AI-framework

「AI-framework」って何？って言われても困るのだが、
「ソフトウェアを開発するうえでの典型的・定型的な処理やデータ構造などをまとめた、枠組みやテンプレートとなるもの」(wikiのフレームワークから)
なそうな.

っでAI開発向けのライブラリ/フレームワークとしては

TensorFlow https://www.tensorflow.org/?hl=ja
PyTorch https://pytorch.org/
NNabla(Neural Network Libraries) https://github.com/sony/nnabla
Theano https://github.com/Theano/Theano
CNTK (Microsoft Cognitive Toolkit) https://github.com/microsoft/CNTK Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit
Keras https://keras.io/
MXnet (Apache MXNet,活動休止中？) https://mxnet.apache.org/versions/1.9.1/

とかなと.

構築†

ここでは pyenv を敷いてその上に anaconda や必要に応じて miniforge を入れ、さらにその上に python 実行環境を用意しています。

[root@rockylinux9 ~]# cat /etc/redhat-release
Rocky Linux release 9.4 (Blue Onyx)
 
[root@rockylinux9 ~]# git clone https://github.com/yyuu/pyenv.git /apps/pyenv
 
[root@rockylinux9 ~]# export PYENV_ROOT=/apps/pyenv
[root@rockylinux9 ~]# export PATH=$PYENV_ROOT/bin:$PATH
 
[root@rockylinux9 ~]# pyenv install --list     <--これでanaconda,minicondaのバージョンが一覧されます.
 
[root@rockylinux9 ~]# pyenv install anaconda3-2024.06-1
 
[root@rockylinux9 ~]# pyenv global anaconda3-2024.06-1
 
[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda env list
# conda environments:
#
base                     /apps/pyenv/versions/anaconda3-2024.06-1
 
[root@rockylinux9 ~]#

っと下地を作ってから各フレームワーク向けの python 実行環境を用意してみます

一応GPU持ちなマシンです

[root@rockylinux9 ~]# cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX x86_64 Kernel Module  535.179  Fri Apr 26 21:43:18 UTC 2024
GCC version:  gcc version 11.4.1 20231218 (Red Hat 11.4.1-3) (GCC)
[root@rockylinux9 ~]#

↑

TensorFlow†

機械学習ライブラリ
keras2も同封される.

[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda create -n tf tensorflow
 
[root@rockylinux9 ~]# conda activate tf
 
(tf) [root@rockylinux9 ~]# conda list
 :
keras                     2.12.0          py311h06a4308_0
keras-preprocessing       1.1.2              pyhd3eb1b0_0
 :
python                    3.11.5               h7a1cb2a_0
 :
tensorflow                2.12.0          mkl_py311h34a0fa1_0
 :

...あらtensorflowがmklでインストールされた。なのでgpu版に差し替えします

(tf) [root@rockylinux9 ~]# conda search tensorflow
 :
tensorflow                    2.12.0 gpu_py311h65739b5_0  pkgs/main
 :
tensorflow                    2.12.0 mkl_py311h34a0fa1_0  pkgs/main
 :
(tf) [root@rockylinux9 ~]# conda install tensorflow=2.12=gpu*

っでGPUチェック

[illya@rockylinux9 ~]$ source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
[illya@rockylinux9 ~]$ conda activate tf
(tf) [illya@rockylinux9 ~]$ python
Python 3.11.9 (main, Apr 19 2024, 16:48:06) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
 
>>> from tensorflow.python.client import device_lib
2024-08-08 00:56:12.151689: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
 
>>> device_lib.list_local_devices()
 :
2024-08-08 00:56:21.938778: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1635] Created device /device:GPU:0 with 4094 MB memory:  -> device: 0, name: NVIDIA RTX A2000, pci bus id: 0000:13:00.0, compute capability: 8.6
[name: "/device:CPU:0"
device_type: "CPU"
memory_limit: 268435456
locality {
}
incarnation: 5687668420694823401
xla_global_id: -1
, name: "/device:GPU:0"
device_type: "GPU"
memory_limit: 4293787648
locality {
  bus_id: 1
  links {
  }
}
incarnation: 10836478890832035215
physical_device_desc: "device: 0, name: NVIDIA RTX A2000, pci bus id: 0000:13:00.0, compute capability: 8.6"
xla_global_id: 416903419
]
>>> quit();
(tf) [illya@rockylinux9 ~]$

↑

PyTorch†

機械学習ライブラリ

[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda create -n pt pytorch
 
[root@rockylinux9 ~]# conda activate pt
 
(pt) [root@rockylinux9 ~]# conda list
 :
cudatoolkit               11.8.0               h6a678d5_0
cudnn                     8.9.2.26               cuda11_0
 :
python                    3.11.9               h955ad1f_0
pytorch                   2.3.0           gpu_cuda118py311hd2d20a8_100
 :
(pt) [root@rockylinux9 ~]#

っでGPUテスト

[illya@rockylinux9 ~]$ source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
[illya@rockylinux9 ~]$ conda activate pt
(pt) [illya@rockylinux9 ~]$ 
(pt) [illya@rockylinux9 ~]$ python
Python 3.11.9 (main, Apr 19 2024, 16:48:06) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
 
>>> import torch
>>> print(torch.cuda.is_available())
True
>>> print(torch.cuda.get_device_name())
NVIDIA RTX A2000
>>> quit();
(pt) [illya@rockylinux9 ~]$

↑

NNabla(Neural Network Libraries)†

pipで簡単に入れられそう.
https://pypi.org/project/nnabla/を読むと python3.5以上が必要なようで3.10でもよさげそれ以上はまだ未対応の様子

[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda create -n NNabla python=3.10 cuda-python=12 libcublas libcusolver libcurand libcufft cudnn=8 nccl -c nvidia -c conda-forge
 
[root@rockylinux9 ~]# conda activate NNabla
(NNabla) [root@rockylinux9 ~]# conda list
 :
cuda-cudart               12.4.127                      0    nvidia
cuda-nvrtc                12.4.127                      0    nvidia
cuda-python               12.6.0          py310h7ffcb74_0    nvidia
cuda-version              11.1                 hdbd7af8_3    conda-forge
cudatoolkit               11.1.74              h6bb024c_0    nvidia
cudnn                     8.0.4                cuda11.1_0    nvidia
 :
nccl                      2.22.3.1             hee583db_1    conda-forge
 :
pip                       24.2               pyhd8ed1ab_0    conda-forge
python                    3.10.14         hd12c33a_0_cpython    conda-forge
 :
(NNabla) [root@rockylinux9 ~]# pip install nnabla
 
(NNabla) [root@rockylinux9 ~]# python -c "import nnabla"
2024-08-09 02:22:58,917 [nnabla][INFO]: Initializing CPU extension...
 
(NNabla) [root@rockylinux9 ~]#
 
(NNabla) [root@rockylinux9 ~]# pip install nnabla-ext-cuda120
(NNabla) [root@rockylinux9 ~]# python
Python 3.10.14 | packaged by conda-forge | (main, Mar 20 2024, 12:45:18) [GCC 12.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import nnabla_ext.cudnn
2024-08-09 03:02:12,603 [nnabla][INFO]: Initializing CPU extension...
2024-08-09 03:02:12,759 [nnabla][INFO]: Initializing CUDA extension...
openmpi library is not found, distribution training feature is disabled.
2024-08-09 03:02:12,768 [nnabla][INFO]: Initializing cuDNN extension...
>>>
>>> device_id = '0'
>>> nnabla_ext.cudnn.check_gpu(device_id)
>>>
>>> quit();
(NNabla) [root@rockylinux9 ~]#
 
(NNabla) [root@rockylinux9 ~]# module load mpi/openmpi-x86_64
(NNabla) [root@rockylinux9 ~]# python
>>> import nnabla_ext.cudnn
2024-08-09 03:11:55,728 [nnabla][INFO]: Initializing CPU extension...
2024-08-09 03:11:55,884 [nnabla][INFO]: Initializing CUDA extension...
2024-08-09 03:11:55,894 [nnabla][INFO]: Initializing cuDNN extension...
>>> device_id = '0'
>>> nnabla_ext.cudnn.check_gpu(device_id)
>>> quit();
(NNabla) [root@rockylinux9 ~]#

mpiのライブラリを噛ますとNVIDIA Collective Communication Library (NCCL) も有効になる様子. これで複数のGPUを使った解析もできるそうな.

↑

Theano†

開発終了っぽい
後継品があってPyTensorだそうなhttps://github.com/pymc-devs/pytensor

[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda create -n theano theano numpy=1.16 nose cuda-nvrtc=11.3 cudnn=8.2 -c nvidia
 
[root@rockylinux9 ~]# conda activate theano
(theano) [root@rockylinux9 ~]# 
(theano) [root@rockylinux9 ~]# conda list
 :
cuda-nvrtc                11.3.58              he300756_0    nvidia
cudatoolkit               11.3.1               ha36c431_9    nvidia
cudnn                     8.2.1                cuda11.3_0
 :
nose                      1.3.7           pyhd3eb1b0_1008
numpy                     1.16.6           py39h7895c89_4
 :
pygpu                     0.7.6            py39hce1f21e_1
python                    3.9.19               h955ad1f_1
 :
theano                    1.0.5            py39h295c915_1
 :
(theano) [root@rockylinux9 ~]#

テストファイルを用意する(gpu_tutorial1.py)

from theano import function, config, shared, tensor
import numpy
import time

vlen = 10 * 30 * 768  # 10 x #cores x # threads per core
iters = 1000

rng = numpy.random.RandomState(22)
x = shared(numpy.asarray(rng.rand(vlen), config.floatX))
f = function([], tensor.exp(x))
print(f.maker.fgraph.toposort())
t0 = time.time()
for i in range(iters):
    r = f()
t1 = time.time()
print("Looping %d times took %f seconds" % (iters, t1 - t0))
print("Result is %s" % (r,))
if numpy.any([isinstance(x.op, tensor.Elemwise) and
              ('Gpu' not in type(x.op).__name__)
              for x in f.maker.fgraph.toposort()]):
    print('Used the cpu')
else:
    print('Used the gpu')

(theano) [root@rockylinux9 ~]# 
(theano) [root@rockylinux9 ~]# THEANO_FLAGS=device=cpu python gpu_tutorial1.py
[Elemwise{exp,no_inplace}(<TensorType(float64, vector)>)]
Looping 1000 times took 1.055831 seconds
Result is [1.23178032 1.61879341 1.52278065 ... 2.20771815 2.29967753 1.62323285]
Used the cpu
 
(theano) [root@rockylinux9 ~]# THEANO_FLAGS=device=cuda0 python gpu_tutorial1.py
Can not use cuDNN on context None: cannot compile with cuDNN. We got this error:
b'/tmp/try_flags_jovwr1bi.c:4:10: fatal error: cudnn.h: No such file or directory\n #include <cudnn.h>\n          ^~~~~~~~~\ncompilation terminated.\n'
Mapped name None to device cuda0: NVIDIA RTX A2000 (0000:13:00.0)
[GpuElemwise{exp,no_inplace}(<GpuArrayType<None>(float64, vector)>), HostFromGpu(gpuarray)(GpuElemwise{exp,no_inplace}.0)]
Looping 1000 times took 0.696092 seconds
Result is [1.23178032 1.61879341 1.52278065 ... 2.20771815 2.29967753 1.62323285]
Used the gpu
 
(theano) [root@rockylinux9 ~]#

↑

CNTK†

Microsoft Cognitive Toolkit 開発終了みたい
https://github.com/microsoft/CNTK

https://pypi.org/project/cntk/から python 3.6 が上限みたい

[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda create -n CNTK python=3.6
 
[root@rockylinux9 ~]# conda activate CNTK
 
(CNTK) [root@rockylinux9 ~]# pip install cntk

↑

Keras3†

機械学習ライブラリ
keras2はTensorFlowに同封される. keras3はTensorFlow, Pytorch, jaxをバックエンドとして選べる

↑

MXnet†

[root@rockylinux9 ~]# bash /Public/cuda/cuda_11.2.2_460.32.03_linux.run --override       *cuda-11.2ライブラリを入れます。このバージョンが必要
[root@rockylinux9 ~]# ldconfig
 
[root@rockylinux9 ~]# source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
 
[root@rockylinux9 ~]# conda create -n mxnet python=3.8 "numpy<1.20" cudnn=8 nccl=2 libcusolver=11 -c nvidia
 
[root@rockylinux9 ~]# conda activate mxnet
(mxnet) [root@rockylinux9 ~]# pip install mxnet-cu112
 
(mxnet) [root@rockylinux9 ~]# conda install pillow=8.0
 
(mxnet) [root@rockylinux9 ~]# conda list
 :
cuda-nvrtc                12.6.20                       0    nvidia
cuda-version              12.6                          3    nvidia
cudnn                     8.9.2.26               cuda12_0
 :
libcublas                 12.6.0.22                     0    nvidia
libcusolver               11.6.4.38                     0    nvidia
libcusparse               12.5.2.23                     0    nvidia
 :
mxnet-cu112               1.9.1                    pypi_0    pypi
nccl                      2.21.5.1             ha515578_0
 :
python                    3.8.19               h955ad1f_0
 :
(mxnet) [root@rockylinux9 ~]#

テスト　参照先https://touch-sp.hatenablog.com/entry/2020/04/24/145406様

[illya@rockylinux9 ~]$ source /apps/pyenv/versions/anaconda3-2024.06-1/etc/profile.d/conda.sh
[illya@rockylinux9 ~]$ conda activate mxnet
(mxnet) [illya@rockylinux9 ~]$ python
import mxnet as mx
from mxnet import image, gluon
 
ctx = mx.gpu() if mx.context.num_gpus() >0 else mx.cpu()
 
gluon.utils.download('https://s3.amazonaws.com/onnx-mxnet/examples/super_res_input.jpg')
 
img = image.imread('super_res_input.jpg').astype('float32')/255
img = mx.nd.transpose(img, (2,0,1))
 
gluon.utils.download('https://raw.githubusercontent.com/WolframRhodium/Super-Resolution-Zoo/master/ARAN/aran_c0_s1_x4-symbol.json')
gluon.utils.download('https://raw.githubusercontent.com/WolframRhodium/Super-Resolution-Zoo/master/ARAN/aran_c0_s1_x4-0000.params')
 
net = gluon.SymbolBlock.imports("aran_c0_s1_x4-symbol.json",['data'], "aran_c0_s1_x4-0000.params")
net.collect_params().reset_ctx(ctx)
 
output = net(img.expand_dims(0).as_in_context(ctx))
output = mx.nd.squeeze(output)
output = (mx.nd.transpose(output, (1,2,0))*255).astype('uint8')
 
from PIL import Image
img = Image.fromarray(output.asnumpy())
img.save('ARAN_4x.jpg')
 
quit();
(mxnet) [illya@rockylinux9 ~]$

犬の画像を大きく拡大して、解像度を上げる操作を行っている。通常なら画像を大きくすればぼやけるが、それを補正している。

御品書き
■samba スキーマ拡張
ユーザ登録(1,2)
win参加 RSAT lin参加

ホームディレクトリ
 共有フォルダ(1)
VSS DC追加 GPO SSO
透過的ファイル圧縮
 Server-Side Copy
ルータ超え時刻同期
 ログオンスクリプト
 NAS参加(1) NFS DFS
Mac参加(NFS)
ownCloud
リモートディスクトップ
 ディスクトップ
 smb3マルチチャンネル
 backup
メモ OpenAM
samba/LDAP XP
nisでsmb
admin-tools