Chainerを使ったディープラーニング【第4回:GPUコンテナで機械学習する】

>>>model = tf.keras.models.Sequential([
tf.keras.layers.Flatten(input_shape=(28, 28)),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dropout(0.2),
tf.keras.layers.Dense(10)
])

>>>predictions = model(x_train[:1]).numpy()
2020-09-04 04:51:53.991893:
I tensorflow/stream_executor/platform/default/dso_loader.cc:44]
Successfully opened dynamic library libcublas.so.10
>>>

>>>predictions
  array([[-0.0729934 , -0.5482341 , 0.3433068 , -0.15514854, 0.44248402,
  　　　　　 0.24744856, -0.15032545, 0.24870253, 0.16561234, -0.56604826]],
  　　　　　 dtype=float32)

>>>tf.nn.softmax(predictions).numpy()

array([[0.15495376, 0.12319947, 0.12194902, 0.07117186, 0.05812015,
　　　　 0.0898917 , 0.08244707, 0.04264582, 0.13558292, 0.12003825]],
　　　　 dtype=float32)

>>>loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
>>>loss_fn(y_train[:1], predictions).numpy()
2.4091496

>>>model.compile(optimizer='adam',
loss=loss_fn,
metrics=['accuracy'])

>>>model.fit(x_train, y_train, epochs=5)
Epoch 1/5
1875/1875 [==============================] - 4s 2ms/step - loss: 0.2965 - accuracy: 0.9138
Epoch 2/5
1875/1875 [==============================] - 3s 2ms/step - loss: 0.1459 - accuracy: 0.9571
Epoch 3/5
1875/1875 [==============================] - 4s 2ms/step - loss: 0.1082 - accuracy: 0.9674
Epoch 4/5
1875/1875 [==============================] - 4s 2ms/step - loss: 0.0881 - accuracy: 0.9724
Epoch 5/5
1875/1875 [==============================] - 4s 2ms/step - loss: 0.0754 - accuracy: 0.9768
<tensorflow.python.keras.callbacks.History object at 0x7f03d0b977b8>

>>>model.evaluate(x_test, y_test, verbose=2)
313/313 - 1s - loss: 0.0737 - accuracy: 0.9774
[0.07369402796030045, 0.977400004863739]

>>>probability_model = tf.keras.Sequential([
　　model,
　　tf.keras.layers.Softmax()
])
>>>probability_model(x_test[:5])

<tf.Tensor: shape=(5, 10), dtype=float32, numpy=
　　　array([[7.41418518e-08, 7.91325783e-10, 3.75941590e-06, 2.61097477e-04,
　　　　　　　 9.48091952e-11, 3.16857694e-07, 4.86965676e-11, 9.99731362e-01,
　　　　　　　 4.52538757e-07, 2.88244155e-06],
　　　　　　　[1.33913982e-11, 2.29599886e-04, 9.99751031e-01, 1.92872030e-05,
　　　　　　　 1.66144136e-15, 1.58800315e-08, 2.13381668e-08, 9.98626540e-16,
　　　　　　　 3.95744095e-08, 4.06508526e-14],
　　　　　　　[4.50110349e-07, 9.99431074e-01, 3.84905434e-05, 1.16055037e-06,
　　　　　　　 3.32167328e-05, 5.31233536e-06, 1.03985847e-04, 3.38415528e-04,
　　　　　　　 4.75658890e-05, 1.63871476e-07],
　　　　　　　[9.99950290e-01, 7.11293871e-12, 3.84013947e-05, 2.12064151e-07,
　　　　　　　 1.02987372e-07, 4.20437999e-07, 2.38990742e-06, 2.65470180e-06,
　　　　　　　 1.09450021e-07, 5.40031533e-06],
　　　　　　　[7.92320952e-06, 2.07563033e-09, 8.12820872e-06, 2.62512536e-08,
　　　　　　　 9.88838971e-01, 1.17090956e-07, 1.72804573e-06, 2.94550318e-05,
　　　　　　　 5.51807261e-06, 1.11081181e-02]], dtype=float32)>

$ pip install chainer

$ docker images chainer/chainer
REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE
chainer/chainer     latest              3ce7a8bfbbab        3 months ago        4.24GB

$ docker run --gpus all chainer/chainer nvidia-smi
Fri Nov 20 05:30:28 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.64.00    Driver Version: 440.64.00    CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla P40           Off  | 00000000:03:00.0 Off |                    0 |
| N/A   38C    P0    50W / 250W |  21697MiB / 22919MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

$ docker run --gpus all -it chainer/chainer:latest bash
root@790e21b683b1:/# pip3 list
chainer (7.7.0)
cupy-cuda92 (7.7.0)
fastrlock (0.5)
filelock (3.0.12)
numpy (1.18.5)
pip (8.1.1)
protobuf (3.12.4)
setuptools (20.7.0)
six (1.15.0)
typing-extensions (3.7.4.2)
wheel (0.29.0)
You are using pip version 8.1.1, however version 20.2.4 is available.
You should consider upgrading via the 'pip install --upgrade pip' command.

root@790e21b683b1:/# pip3 install --upgrade pip
Collecting pip
  Downloading https://files.pythonhosted.org/packages/cb/28/91f26bd088ce8e22169032100d4260614fc3da435025ff389ef1d396a433/pip-20.2.4-py2.py3-none-any.whl (1.5MB)
    100% |################################| 1.5MB 652kB/s
Installing collected packages: pip
  Found existing installation: pip 8.1.1
    Not uninstalling pip at /usr/lib/python3/dist-packages, outside environment /usr
Successfully installed pip-20.2.4

root@790e21b683b1:/# pip3 list
WARNING: pip is being invoked by an old script wrapper. This will fail in a future version of pip.
Please see https://github.com/pypa/pip/issues/5599 for advice on fixing the underlying issue.
To avoid this problem you can invoke Python with '-m pip' instead of running pip directly.
DEPRECATION: Python 3.5 reached the end of its life on September 13th, 2020. Please upgrade your Python as Python 3.5 is no longer maintained. pip 21.0 will drop support for Python 3.5 in January 2021. pip 21.0 will remove support for this functionality.
Package           Version
----------------- -------
chainer           7.7.0
cupy-cuda92       7.7.0
fastrlock         0.5
filelock          3.0.12
numpy             1.18.5
pip               20.2.4
protobuf          3.12.4
setuptools        20.7.0
six               1.15.0
typing-extensions 3.7.4.2
wheel             0.29.0

root@790e21b683b1:/# pip3 install matplotlib
WARNING: pip is being invoked by an old script wrapper. This will fail in a future version of pip.
Please see https://github.com/pypa/pip/issues/5599 for advice on fixing the underlying issue.
To avoid this problem you can invoke Python with '-m pip' instead of running pip directly.
DEPRECATION: Python 3.5 reached the end of its life on September 13th, 2020. Please upgrade your Python as Python 3.5 is no longer maintained. pip 21.0 will drop support for Python 3.5 in January 2021. pip 21.0 will remove support for this functionality.
Collecting matplotlib
  Downloading matplotlib-3.0.3-cp35-cp35m-manylinux1_x86_64.whl (13.0 MB)
     |################################| 13.0 MB 13.3 MB/s
Collecting kiwisolver>=1.0.1
  Downloading kiwisolver-1.1.0-cp35-cp35m-manylinux1_x86_64.whl (90 kB)
     |################################| 90 kB 10.5 MB/s
（略）
Successfully installed cycler-0.10.0 kiwisolver-1.1.0 matplotlib-3.0.3 pyparsing-2.4.7 python-dateutil-2.8.1

root@790e21b683b1:/# pip3 install scikit-learn
WARNING: pip is being invoked by an old script wrapper. This will fail in a future version of pip.
Please see https://github.com/pypa/pip/issues/5599 for advice on fixing the underlying issue.
To avoid this problem you can invoke Python with '-m pip' instead of running pip directly.
DEPRECATION: Python 3.5 reached the end of its life on September 13th, 2020. Please upgrade your Python as Python 3.5 is no longer maintained. pip 21.0 will drop support for Python 3.5 in January 2021. pip 21.0 will remove support for this functionality.
Collecting scikit-learn
  Downloading scikit_learn-0.22.2.post1-cp35-cp35m-manylinux1_x86_64.whl (7.0 MB)
     |################################| 7.0 MB 8.6 MB/s
Collecting scipy>=0.17.0
  Downloading scipy-1.4.1-cp35-cp35m-manylinux1_x86_64.whl (26.0 MB)
     |################################| 26.0 MB 53.8 MB/s
（略）
Successfully installed joblib-0.14.1 scikit-learn-0.22.2.post1 scipy-1.4.

root@790e21b683b1:/# pip3 list
WARNING: pip is being invoked by an old script wrapper. This will fail in a future version of pip.
Please see https://github.com/pypa/pip/issues/5599 for advice on fixing the underlying issue.
To avoid this problem you can invoke Python with '-m pip' instead of running pip directly.
DEPRECATION: Python 3.5 reached the end of its life on September 13th, 2020. Please upgrade your Python as Python 3.5 is no longer maintained. pip 21.0 will drop support for Python 3.5 in January 2021. pip 21.0 will remove support for this functionality.
Package           Version
----------------- ------------
chainer           7.7.0
cupy-cuda92       7.7.0
cycler            0.10.0
fastrlock         0.5
filelock          3.0.12
joblib            0.14.1
kiwisolver        1.1.0
matplotlib        3.0.3
numpy             1.18.5
pip               20.2.4
protobuf          3.12.4
pyparsing         2.4.7
python-dateutil   2.8.1
scikit-learn      0.22.2.post1
scipy             1.4.1
setuptools        20.7.0
six               1.15.0
typing-extensions 3.7.4.2
wheel             0.29.0

root@790e21b683b1:/# python3
Python 3.5.2 (default, Oct  8 2019, 13:06:37)
[GCC 5.4.0 20160609] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import chainer
>>> chainer.print_runtime_info()
Platform: Linux-4.4.0-154-generic-x86_64-with-Ubuntu-16.04-xenial
Chainer: 7.7.0
ChainerX: Available
NumPy: 1.18.5
CuPy:
  CuPy Version          : 7.7.0
  CUDA Root             : /usr/local/cuda
  CUDA Build Version    : 9020
  CUDA Driver Version   : 10020
  CUDA Runtime Version  : 9020
  cuBLAS Version        : 9020
  cuFFT Version         : 9020
  cuRAND Version        : 9020
  cuSOLVER Version      : (9, 2, 0)
  cuSPARSE Version      : 9020
  NVRTC Version         : (9, 2)
  cuDNN Build Version   : 7600
  cuDNN Version         : 7605
  NCCL Build Version    : 2408
  NCCL Runtime Version  : 2408
  CUB Version           : None
  cuTENSOR Version      : None
iDeep: Not Available

>>> from sklearn.datasets import load_iris
>>> iris = load_iris()

>>> print(iris.DESCR)
.. _iris_dataset:

Iris plants dataset
--------------------

**Data Set Characteristics:**

    :Number of Instances: 150 (50 in each of three classes)
    :Number of Attributes: 4 numeric, predictive attributes and the class
    :Attribute Information:
        - sepal length in cm
        - sepal width in cm
        - petal length in cm
        - petal width in cm
        - class:
                - Iris-Setosa
                - Iris-Versicolour
                - Iris-Virginica

    :Summary Statistics:

    ============== ==== ==== ======= ===== ====================
                    Min  Max   Mean    SD   Class Correlation
    ============== ==== ==== ======= ===== ====================
    sepal length:   4.3  7.9   5.84   0.83    0.7826
    sepal width:    2.0  4.4   3.05   0.43   -0.4194
    petal length:   1.0  6.9   3.76   1.76    0.9490  (high!)
    petal width:    0.1  2.5   1.20   0.76    0.9565  (high!)
    ============== ==== ==== ======= ===== ====================

    :Missing Attribute Values: None
    :Class Distribution: 33.3% for each of 3 classes.
    :Creator: R.A. Fisher
    :Donor: Michael Marshall (MARSHALL%PLU@io.arc.nasa.gov)
    :Date: July, 1988

>>> x, t = load_iris(return_X_y=True)
>>> print('x:', x.shape)
x: (150, 4)
>>> print('t:', t.shape)
t: (150,)

>>> x = x.astype('float32')
>>> t = t.astype('int32')

>>> from sklearn.model_selection import train_test_split
>>> x_train_val, x_test, t_train_val, t_test = train_test_split(x, t, test_size=0.3, random_state=0)

>>> x_train, x_val, t_train, t_val = train_test_split(x_train_val, t_train_val, test_size=0.3, random_state=0)

>>> import chainer.links as L
>>> import chainer.functions as F

>>> l = L.Linear(3, 2)

>>> from chainer import Sequential
>>>
>>> n_input = 4
>>> n_hidden = 10
>>> n_output = 3
>>>
>>> net = Sequential(
... L.Linear(n_input, n_hidden), F.relu,
... L.Linear(n_hidden, n_hidden), F.relu,
... L.Linear(n_hidden, n_output)
... )
>>>

>>> optimizer = chainer.optimizers.SGD(lr=0.01)

>>> optimizer.setup(net)

>>> n_epoch = 30
>>> n_batchsize = 16

>>> import numpy as np
>>>

>>> iteration = 0
>>># ログの保存用
>>> results_train = {
...     'loss': [],
...     'accuracy': []
... }
>>> results_valid = {
...     'loss': [],
...     'accuracy': []
... }
>>>
>>> for epoch in range(n_epoch):
... # データセット並べ替えた順番を取得
...     order = np.random.permutation(range(len(x_train)))
... # 各バッチ毎の目的関数の出力と分類精度の保存用
...     loss_list = []
...     accuracy_list = []
... #
...     for i in range(0, len(order), n_batchsize):
...         # バッチを準備
...         index = order[i:i+n_batchsize]
...         x_train_batch = x_train[index,:]
...         t_train_batch = t_train[index]
...         # 予測値を出力
...         y_train_batch = net(x_train_batch)
...         # 目的関数を適用し、分類精度を計算
...         loss_train_batch = F.softmax_cross_entropy(y_train_batch, t_train_batch)
...         accuracy_train_batch = F.accuracy(y_train_batch, t_train_batch)
... #
...         loss_list.append(loss_train_batch.array)
...         accuracy_list.append(accuracy_train_batch.array)
...         # 勾配のリセットと勾配の計算
...         net.cleargrads()
...         loss_train_batch.backward()
...         # パラメータの更新
...         optimizer.update()
...         # カウントアップ
...         iteration += 1
...     # 訓練データに対する目的関数の出力と分類精度を集計
...     loss_train = np.mean(loss_list)
...     accuracy_train = np.mean(accuracy_list)
...     # 1エポック終えたら、検証データで評価
...     # 検証データで予測値を出力
...     with chainer.using_config('train', False), chainer.using_config('enable_backprop', False):
...         y_val = net(x_val)
...     # 目的関数を適用し、分類精度を計算
...     loss_val = F.softmax_cross_entropy(y_val, t_val)
...     accuracy_val = F.accuracy(y_val, t_val)
...     # 結果の表示
...     print('epoch: {}, iteration: {}, loss (train): {:.4f}, loss (valid): {:.4f}'.format(
...         epoch, iteration, loss_train, loss_val.array))
...     # ログを保存
...     results_train['loss'] .append(loss_train)
...     results_train['accuracy'] .append(accuracy_train)
...     results_valid['loss'].append(loss_val.array)
...     results_valid['accuracy'].append(accuracy_val.array)
...

epoch: 0, iteration: 5, loss (train): 2.0148, loss (valid): 1.3990
epoch: 1, iteration: 10, loss (train): 1.3596, loss (valid): 1.2077
epoch: 2, iteration: 15, loss (train): 1.1410, loss (valid): 1.1454
epoch: 3, iteration: 20, loss (train): 1.1005, loss (valid): 1.1080
epoch: 4, iteration: 25, loss (train): 1.0440, loss (valid): 1.0905
epoch: 5, iteration: 30, loss (train): 1.0267, loss (valid): 1.0754
epoch: 6, iteration: 35, loss (train): 1.0065, loss (valid): 1.0648
epoch: 7, iteration: 40, loss (train): 0.9888, loss (valid): 1.0561
epoch: 8, iteration: 45, loss (train): 0.9728, loss (valid): 1.0495
epoch: 9, iteration: 50, loss (train): 0.9644, loss (valid): 1.0435
epoch: 10, iteration: 55, loss (train): 0.9523, loss (valid): 1.0390
epoch: 11, iteration: 60, loss (train): 0.9513, loss (valid): 1.0311
epoch: 12, iteration: 65, loss (train): 0.9559, loss (valid): 1.0225
epoch: 13, iteration: 70, loss (train): 0.9353, loss (valid): 1.0170
epoch: 14, iteration: 75, loss (train): 0.9267, loss (valid): 1.0104
epoch: 15, iteration: 80, loss (train): 0.9171, loss (valid): 1.0026
epoch: 16, iteration: 85, loss (train): 0.9142, loss (valid): 0.9918
epoch: 17, iteration: 90, loss (train): 0.9190, loss (valid): 0.9785
epoch: 18, iteration: 95, loss (train): 0.9013, loss (valid): 0.9652
epoch: 19, iteration: 100, loss (train): 0.8918, loss (valid): 0.9508
epoch: 20, iteration: 105, loss (train): 0.8834, loss (valid): 0.9364
epoch: 21, iteration: 110, loss (train): 0.8678, loss (valid): 0.9217
epoch: 22, iteration: 115, loss (train): 0.8653, loss (valid): 0.9038
epoch: 23, iteration: 120, loss (train): 0.8406, loss (valid): 0.8899
epoch: 24, iteration: 125, loss (train): 0.8401, loss (valid): 0.8718
epoch: 25, iteration: 130, loss (train): 0.8218, loss (valid): 0.8577
epoch: 26, iteration: 135, loss (train): 0.8110, loss (valid): 0.8416
epoch: 27, iteration: 140, loss (train): 0.8112, loss (valid): 0.8259
epoch: 28, iteration: 145, loss (train): 0.7880, loss (valid): 0.8126
epoch: 29, iteration: 150, loss (train): 0.7768, loss (valid): 0.7979
>>>

>>> 'exec(%matplotlib inline)'
'exec(%matplotlib inline)'
>>> import matplotlib
>>> matplotlib.use('Agg')
>>> import matplotlib.pyplot as plt
>>>

>>>  # 目的関数の出力 (loss)
>>> plt.figure() # 何も描画されていない新しいウィンドウを作成
<Figure size 640x480 with 0 Axes>
>>> plt.plot(results_train['loss'], label='train') # label で凡例の設定
[<matplotlib.lines.Line2D object at 0x7f6b5e656dd8>]
>>> plt.plot(results_valid['loss'], label='valid') # label で凡例の設定
[<matplotlib.lines.Line2D object at 0x7f6b5e5e2198>]
>>> plt.legend() # 凡例の表示
<matplotlib.legend.Legend object at 0x7f6b5e8e9b38>
>>> plt.savefig( 'fig1_Loss.png' ) # fig1_Loss.png として保存

>>> # 分類精度の出力 (accuracy)
>>> plt.figure()
<Figure size 640x480 with 0 Axes>
>>> plt.plot(results_train['accuracy'], label='train')
[<matplotlib.lines.Line2D object at 0x7f6b5c5ccc88>]
>>> plt.plot(results_valid['accuracy'], label='valid') 
[<matplotlib.lines.Line2D object at 0x7f6b5c5d2080>]
>>> plt.legend()
<matplotlib.legend.Legend object at 0x7f6b5c5a8e10>
>>> plt.savefig( 'fig2_Accuracy.png' )

>>> with chainer.using_config('train', False), chainer.using_config('enable_backprop', False):
...     y_test = net(x_test)

>>> accuracy_test = F.accuracy(y_test, t_test)
>>> accuracy_test.array
array(0.5555556, dtype=float32)

>>> chainer.serializers.save_npz('my_iris.net', net)

root@790e21b683b1:/# ls -a
.  ..  .dockerenv  bin  boot  dev  etc  home  lib  lib64  media  mnt  my_iris.net  opt  proc  root  run  sbin  srv  sys  tmp  usr  var
root@790e21b683b1:/#

>>> loaded_net = Sequential(
...     L.Linear(n_input, n_hidden), F.relu,
...     L.Linear(n_hidden, n_hidden), F.relu,
...     L.Linear(n_hidden, n_output)
... )

>>> chainer.serializers.load_npz('my_iris.net', loaded_net)

>>> with chainer.using_config('train', False), chainer.using_config('enable_backprop', False):
...  y_test = loaded_net(x_test)
...

>>> np.argmax(y_test[0,:].array)
2