So I'm trying set run OpenAI gym in a docker container, but it looks like this:
Notice the pong window has a weird render issue where it's repeating things and the colors are off. Here is space invaders:
NOTE FOR "NOT A PROGRAMMING ISSUE" PEOPLE: The solution involves the correct bash script code to call the right API methods to render the arrays of pixels correctly. Also only a graphics programmer is likely to "recognize the render glitch".
My setup is very simple.
- I'm on a local ubuntu 16.04 install with an Nvidia gtx1060 and corei7
- I installed nvida runfile driver with --no-opengl-files (as per instructions from Nvidia and many place).
- Specifically, I'm running floydhub/pytorch docker image.
Does anyone recognize the particular render glitch and what it could mean? It almost looks like a StackOverflow of a frame buffer! What can I do to track down the bug?
EDIT: I have eliminated all the extra dependencies I had been installing and am just doing simple x-forwarding according to the ROS GUI guide.
You can easily reproduce this as follows:
docker run -it --user=$(id -u) --env="DISPLAY" --workdir="/home/$USER" --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" floydhub/pytorch:0.1.11-gpu-py3.6 bash
Now in the image, type python and then the following:
import gym
gym.make('Pong-v0').render()
That should open up an x-forwarded window on your machine, but the display is corrupt (at least for me)
Above I actually used SpaceInvaders-v0
Related
Im working on Ubuntu 20. I've installed docker, nvidia-docker2. On Pycharm, I've followed jetbrain guide, but in the advanced steps it isn't consistent with what I see in my setup. I use PyCharm Proffesional 2022.2.
In this step:
in the run options I put additionally --runtime=nvidia and --gpus=all.
Step 4 finishes as same as in the guide (almost, but it seems that it doesn't bother anything so on that later) and on step 5 I put manually the path to the interpreter in the virtual environment I've created using the Dockerfile.
In that way I am able to run the command of nvidia-smi and see correctly the GPU, but I don't see any packages I've installed during the Dockerfile build.
There is another option to connect the interpreter a little bit differently in which I do see the packages, but I can't run the nvidia-smi command and the torch.cuda.is_availble return False.
The way is instead of doing this as in the guide:
I press on the little down arrow in left of the Add Interpreter button and then click on Show all:
After which I can press the + button :
works, so it might be PyCharm "Python Console" issue.
and then I can choose Docker:
which will result in the difference mentioned above in functionality and also in the path dispalyed (the first one is the first remote interpreter top to bottom direction and the second is the second correspondingly):
Here of course the effect of the first and the second correspondingly:
Here is the results of the interpreter run with the first method connected interpreter:
and here is the second:
Of the following code:
Here is the Dockerfile file if you want to take a look:
Anyone configured it correctly and can help ?
Thank you in advance.
P.S: if I run the docker from services and enter the terminal the command nvidia-smi works fine and also the import of torch and the command torch.cuda.is_available return True.
P.S.2:
The thing that has worked for me for now is to change the Dockerfile to install directly torch with pip without create conda environement.
Then I set the path to the python2.7 and I can run the code, but not debug it.
for run the result is as expected (the packages list as was shown before is still empty, but it works, I guess somehow my IDE cannot access the packages list of the remote interpreter in that case, I dont know why):
But the debugger outputs the following error:
Any suggestions for the debugger issue also will be welcome, although it is a different issue.
Please update to 2022.2.1 as it looks like a known regression that has been fixed.
Let me know if it still does not work well.
Been playing with pyspark on juypter all day with no issues. Just by simply using the docker image juypter/pyspark-notebook, 90% of everything I need is packaged (YAY!)
I would like to start exploring using GraphFrames, which sits on top of GraphX which sits on top of Spark. Has anyone gotten this combination to work?
Essentially, according to the documentation, I just need to pass "--packages graphframes:xxyyzz" when running pyspark to download and run graphframes. Problem is that juypter is already running as soon as the container comes up.
I've tried passing the "--packages" line as an environment variable (-e) for both JUYPTER_SPARK_OPTS and SPARK_OPTS when running docker run and that didn't work. I found that I can do pip install graphframes from a terminal, which gets me part of the way -- the python libraries are installed, but the java ones are not "java.lang.ClassNotFoundException: org.graphframes.GraphFramePythonAPI".
The image specifics documentation does not appear to offer any insights on how to deploy a Spark Package to the image.
Is there a certain place to throw the graphframes .jar? Is there a command to install a spark package post-docker? Is there a magic argument to docker run that would install this?
I bet there's a really simple answer to this --Or am I in high cotton here?
References:
No module named graphframes Jupyter Notebook
How do I run pyspark with jupyter notebook?
So the answer was quite simple:
From the gist here, we need to simply tell juypter to add the --packages line to the SPARK_SUBMIT with something like this to the top of my notebook. Spark goes out and installs the package when grabbing the context:
import os
os.environ['PYSPARK_SUBMIT_ARGS'] = '--packages graphframes:graphframes:0.8.1-spark3.0-s_2.12 pyspark-shell'
Keep a watch on the versions available at the graphframes package, which for now, means graphframes 0.8.1 on spark 3.0 on scala 2.12.
Jupyter Lab application features nice Terminas with in-browser terminal shell that support colours, navigation keys, and pretty much all standard features of a terminal application. In this question I mean /lab app, not classic Notebook (/tree) app.
If I launch Jupyter server using this Docker image it works great. I need to build my own image, and preferably not based on that. I do it simply as documented:
docker run -it --rm -p 8888:8888 -v "$PWD":/jupyter python:3.8 bash
# pip install jupyterlab
# jupyter lab --config=/jupyter/jupyter_notebook_config.py
The above is assuming I have a jupyter_notebook_config.py in the current directory:
c.NotebookApp.ip = '0.0.0.0'
c.NotebookApp.port = 8888
c.NotebookApp.password = 'sha1:<salt>:<hash>'
c.NotebookApp.allow_password_change = False
c.NotebookApp.allow_root = True
c.NotebookApp.open_browser = False
Everything works, but Terminal performs very poor, it does not support colours and send codes (like ^[[A, ^[[B) instead of arrow keys. Line-by-line investigation of the Dockerfile is not so exciting endeavour, may be somebody can point me to what I am mising?
EDIT: I was little bit wrong about colours (was confused by the default green prompt in the jupyter/base-notebook image) and the overall issue description. The root cause was that the shell that is started in my image is sh while in the official image it is bash. But nevertheless Terminal is not fully functional, e.g. if I launch nano, it starts only in 80x25 characters area and does not stretch to the actual size of the terminal). Arrows work in nano though.
So I figured out the reason. Apparently the Terminal web app just replicates the behaviour of the default shell of the user under which Jupyter is run. In this image they enable colouring in .bashrc template and then create a new user specifying a shell for him (lines 52 and 59).
EDIT: also SHELL=/bin/bash must be set in environment.
I'm trying to learn Pytorch, but whenever I seem to try any online tutorial (https://pytorch.org/tutorials/beginner/blitz/tensor_tutorial.html#sphx-glr-beginner-blitz-tensor-tutorial-py), I get errors when trying to run certain functions, but only in Jupyter Notebook.
When running
x = torch.empty(5, 3)
I get an error:
module 'torch' has no attribute 'empty'
Furthermore, when running
x = torch.zeros(5, 3, dtype=torch.long)
I get the error:
module 'torch' has no attribute 'long'
Some other functions work fine like:
x = torch.rand(5, 3)
But generally, most code I try to run seems to run into an error really quickly. I couldn't find any resolution online.
When I go into my docker container and simply run python in the shell, I can run these lines just fine with no errors.
I'm running pytorch in a Docker image that I extended from a fastai image, as it already included things like jupyter notebook and pytorch. I used anaconda to update everything, and committed it to a new image for myself.
I have absolutely no idea what the issue could be. I've tried updating packages through anaconda, pip, aptitude in my docker container, and making sure to commit my changes, but nothing seems to work. I also tried creating a new kernel with python 3.7 as I noticed that my Jupyter Notebook only runs in 3.6.4, and when I run python in the shell it is at 3.7.
I've also tried getting different docker images and extending them with what I need, but all images that I've tried have had errors with anaconda where it gets stuck on "Solving environment" step.
Ok, so the fix for me was to either update pytorch through conda using the following command
conda update pytorch
If it's not installed yet, I've gotten it to work in other environments by simply installing it through conda
conda install pytorch
Kind of stupid that I didn't try this earlier, but I was confused on the difference between conda and pip.
I would like to build a Dockerfile in linux which
1. compiles vim with python
2. installs python stack (such as numpy, scipy, ipython, etc)
3. creates ssl certificate for ipython-notebook, to view the notebooks on host machine
It seemed straightforward enough. But I have run into problems despite a variety of approaches, such as linking separate containers, using anaconda, as well as with a single unified image vs separate layers, or creating a user or running all as a root.
In order to run vim, simply installing to root, does not activate pathogen bundle/vim-ipython. Creating a user allows pathogen bundles (ie nerdtree works) to install, but :IPython throws error.
:IPython failed
^-- failed '' not found .
Ive tried the above with no layers/1 large Dockerfile, and with different layers for the python stack, vim, and the ipython notebook.
Dockerfile
What am I not seeing here ?
what does the ^-- failed '' not found referring to?
Ive tried running the ipython notebook using --no-browser & and then running vim, or using running two shells on the same container... but cant get past this error.
Here is a working Dockerfile for anyone trying to get vim-ipython working in Docker.
issues:
user/shared home needed to for vim, despite runtimepath in .vimrc to pathogen/bundle
%connect_info >> required with containers
I am running in root, not sure why vim required a USER to install packages, but changing to USER would throw errors with CMD
--best