AI 事始め12：Get Started with TensorFlow : Keras (κέρας)

Keras: The Python Deep Learning library によると、Keras (κέρας) means horn in Greek. だそうです。脇道に逸れないで、本題に入ると、Keras is a high-level neural networks API, written in Python and capable of running on top of TensorFlow, CNTK, or Theano.だそうです。しかし、TensorFlow Keras guideによれば、tf.keras is TensorFlow's implementation of the Keras API specification. なので、

tf.keras can run any Keras-compatible code, but keep in mind:

　　　・The tf.keras version in the latest TensorFlow release might not be the same as the latest keras version from PyPI. Check tf.keras.version.
　　　・When saving a model's weights, tf.keras defaults to the checkpoint format. Pass save_format='h5' to use HDF5.

です。つまり、

model = keras.Sequential([　という表記より、TensorFlow を使う限りは、

model = tf.keras.models.Sequential([　という表記を使った方が安全ということですか。

次に、下記のようなコードを読み解くには、それぞれのリンクを辿るしかありません。

model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(512, activation=tf.nn.relu),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation=tf.nn.softmax)

])

model.compile(optimizer='adam',

                        loss='sparse_categorical_crossentropy',
                        metrics=['accuracy'])
model.fit(train_images, train_labels, epochs=5)
model.evaluate(test_images, test_labels)

理解できるかどうかは、努力次第とは思いますが、どこまで理解しないといけないかが不明なので、斜め読みして、先に進みます。

Get Started with TensorFlow では、ここまでに二つのモデルが例示されていて、違いは

    tf.keras.layers.Dense(512, activation=tf.nn.relu),
    tf.keras.layers.Dropout(0.2),

の部分のみです。neural networkのoutput arrays of shape が(*, 128)のモデルと(*, 512)のモデルが例示されています。また、Dropoutのあるモデルと無いモデルがあります。Dropoutとは、

Dropout consists in randomly setting a fraction rate of input units to 0 at each update during training time, which helps prevent overfitting.

と書かれていますが、知識ゼロでは、overfittingを抑制するために、input units？を操作するくらいしか理解できませんので、飛ばします。

計算時間の速いモデルを、tf.kerasで書き換えて実行してみました。ReStartしてRunさせる毎に結果が微妙に変わっているのは、多分、学習に使うデータをランダムに選んでいるためかな。とにかく、仕組みがわかりません。

In [1]:

# TensorFlow and tf.keras
import tensorflow as tf
from tensorflow import keras  # 最新のtensorflowにはkerasが同梱されているのかな？

# Helper libraries
import numpy as np
import matplotlib.pyplot as plt

print(tf.__version__)

 

1.11.0

In [2]:

fashion_mnist = keras.datasets.fashion_mnist

(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()

In [3]:

class_names = ['T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 
               'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle boot']

In [4]:

train_images.shape

Out[4]:

(60000, 28, 28)

In [5]:

len(train_labels)

Out[5]:

60000

In [6]:

train_labels

Out[6]:

array([9, 0, 0, ..., 3, 0, 5], dtype=uint8)

In [7]:

test_images.shape

Out[7]:

(10000, 28, 28)

In [8]:

len(test_labels)

Out[8]:

10000

In [9]:

plt.figure()
plt.imshow(train_images[0])
plt.colorbar()
plt.grid(False)

 

In [10]:

train_images = train_images / 255.0

test_images = test_images / 255.0

In [11]:

plt.figure(figsize=(10,10))
for i in range(25):
    plt.subplot(5,5,i+1)
    plt.xticks([])
    plt.yticks([])
    plt.grid(False)
    plt.imshow(train_images[i], cmap=plt.cm.binary)
    plt.xlabel(class_names[train_labels[i]])

 

In [12]:

# tf.keras.で書き換え
model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(128, activation=tf.nn.relu),
    tf.keras.layers.Dense(10, activation=tf.nn.softmax)
])

In [13]:

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

In [14]:

model.fit(train_images, train_labels, epochs=5)
test_loss, test_acc = model.evaluate(test_images, test_labels)

print('Test accuracy:', test_acc)

 

Epoch 1/5
60000/60000 [==============================] - 4s 58us/step - loss: 0.5057 - acc: 0.8219
Epoch 2/5
60000/60000 [==============================] - 3s 52us/step - loss: 0.3768 - acc: 0.8636
Epoch 3/5
60000/60000 [==============================] - 3s 51us/step - loss: 0.3408 - acc: 0.8760
Epoch 4/5
60000/60000 [==============================] - 3s 52us/step - loss: 0.3149 - acc: 0.8849
Epoch 5/5
60000/60000 [==============================] - 3s 50us/step - loss: 0.2961 - acc: 0.8904
10000/10000 [==============================] - 0s 24us/step
Test accuracy: 0.869

In [15]:

predictions = model.predict(test_images)

In [16]:

predictions[0]

Out[16]:

array([5.0372423e-06, 8.1114271e-09, 8.7406681e-07, 2.6778051e-09,
       4.5041918e-07, 3.0179184e-03, 1.2975282e-06, 2.3487598e-02,
       7.7159930e-06, 9.7347915e-01], dtype=float32)

In [17]:

np.argmax(predictions[0])

Out[17]:

9

In [18]:

test_labels[0]

Out[18]:

9

In [19]:

def plot_image(i, predictions_array, true_label, img):
  predictions_array, true_label, img = predictions_array[i], true_label[i], img[i]
  plt.grid(False)
  plt.xticks([])
  plt.yticks([])
  
  plt.imshow(img, cmap=plt.cm.binary)

  predicted_label = np.argmax(predictions_array)
  if predicted_label == true_label:
    color = 'blue'
  else:
    color = 'red'
  
  plt.xlabel("{} {:2.0f}% ({})".format(class_names[predicted_label],
                                100*np.max(predictions_array),
                                class_names[true_label]),
                                color=color)

def plot_value_array(i, predictions_array, true_label):
  predictions_array, true_label = predictions_array[i], true_label[i]
  plt.grid(False)
  plt.xticks([])
  plt.yticks([])
  thisplot = plt.bar(range(10), predictions_array, color="#777777")
  plt.ylim([0, 1]) 
  predicted_label = np.argmax(predictions_array)
 
  thisplot[predicted_label].set_color('red')
  thisplot[true_label].set_color('blue')

In [20]:

i = 0
plt.figure(figsize=(6,3))
plt.subplot(1,2,1)
plot_image(i, predictions, test_labels, test_images)
plt.subplot(1,2,2)
plot_value_array(i, predictions,  test_labels)

 

In [21]:

i = 12
plt.figure(figsize=(6,3))
plt.subplot(1,2,1)
plot_image(i, predictions, test_labels, test_images)
plt.subplot(1,2,2)
plot_value_array(i, predictions,  test_labels)

 

In [22]:

# Plot the first X test images, their predicted label, and the true label
# Color correct predictions in blue, incorrect predictions in red
num_rows = 5
num_cols = 3
num_images = num_rows*num_cols
plt.figure(figsize=(2*2*num_cols, 2*num_rows))
for i in range(num_images):
  plt.subplot(num_rows, 2*num_cols, 2*i+1)
  plot_image(i, predictions, test_labels, test_images)
  plt.subplot(num_rows, 2*num_cols, 2*i+2)
  plot_value_array(i, predictions, test_labels)

 

In [23]:

# Grab an image from the test dataset
img = test_images[0]

print(img.shape)

 

(28, 28)

In [24]:

# Add the image to a batch where it's the only member.
img = (np.expand_dims(img,0))

print(img.shape)

 

(1, 28, 28)

In [25]:

predictions_single = model.predict(img)

print(predictions_single)

 

5.0372469e-06 8.1114271e-09 8.7406596e-07 2.6777998e-09 4.5041918e-07
  3.0179170e-03 1.2975271e-06 2.3487566e-02 7.7159930e-06 9.7347915e-01

In [26]:

plot_value_array(0, predictions_single, test_labels)
_ = plt.xticks(range(10), class_names, rotation=45)

 

In [27]:

np.argmax(predictions_single[0])

Out[27]:

9

In [28]:

#@title MIT License
#
# Copyright (c) 2017 François Chollet
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.