Public | Automated Build

Last pushed: 4 months ago
Short Description
Ubuntu Core 14.04 + CUDA + Pycaffe + DIGITS.
Full Description


cuda-digits

Ubuntu Core 14.04 + CUDA 7.5 + cuDNN v5 + Caffe (NVIDIA fork) + DIGITS.

Requirements

Usage

Use NVIDIA Docker: nvidia-docker run -dP kaixhin/cuda-digits.

For automatically mapping the DIGITS server port use nvidia-docker run -dP kaixhin/cuda-digits and docker port <id> to retrieve the port.
For specifying the port manually use nvidia-docker run -d -p <port>:5000 kaixhin/cuda-digits.
The shell can be entered as usual using nvidia-docker run -it kaixhin/cuda-digits bash.

For more information on CUDA on Docker, see the repo readme.

Citation

If you find this useful in research please consider citing this work.

Docker Pull Command
Owner
kaixhin
Source Repository

Comments (5)
reeseak
4 months ago

I did the following to fix the issue people have had in previous comments [ERROR] Train Caffe Model: Check failed: error == cudaSuccess (8 vs. 0) invalid device function:

nvidia-docker run -it -p 5000:5000 kaixhin/cuda-digits bash

cd to ~/caffe/build
make test
make runtest

Should be good to go after that. I tested DIGITS on the mnist dataset. You can get that by doing:
cd ~/digits
python -m digits.download_data mnist ~/mnist

./digits-devserver

follow the DIGITS walkthrough from there and you should be good. I made a commit of the image and use that so I do not have to run the caffe test each time.

rmitch
9 months ago

Hello:

Thanks very much for this very valuable project! I am using your Digits CUDA Docker image for medical imaging research at Mayo Clinic.

However, I have run into the issue mentioned by ten2net earlier:

[ERROR] Train Caffe Model: Check failed: error == cudaSuccess (8 vs. 0) invalid device function

I did some research, and apparently this error indicates that code for my GPU architecture is not present in your Caffe build.

I am using a Pascal GPU, a GTX 1080. This device appears to have an arch code of 61, i.e. The following flags should be passed to nvcc when compiling:

-gencode arch=compute_61,code=compute_61

I checked the build logs for your latest Docker image, and saw the following entries:

-- Automatic GPU detection failed. Building for all known architectures.

-- Added CUDA NVCC flags for: sm_20 sm_21 sm_30 sm_35 sm_50

Would it be possible to add sm_61 to the list of known architectures for your build?

Any help appreciated.

Sincerely,

Ross Mitchell

ten2net
10 months ago

maybe json parse error. UI display raw json

{[ jobs = (jc.model_jobs | filter:search_text | sort_with_empty_at_end:this:jc.storage.show_groups); '' ]}

ex:
Filter
{[enabled = any_selected();'']} {[group_enabled = jc.storage.show_groups && enabled;'']}

ten2net
10 months ago

[ERROR] Train Caffe Model: Check failed: error == cudaSuccess (8 vs. 0) invalid device function

ten2net
10 months ago

not support multi-GPU :[ERROR] Train Caffe Model: USE_NCCL := 1 must be specified for multi-GPU