
tornado                   6.0.2            py36h7b6447c_0  
typed-ast                 1.3.4            py36h7b6447c_0  
wheel                     0.33.1                   py36_0  
wrapt                     1.11.1           py36h7b6447c_0  
xz                        5.2.4                h14c3975_4  
zlib                      1.2.11               h7b6447c_3  
zstd                      1.3.7                h0b5b093_0 

Any advice? thanks.

Hi Rana,

If I am not wrong, the function has_boundaries determines whether an edge is a boundary edge or not. So why are we checking the connectivity of the adjacent edges of a given edge (having edge_id = edge_id)?

Hi,
I've executed your source code without changing any settings. But, at 200 epoch, I'm getting 97.5% accuracy on shrec16 data set, but you've mentioned 98.6% accuracy in your paper. could you please tell me, is this 200th epoch accuracy or a highest value at any epoch of training model.

Thanks,

i was working with your cube dataset to benchmark my dataset later. Each and every mesh had the same number of triangles and vertices which somehow won't be the case for me. Will it be possible to work on training data with varied number of triangles and vertices ?

Hi ranahanocka,

Cheers for the creativity. I am sorry if here is not the right place to ask the questions. I am looking at the code:

1. in the implementation, inside the build_gemm function, what does the "sides" list here mean? Why do we need this information?
2. After unpooling of the mesh, are the mesh exactly the same as the original mesh, i mean the vertex position as well as the mesh connectivity? Will they be different? How can i export the reconstructed mesh?

Thanks & Regards,
zg

Hi Rana,

first of all, thanks for this contribution. It really looks promising!

I am working with your code in order to analyze a set of meshes formed by molecular shapes. The objects I work with are all genus 0 manifolds (i.e. no holes). I am running your code with segmentation options (although I don't really have any segments defined).

I could run the code with some dummy examples. But, when I try to extend it to a more general set, I run into the following error:

Traceback (most recent call last):
  File "train.py", line 36, in <module>
    model.optimize_parameters()
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/mesh_classifier.py", line 72, in optimize_parameters
    out = self.forward()
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/mesh_classifier.py", line 59, in forward
    out = self.net(self.edge_features, self.mesh)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/networks.py", line 209, in __call__
    return self.forward(x, meshes)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/networks.py", line 202, in forward
    fe, before_pool = self.encoder((x, meshes))
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/networks.py", line 363, in __call__
    return self.forward(x)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/networks.py", line 347, in forward
    fe, before_pool = conv((fe, meshes))
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/networks.py", line 228, in __call__
    return self.forward(x)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/networks.py", line 250, in forward
    x2 = self.pool(x2, meshes)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/layers/mesh_pool.py", line 21, in __call__
    return self.forward(fe, meshes)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/layers/mesh_pool.py", line 34, in forward
    self.__pool_main(mesh_index)
  File "/Users/ismael.gomez/Documents/Software/MeshCNN/MeshCNN-unsupervised v2/models/layers/mesh_pool.py", line 50, in __pool_main
    value, edge_id = heappop(queue)
IndexError: index out of range

Adding some marks within your code, I could identify that his happens in cases where there are -1's in the gemm-edges structure, but I don't understand why is this. My surfaces are all (allegedly) closed, so no boundary edges exist.

Do you have any idea of why this is happening?

Thanks in advance,
Ismael.

Hi, first of all, thanks for sharing this code public!

I've discovered a problem with the README. It says it works with pytorch 1.0, but it won't because it is using torch.BoolTensor in model/layers/mesh.py. torch.BoolTensor is supported since 1.2.0.

I confirmed the test code for scripts/human_seg is working fine with pytorch 1.2.0. I hope this information helps other guys testing this code.

I want to know that what does sides (which defined in file 'mesh_prepare.py') stand for? Thank you

Can you please explain what self.sides actually signify? It can be found in mesh_prepare.py

hi,
While running for split 10 classification on SHREC do we need to change the batch_size to 10 or there are some other settings? shrec16 has 16 training samples for each class, do we need shrec10 of 10 samples for each class?

Kind regards,
Rana Kamran

@ranahanocka
Hi,

Once we segment the object, how can we view each segmented component independently ?

Hi,

The images regarding visualization of 3d objects in your paper are elgent. Can you talk about how did you draw them?

Thanks.

bash ./scripts/shrec/train.sh
THCudaCheck FAIL file=/opt/conda/conda-bld/pytorch_1549630534704/work/torch/csrc/cuda/Module.cpp line=34 error=35 : CUDA driver version is insufficient for CUDA runtime version
Traceback (most recent call last):
File "train.py", line 9, in
opt = TrainOptions().parse()
File "/home/peichen/python_test/MeshCNN/options/base_options.py", line 54, in parse
torch.cuda.set_device(self.opt.gpu_ids[0])
File "/home/peichen/anaconda3/envs/meshcnn/lib/python3.6/site-packages/torch/cuda/init.py", line 264, in set_device
torch._C._cuda_setDevice(device)
RuntimeError: cuda runtime error (35) : CUDA driver version is insufficient for CUDA runtime version at /opt/conda/conda-bld/pytorch_1549630534704/work/torch/csrc/cuda/Module.cpp:34

I attempted to use the pre-trained network on personally generated data and the segmentation results were quite poor. The mesh is decimated down to ~1500 verts and 3000 faces, manifold, and watertight. I attached the OBJ (inside zip) I used.
mesh.zip

So I'm trying to use MeshCNN on some quite large meshes, around ~10000 faces, ~16000 edges. When running the train script however, I get the following error:

File "C:\MeshCNN\models\layers\mesh_pool.py", line 50, in __pool_main
    value, edge_id = heappop(queue)
IndexError: index out of range

It seems that the queue of edges runs out of edges before it has reached the number of edges required by the pool_res argument? Is this correct?

I'm also guessing this is happening because __pool_edge in the same file returns false for too many of the edges in my mesh.

Does anyone have any idea how I might rectify this? Is there a way to make sure that the queue does not run out of edges so that I can continue removing edges, regardless of how many edges I started with?

I want to generate .seg files so that I can use my data to train this network. Could you please suggest any software that is better for this issue?

Running the blender script, I sometimes get the following messages on some files:

Info: Applied modifier was not first, result may not be as expected
Info: Applied modifier was not first, result may not be as expected
Info: Applied modifier was not first, result may not be as expected
Info: Applied modifier was not first, result may not be as expected

Can this be ignored?

Some of the larger datasets out there (such as Shapenet and ABC) have edge counts ranging from 100's to 100000's.

Do you have any advice on selecting a pool_res and ncf for datasets with this sort of edge count variation? And/or any thoughts on the effect this would have on your research?

I used classification pretrained parameters, the test works well for edges 750, but then I subdivide the mesh to 3000 edges(apply 1 loop, still manifold), and set --pool_res 2500 2000 1500 1000, error occurs below, do I have to use the fixed edges(like 750) as input? How to deal with dense meshes, do we need to retrain it?

Running Test 
loaded mean / std from cache  
loading the model from ./checkpoints\mytest\latest_net.pth  
Traceback (most recent call last):  
  File "test.py", line 25, in <module>  
    run_test()  
  File "test.py", line 16, in run_test  
    for i, data in enumerate(dataset):  
  File "E:\Github\MeshCNN\data\__init__.py", line 33, in __iter__
    for i, data in enumerate(self.dataloader):
  File "D:\DevTool\anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 582, in __next__
    return self._process_next_batch(batch)
  File "D:\DevTool\anaconda3\lib\site-packages\torch\utils\data\dataloader.py", line 608, in _process_next_batch
    raise batch.exc_type(batch.exc_msg)
ValueError: Traceback (most recent call last):
  File "D:\DevTool\anaconda3\lib\site-packages\torch\utils\data\_utils\worker.py", line 99, in _worker_loop
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "D:\DevTool\anaconda3\lib\site-packages\torch\utils\data\_utils\worker.py", line 99, in <listcomp>
    samples = collate_fn([dataset[i] for i in batch_indices])
  File "E:\Github\MeshCNN\data\classification_data.py", line 31, in __getitem__
    edge_features = pad(edge_features, self.opt.ninput_edges)
  File "E:\Github\MeshCNN\util\util.py", line 22, in pad
    return np.pad(input_arr, pad_width=npad, mode='constant', constant_values=val)
  File "D:\DevTool\anaconda3\lib\site-packages\numpy\lib\arraypad.py", line 1172, in pad
    pad_width = _as_pairs(pad_width, narray.ndim, as_index=True)
  File "D:\DevTool\anaconda3\lib\site-packages\numpy\lib\arraypad.py", line 949, in _as_pairs
    raise ValueError("index can't contain negative values")
ValueError: index can't contain negative values