error importation cudnn

[theoeiferman]

when launching :

th src/train.lua -phase train -gpu_id 1
-model_dir model
-input_feed -prealloc
-data_base_dir data/sample/images_processed/
-data_path data/sample/train_filter.lst
-val_data_path data/sample/validate_filter.lst
-label_path data/sample/formulas.norm.lst
-vocab_file data/sample/latex_vocab.txt
-max_num_tokens 150 -max_image_width 500 -max_image_height 160
-batch_size 20 -beam_size 1

cudnn is not found....
I tried" luarocks install cudnn"
still doesen't work

[da03]

Hmm can you try "require cudnn" in a torch prompt (by command "th") and see if that works?

[theoeiferman]

I have this as an answer, I m not very familiar with lua language

[da03]

sorry I meant entering "th" first,

then in the prompt, enter "require cudnn"

[theoeiferman]

thanks for the advice ! but it still doesn't work . It seems that cutorch is also not install . I tried "luarocks install cutorch " but I have failed building..!

[da03]

Oh I think it seems to be some issue with installing cudnn. Installing torch correctly might be hard, would you mind using docker? Here is a docker file that can be directly used: https://github.com/OpenNMT/OpenNMT/blob/master/Dockerfile

[theoeiferman]

Ok I didn't know about the existence of Docker, thanks for the tips it looks great!
I tried to build an image by copying the content of the dockerfile you sent me but I get "unable to prepare context: context must be a directory: /Users/teiferman27/Dockerfile"
When I tried to launch directly "docker build https://github.com/OpenNMT/OpenNMT/blob/master/Dockerfile#L6"

I have one more question if I succeed to launch this dockerfile. Then I can use the Lua language with file on my computer? or just inside the "container". Thanks you for the advice already !

[da03]

For the first question, I think you need to put the dockerfile inside a folder, then inside this folder do docker build . will generate the image.

For the second question, it allows using Lua inside docker container only.

[da03]

BTW, I think you might need to use nvidia-docker (https://github.com/NVIDIA/nvidia-docker) to support using GPUs inside docker container.

[theoeiferman]

Ok I succeeded to use docker build -t operating_lua .
After running all the afternoon to build the image, I then have tried to launch the command docker run operating_lua but it is just opening and closing on the docker dashboard ..... You think docker doesn't support the container and I need to use nvidia-docker ? thanks for the respond

[da03]

I think it should be nvidia-docker run -it operating_lua /bin/bash, but it might be better to directly check docker documentation.

[theoeiferman]

I wanted to install nvidia-docker but I needed to install NVIDIA driver first. but it seems that this step require Linux operating system and I am on MacOs .... but I was surprised because one good point for docker was that everyone can run it from every operating system !

Then I tried docker run -dit operating_lua and I was able to open the container and read in it :

Moreover when I tried to see if the module 'cudnn' is in the system I get :

I am still confused about how to approach the big picture... Can I import files into the container? Do I really need nvidia-docker ?

Thank you again for your time @da03 .

[da03]

Hmm I suspect that your CUDA driver version might be too outdated (what's the output of nvcc --version and nvidia-smi?), which caused issues both for require cudnn and for installing nvidia-docker. There are actually CUDA drivers available for mac: https://www.nvidia.com/en-us/drivers/cuda/mac-driver-archive/. Fixing the driver version issue might solve all problems.

[theoeiferman]

On the link https://github.com/NVIDIA/nvidia-docker , they talk about Linux

nvcc --version and nvidia-smi are unkown for now but probably because I didn't install nvidia-docker yet ? I am going to install CUDA drivers and then nvidia-docker.

[da03]

Oh no, so it seems nvidia-docker would not work on Mac... I have never used GPUs on Mac, but I think with a proper CUDA installation (https://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html), you should get both nvcc --version and nvidia-smi working.

[theoeiferman]

I can't install on my mac because Nvidia doesen't support mac system anymore.
I think my mac may be it is too old.
I check my graphics on the system information as shown in
https://www.quantstart.com/articles/Installing-Nvidia-CUDA-on-Mac-OSX-for-GPU-Based-Parallel-Computing/
But I don't have NVIDIA graphic card on my computer.

I think the incompatibility is also mentionned there
https://developer.nvidia.com/nvidia-cuda-toolkit-developer-tools-mac-hosts

I am surprised of this CUDA/NVIDIA requirement to use the container though.

[da03]

Oh that explains why: this code (or cudnn) only supports CUDA and cannot run on systems without GPUs. While this version (https://opennmt.net/OpenNMT-py/im2text.html, code can be found at https://github.com/OpenNMT/OpenNMT-py) supports CPU only training, doing so would be extremely slow without the parallelism provided by GPUs. Another way might be using cloud computes such as Amazon EC2 or Google GCE or Microsoft Azure, and rent a GPU instance.

[theoeiferman]

I manage to get another computer but the GPU is AMD Radeon and so I can't use the cudnn module ... I think it should be mentioned on the prerequisites since docker can't solve this hardware issue.

I was about to try CPU but I think that on the link you gave me (https://opennmt.net/OpenNMT-py/im2text.html) there is dependencies like torch vision and pytorch is required ( and so CUDA-enabled GPU are needed again no ? )

I try to follow the steps from https://opennmt.net/OpenNMT-py/im2text.html
but the command onmt_preprocess is not found. There is a step I have missed ?

I will try to use cloud computes probably.

But just to be sure ( correct me if I am wrong) :

• OpenNMT-py project works with pytorch (https://github.com/OpenNMT/OpenNMT-py#requirements)
• this project works with torch ( https://github.com/harvardnlp/im2markup)

[da03]

Yes you are right that OpenNMT-py uses PyTorch and this project uses LuaTorch. PyTorch does not require GPUs (you can do CPU-only installation), but again, it might be extremely slow without using GPUs.

For the onmt_preprocess missing issue, have you installed OpenNMT-py following the instructions here? https://github.com/OpenNMT/OpenNMT-py

[theoeiferman]

I had issues with onmt commands because I use python environment using google collab ( You can activate GPU on the settings and it seems to be a good free solution )
Installing OpenNMT-py with pip instead of clonings the project worked for the onmt command.

Is Using Google-collab a good way to perform GPU calculations ? I am trying to train the model, but it takes a lot of times, do you now how much ?

[da03]

Yeah I think so! The only problem is that the runtime would be disconnected if it's idle for a certain period of time, and the instance would be freed so all progress would be lost. Therefore, you might want to connect to your google drive, and save progress (checkpoints) to your google drive.