ajbrock / generative-and-discriminative-voxel-modeling Goto Github PK

Great work! A small request:

My understanding is that cuDNN is just for a speed up. Could this be made an optional dependency? (I don't have a nVidia GPU).

Hi, I'm trying to run your classification code. Is the current configuration in VRN.py correct? Now the batch size is 1 (which makes it more than 2 hours for training one epoch), and the learning rate is 0.002 and 0.0002 after 12 epochs. What's the batch size and learning rate and training epochs you used in your paper? I couldn't find it in the paper. Thanks very much!

Hello,

Thank you for your kind replies.

I see your pre-trained models work perfectly

I made a simple classifier using the VRN model. (https://www.youtube.com/watch?v=a246GAffWZk)

For now, I want to re-train this network using my custom data. (greyscaled dicom volume)

Before that, I just tried to run the train.py script using the initial configuration here.

I had an error and crashed when first checkpoint was being saved.

The problem is, it took almost 48 hours to get this error.

That means, it takes about 48 hours for training a single epoch, and the configuration shows that I need 250 epoches to get the whole training to be done.

I think I need about 500 days to finish the training script using my computer..

My computer specification is :
Windows 10 x64 OS
Intel(R) Core(TM) i7-2600 CPU @ 3.40GHZ x 2 CPU
8.00GB RAM
NVIDIA GeForce GTX 1060 6GB graphics

If you don't mind, could you let me know your training time and computer specification when you trained your VRN network?, and are 250 epoches mandatory to expect the performance as the VRN model here?

Or,, is there any other configuration(like batch size, I do not know much about this.. the current batch size of VRN.py here is 1) that might be able to reduce the training time?

Thank you, and please excuse my English.

I have added a pytorch implementation of the voxception-resnet. It might be useful for others.

I'm trying to run the train.py file of the Discriminative folder, but a problem arises at "from utils import checkpoints, metrics_logging", the problem is " ImportError: cannot import name checkpoints ".

Does anyone know how to fix it ?

Hi guys,

I'm trying to get the GUI to run with following setup:

Windows 10
Python 2.7
VTK 7.0

Inside GUI.py, when the following line is called

renderWin.Render()

I get the following output from the output window of VTK:

The rendering window opens up but immediatelly crashes after that.

Hi,
I am getting an error whenever trying to run the GUI or train the VAE. I can understand this issue is related to CUDA but I would like to confirm the exact error and would appreciate any comments to help the problem solved. The error is following -

ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -O3 -m64 -Xcompiler -DCUDA_NDARRAY_CUH=mc72d035fdf91890f3b36710688069b2e,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,-fPIC,-fvisibility=hidden -Xlinker -rpath,/home/tansin/.theano/compiledir_Linux-4.15--generic-x86_64-with-Ubuntu-18.04-bionic-x86_64-3.6.7-64/cuda_ndarray -I/usr/lib/python3/dist-packages/theano/sandbox/cuda -I/home/tansin/.local/lib/python3.6/site-packages/numpy/core/include -I/usr/include/python3.6m -I/usr/lib/python3/dist-packages/theano/gof -L/usr/lib -o /home/tansin/.theano/compiledir_Linux-4.15--generic-x86_64-with-Ubuntu-18.04-bionic-x86_64-3.6.7-64/cuda_ndarray/cuda_ndarray.so mod.cu -lcublas -lpython3.6m -lcudart')
Traceback (most recent call last):
File "Generative/train_VAE.py", line 522, in
main(args)
File "Generative/train_VAE.py", line 336, in main
config_module = imp.load_source('config', args.config_path)
File "/usr/lib/python3.6/imp.py", line 172, in load_source
module = _load(spec)
File "", line 684, in _load
File "", line 665, in _load_unlocked
File "", line 674, in exec_module
File "", line 786, in get_code
File "", line 503, in _code_to_bytecode
ValueError: unmarshallable object

Hello, I seek your advice

Could be there any problem if I re-train this network using higher dimension , such as 256x256x256

or 500x500x20?

Furthermore, can grayscale(not -1,3 binary) image volume also be used?

Hi, I am trying to re-implement your nice work in tensorflow and I encounter a question.
I need to use make_mats.m to generate 24 rotations, but it seems missing some dependencies, which must be found in princeton toolkits, especially read_off.m, would you provide more details about it?
best

Currently, I am using

Windows 10
Python 3.6
VTK 7.1.0

If i try to run GUI.py, I get Error from

config_module = imp.load.source('config', args.config_path)

saying "unmarshallable object"

which version of theano and lasagne is the code created?