This repo contains the tutor code for tensorflow beginners. The documentation is at this link.
The test environment is
python 2.7.14
tensorflow 1.5.0
Clone or fork the code to your own machine. Type in
python mnist.py
It will automatically train the model and print out loss values and test accuracies. Feel free to change the code and play around with it.
Have fun!
- Try with the default environment. It should work. There is one Tesla p100 on dll-00, and four GTX 1080 Ti on dll-01. The CUDA version is 9.0 on both machines.
python 2.7.14
tensorflow 1.6.0
CUDA 9.0
- If the tutor code can not run in the default environment, go through the following steps.
Step 1: Connect to the GPU server.
Type the following command in the terminal. Replace <username> with your own username. It should be your Cornell ID.
ssh <username>@dll-00.ece.cornell.edu
or
ssh <username>@dll-01.ece.cornell.edu
Step 2: Configure the environment. Run the following commands to configure environment variables:
export PATH=/usr/local/cuda-9.0/bin:${PATH}
export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64:${LD_LIBRARY_PATH}
Step 3: Run the example.
Now, if you want to run some example codes, they are at /usr/local/cuda-9.0/samples/
Copy the samples to your personal folder first:
cp -rp /usr/local/cuda-9.0/samples/ ~/cuda_samples/
Then, cd to a folder that contains a specific example you want to run.
cd ~/cuda_samples/<path_to_the_sample>
A good one to start is deviceQuery, which shows you the properties of the GPU and whether it's working properly. Find it at ~/cuda_samples/1_Utilities/deviceQuery/, and type in
make
The example will be compiled. Run the excutable file using
./<name_of_the_binary>
If everything works fine, CUDA should be working. With the default tensorflow package and python version 2.7.14, you should now be able to run the tutor code.
- Yichi Zhang - Web