Tensorflow and Keras for the Nvidia Geforce GT 710

 

... alas! either the locks were too large, or the key was too small, but at any rate it would not open any of them. 

Seemingly against the odds, CUDA and cudaDNN ran on the GT 710, and I could run an AI super resolution inference program to upscale images and video. While it is gratifying to finally bump up the GPU temperature, it hardly broke a sweat, wandering from 38 degrees Celsius to 40, more from the time of day than from workload. After all, my Raspberry Pi could do the same.

Training an AI might stretch it a little more, one of the biggest and baddest of them all, an SRGAN might make an impression. There seems to be two main frameworks, Pytorch and Tensorflow. A very cursory search shows that Pytorch may require my Ubuntu 18.04 python 3.6 to be first upgraded to 3.8. This is quite possible, but having spent 3 weeks building python 3.6 for CUDA I decided it might be time to try Tensorflow.

There is the usual dilemma of juggling Tensorflow, CUDA, gcc and python versions, but using Fan Leng-Yoon and the tensorflow sites, I settled on Tensorflow 2.2.0.

The instructions culled from tensorflow site are:

$sudo apt-get update

$sudo pip3 install "tensorflow==2.2.*"

$sudo pip3 install keras

It installed surprisingly smoothly, except when it was time for the tensorflow test:

$python3

Python 3.6.9 (default, Mar 10 2023, 16:46:00)

[GCC 8.4.0] on linux

Type "help", "copyright", "credits" or "license" for more information.

>>> import tensorflow as tf

Illegal instruction (core dumped)

That was not good. A hint came from the tensorflow repository: I may be missing the AVX instruction. And indeed my CPU, an Athlon II X3 440 did not have it

$cat /proc/cpuinfo

wp              : yes

flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov

pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp

 lm 3dnowext 3dnow constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid pni

monitor cx16 popcnt lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalign

sse 3dnowprefetch osvw ibs skinit wdt nodeid_msr hw_pstate vmmcall npt lbrv svm_

lock nrip_save

bugs            : tlb_mmatch fxsave_leak sysret_ss_attrs null_seg spectre_v1 spectre_v2

bogomips        : 6020.19

TLB size        : 1024 4K pages

clflush size    : 64

cache_alignment : 64

address sizes   : 48 bits physical, 48 bits virtual

power management: ts ttp tm stc 100mhzsteps hwpstate

The obvious thing to do would be to downgrade tensorflow to version 1.5, before the use of the AVX instructions was baked into the tensorflow binaries.

$sudo pip3 uninstall tensorflow

$sudo pip3 install "tensorflow_gpu==1.5.*"

But now I get a different failure:

$python3

Python 3.6.9 (default, Mar 10 2023, 16:46:00)

[GCC 8.4.0] on linux

Type "help", "copyright", "credits" or "license" for more information.

>>> import tensorflow as tf

ImportError: libcublas.so.9.0: cannot open shared object file: No such file or directory

 It looks like it wants an older CUDA:

$sudo ls -lR /usr/ | grep -e libcublas

[sudo] password for heong:

lrwxrwxrwx  1 root root        15 Aug 10  2019 libcublas.so -> libcublas.so.10

lrwxrwxrwx  1 root root        23 Aug 10  2019 libcublas.so.10 -> libcublas.so.1

0.2.1.243

-rw-r--r--  1 root root  62459056 Aug 10  2019 libcublas.so.10.2.1.243

I guess I will be needing to build tensorflow without the AVX instruction.

$sudo -H pip3 uninstall tensorflow_gpu

 To build from source I used the tensorflow instructions, which called for first installing an enormous installer, Bazel. And since I recently got chatGPT why not give it a try:

chatGPT convincingly gave the wrong answer

chatGPT very convincingly gave the wrong answer, version 0.26.0. The correct answer is 3.1.0. When I pointed this out it immediately gave another equally convincing (and correct) answer:

chatGPT quickly changed its mind to version 3.1.0

For now chatGPT 3 seems to have the credibility of a used-car salesman. They say an SRGAN hallucinates all those extra pixels in an up-scaled image.  ChatGPT is known to a few flights of fancy like ... the Mad Hatter?

“Have I gone mad? I'm afraid so.You're entirely Bonkers.But I will tell you a secret,All the best people are.”

$sudo pip3 uninstall keras

$sudo apt-get update
$sudo apt-get install curl gnupg
$curl -fsSL https://bazel.build/bazel-release.pub.gpg | gpg --dearmor > bazel.gp

$sudo mv bazel.gpg /etc/apt/trusted.gpg.d/
$echo "deb [arch=amd64] https://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list

$cat  /etc/apt/sources.list.d/bazel.list
deb [arch=amd64] https://storage.googleapis.com/bazel-apt stable jdk1.8

$sudo apt-get install bazel-2.0.0

$bazel --version
bazel 2.0.0

git clone https://github.com/tensorflow/tensorflow.git

$cd tensorflow

$git checkout v2.2.0
$./configure

Extracting Bazel installation...
You have bazel 2.0.0 installed.
Please specify the location of python. [Default is /usr/bin/python3]:


Found possible Python library paths:
 /usr/lib/python3/dist-packages
 /home/heong/opencv_build/opencv/build/lib/python3/
 /usr/local/lib/python3.6/dist-packages
Please input the desired Python library path to use.  Default is [/usr/lib/python3/dist-packages]
/usr/local/lib/python3.6/dist-packages
Do you wish to build TensorFlow with OpenCL SYCL support? [y/N]:
No OpenCL SYCL support will be enabled for TensorFlow.

Do you wish to build TensorFlow with ROCm support? [y/N]:
No ROCm support will be enabled for TensorFlow.

Do you wish to build TensorFlow with CUDA support? [y/N]: y

Do you wish to build TensorFlow with TensorRT support? [y/N]:
No TensorRT support will be enabled for TensorFlow.

Found CUDA 10.1 in:
   /usr/local/cuda/lib64
   /usr/local/cuda/include
Found cuDNN 7 in:
   /usr/lib/x86_64-linux-gnu
   /usr/include

Do you want to use clang as CUDA compiler? [y/N]:

nvcc will be used as CUDA compiler.

$bazel build --config=opt --config=cuda --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0"//tensorflow/tools/pip_package:build_pip_package

$sudo link /usr/bin/python3 /usr/bin/python

$bazel build  //tensorflow/tools/pip_package:build_pip_package

The build took the Athlon all night but completed successfully.

With a little bit of help from Isaac Lascasas, 

$./tensorflow/tools/pip_package/build_pip_package.sh /tmp/tensorflow_pkg

$pip3 install --upgrade --force-reinstall /tmp/tensorflow_pkg/tensorflow-2.2.0-cp36-cp36m-linux_x86_64.whl

And it worked:

$cd ..
$python3
Python 3.6.9 (default, Mar 10 2023, 16:46:00)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorflow as tf
>>> tf.__version__
'2.2.0'

$sudo pip3 install keras

Training HasnainRaz's Fast-SRGAN bumped the GT 710's temperature up to 60 degrees Celsius. Running inference on it on several hundred frames raised it further to 67. Somehow that felt more like real work.

Happy Trails.