tldr

KggleのAGE, GENDER AND ETHNICITY (FACE DATA) CSVをFace Image Classification - Data Every Day #042に沿ってやっていきます。

実行環境はGoogle Colaboratorです。

インポート

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

import sklearn.preprocessing as sp
from sklearn.model_selection import train_test_split
import sklearn.linear_model as slm

import tensorflow as tf

データのダウンロード

Google Driveをマウントします。

from google.colab import drive
drive.mount('/content/drive')

Mounted at /content/drive

KaggleのAPIクライアントを初期化し、認証します。
認証情報はGoogle Drive内（/content/drive/My Drive/Colab Notebooks/Kaggle）にkaggle.jsonとして置いてあります。

import os
kaggle_path = "/content/drive/My Drive/Colab Notebooks/Kaggle"
os.environ['KAGGLE_CONFIG_DIR'] = kaggle_path

from kaggle.api.kaggle_api_extended import KaggleApi
api = KaggleApi()
api.authenticate()

Kaggle APIを使ってデータをダウンロードします。

dataset_id = 'nipunarora8/age-gender-and-ethnicity-face-data-csv'
dataset = api.dataset_list_files(dataset_id)
file_name = dataset.files[0].name
file_path = os.path.join(api.get_default_download_dir(), file_name)
file_path

'/content/age_gender.csv'

api.dataset_download_file(dataset_id, file_name, force=True, quiet=False)

  8%|▊         | 5.00M/63.2M [00:00<00:01, 30.9MB/s]

Downloading age_gender.csv.zip to /content


100%|██████████| 63.2M/63.2M [00:00<00:00, 83.8MB/s]










True

データの読み込み

Pedumagalhaes/quality-prediction-in-a-mining-processadasを使ってダウンロードしてきたCSVファイルを読み込みます。

data = pd.read_csv(file_path + '.zip')

data

	age	ethnicity	gender	img_name	pixels
0	1	2	0	20161219203650636.jpg.chip.jpg	129 128 128 126 127 130 133 135 139 142 145 14...
1	1	2	0	20161219222752047.jpg.chip.jpg	164 74 111 168 169 171 175 182 184 188 193 199...
2	1	2	0	20161219222832191.jpg.chip.jpg	67 70 71 70 69 67 70 79 90 103 116 132 145 155...
3	1	2	0	20161220144911423.jpg.chip.jpg	193 197 198 200 199 200 202 203 204 205 208 21...
4	1	2	0	20161220144914327.jpg.chip.jpg	202 205 209 210 209 209 210 211 212 214 218 21...
...	...	...	...	...	...
23700	99	0	1	20170120221920654.jpg.chip.jpg	127 100 94 81 77 77 74 99 102 98 128 145 160 1...
23701	99	1	1	20170120134639935.jpg.chip.jpg	23 28 32 35 42 47 68 85 98 103 113 117 130 129...
23702	99	2	1	20170110182418864.jpg.chip.jpg	59 50 37 40 34 19 30 101 156 170 177 184 187 1...
23703	99	2	1	20170117195405372.jpg.chip.jpg	45 108 120 156 206 197 140 180 191 199 204 207...
23704	99	0	1	20170110182052119.jpg.chip.jpg	156 161 160 165 170 173 166 177 183 191 187 18...

23705 rows × 5 columns

下準備

data.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 23705 entries, 0 to 23704
Data columns (total 5 columns):
 #   Column     Non-Null Count  Dtype 
---  ------     --------------  ----- 
 0   age        23705 non-null  int64 
 1   ethnicity  23705 non-null  int64 
 2   gender     23705 non-null  int64 
 3   img_name   23705 non-null  object
 4   pixels     23705 non-null  object
dtypes: int64(3), object(2)
memory usage: 926.1+ KB

data = data.drop('img_name', axis=1)

{column: list(data[column].unique()) for column in ['gender', 'ethnicity', 'age']}

{'age': [1,
  10,
  100,
  101,
  103,
  105,
  11,
  110,
  111,
  115,
  116,
  12,
  13,
  14,
  15,
  16,
  17,
  18,
  19,
  2,
  20,
  21,
  22,
  23,
  24,
  25,
  26,
  27,
  28,
  29,
  3,
  30,
  31,
  32,
  33,
  34,
  35,
  36,
  37,
  38,
  39,
  4,
  40,
  41,
  42,
  43,
  44,
  45,
  46,
  47,
  48,
  49,
  5,
  50,
  51,
  52,
  53,
  54,
  55,
  56,
  57,
  58,
  59,
  6,
  60,
  61,
  62,
  63,
  64,
  65,
  66,
  67,
  68,
  69,
  7,
  70,
  71,
  72,
  73,
  74,
  75,
  76,
  77,
  78,
  79,
  8,
  80,
  81,
  82,
  83,
  84,
  85,
  86,
  87,
  88,
  89,
  9,
  90,
  91,
  92,
  93,
  95,
  96,
  99],
 'ethnicity': [2, 3, 4, 0, 1],
 'gender': [0, 1]}

data['age'] = pd.qcut(data['age'], q=4, labels=[0, 1, 2, 3])

data

	age	ethnicity	gender	pixels
0	0	2	0	129 128 128 126 127 130 133 135 139 142 145 14...
1	0	2	0	164 74 111 168 169 171 175 182 184 188 193 199...
2	0	2	0	67 70 71 70 69 67 70 79 90 103 116 132 145 155...
3	0	2	0	193 197 198 200 199 200 202 203 204 205 208 21...
4	0	2	0	202 205 209 210 209 209 210 211 212 214 218 21...
...	...	...	...	...
23700	3	0	1	127 100 94 81 77 77 74 99 102 98 128 145 160 1...
23701	3	1	1	23 28 32 35 42 47 68 85 98 103 113 117 130 129...
23702	3	2	1	59 50 37 40 34 19 30 101 156 170 177 184 187 1...
23703	3	2	1	45 108 120 156 206 197 140 180 191 199 204 207...
23704	3	0	1	156 161 160 165 170 173 166 177 183 191 187 18...

23705 rows × 4 columns

画像のサイズを計算します。

length = len(data['pixels'][0].split(' '))
np.sqrt(length)

48.0

num_pixels = 2304
img_height = 48
img_width = 48

target_columns = ['age', 'ethnicity', 'gender']

y = data[target_columns]
X = data.drop(target_columns, axis=1)

intのリストに変換

X = pd.Series(X['pixels'])

X = X.apply(lambda x: x.split(' '))

X = X.apply(lambda x: np.array(list(map(lambda z: np.int(z), x))))

画像データなので書く値を2Dにする

X = np.stack(np.array(X), axis=0)

X.shape

(23705, 2304)

X = np.reshape(X, (-1, 48, 48))

X.shape

(23705, 48, 48)

可視化

plt.figure(figsize=(10, 10))

for index, image in enumerate(np.random.randint(0, 1000, 9)):
    plt.subplot(3, 3, index + 1)
    plt.imshow(X[image])
    plt.axis('off')

plt.show()

トレーニング

y_gender = np.array(y['gender'])
y_age = np.array(y['age'])
y_ethnicity = np.array(y['ethnicity'])

def build_model(num_classes, activation='softmax', loss='sparse_categorical_crossentropy'):
    model = tf.keras.Sequential([
        tf.keras.layers.experimental.preprocessing.Rescaling(1./255, input_shape=(img_height, img_width, 1)),
        tf.keras.layers.Conv2D(16, 3, padding='same', activation='relu'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(32, 3, padding='same', activation='relu'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Conv2D(64, 3, padding='same', activation='relu'),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(128, activation='relu'),
        tf.keras.layers.Dense(num_classes, activation=activation),
    ])
    model.summary()

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

X_gender_train, X_gender_test, y_gender_train, y_gender_test = train_test_split(X, y_gender, train_size=0.7)
X_age_train, X_age_test, y_age_train, y_age_test = train_test_split(X, y_age, train_size=0.7)
X_ethnicity_train, X_ethnicity_test, y_ethnicity_train, y_ethnicity_test = train_test_split(X, y_ethnicity, train_size=0.7)

batch_size = 64
epochs = 20
gender_model = build_model(1, activation='sigmoid', loss='binary_crossentropy')

gender_history = gender_model.fit(
    X_gender_train,
    y_gender_train,
    validation_split=0.2,
    batch_size=batch_size,
    epochs=epochs,
    callbacks=[tf.keras.callbacks.ReduceLROnPlateau()],
    verbose=0
)

Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
rescaling_2 (Rescaling)      (None, 48, 48, 1)         0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 48, 48, 16)        160       
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 24, 24, 16)        0         
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 24, 24, 32)        4640      
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 12, 12, 32)        0         
_________________________________________________________________
conv2d_8 (Conv2D)            (None, 12, 12, 64)        18496     
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 6, 6, 64)          0         
_________________________________________________________________
flatten_2 (Flatten)          (None, 2304)              0         
_________________________________________________________________
dense_4 (Dense)              (None, 128)               295040    
_________________________________________________________________
dense_5 (Dense)              (None, 1)                 129       
=================================================================
Total params: 318,465
Trainable params: 318,465
Non-trainable params: 0
_________________________________________________________________

def plot_show(history):
    plt.figure(figsize=(14, 10))
    
    epochs_range = range(1, epochs+1)
    train_loss = history.history['loss']
    val_loss = history.history['val_loss']
    
    plt.plot(epochs_range, train_loss, label='Training Loss')
    plt.plot(epochs_range, val_loss, label='Validation Loss')
    
    plt.title('Training and Validation Loss')
    plt.xlabel('Epoch')
    plt.ylabel('Loss')
    plt.legend()
    
    plt.show()

plot_show(gender_history)

gender_acc = gender_model.evaluate(X_gender_test, y_gender_test)[1]

223/223 [==============================] - 1s 3ms/step - loss: 0.1873 - accuracy: 0.9515

age_model = build_model(4, activation='softmax', loss='sparse_categorical_crossentropy')

age_history = age_model.fit(
    X_age_train,
    y_age_train,
    validation_split=0.2,
    batch_size=batch_size,
    epochs=epochs,
    callbacks=[tf.keras.callbacks.ReduceLROnPlateau()],
    verbose=0
)
plot_show(gender_history)

Model: "sequential_11"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
rescaling_11 (Rescaling)     (None, 48, 48, 1)         0         
_________________________________________________________________
conv2d_33 (Conv2D)           (None, 48, 48, 16)        160       
_________________________________________________________________
max_pooling2d_33 (MaxPooling (None, 24, 24, 16)        0         
_________________________________________________________________
conv2d_34 (Conv2D)           (None, 24, 24, 32)        4640      
_________________________________________________________________
max_pooling2d_34 (MaxPooling (None, 12, 12, 32)        0         
_________________________________________________________________
conv2d_35 (Conv2D)           (None, 12, 12, 64)        18496     
_________________________________________________________________
max_pooling2d_35 (MaxPooling (None, 6, 6, 64)          0         
_________________________________________________________________
flatten_11 (Flatten)         (None, 2304)              0         
_________________________________________________________________
dense_22 (Dense)             (None, 128)               295040    
_________________________________________________________________
dense_23 (Dense)             (None, 4)                 516       
=================================================================
Total params: 318,852
Trainable params: 318,852
Non-trainable params: 0
_________________________________________________________________

age_acc = age_model.evaluate(X_age_test, y_age_test)[1]

223/223 [==============================] - 1s 2ms/step - loss: 1.0786 - accuracy: 0.6209

ethnicity_model = build_model(5, activation='softmax', loss='sparse_categorical_crossentropy')

ethnicity_history = ethnicity_model.fit(
    X_ethnicity_train,
    y_ethnicity_train,
    validation_split=0.2,
    batch_size=batch_size,
    epochs=epochs,
    callbacks=[tf.keras.callbacks.ReduceLROnPlateau()],
    verbose=0
)
plot_show(ethnicity_history)

Model: "sequential_12"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
rescaling_12 (Rescaling)     (None, 48, 48, 1)         0         
_________________________________________________________________
conv2d_36 (Conv2D)           (None, 48, 48, 16)        160       
_________________________________________________________________
max_pooling2d_36 (MaxPooling (None, 24, 24, 16)        0         
_________________________________________________________________
conv2d_37 (Conv2D)           (None, 24, 24, 32)        4640      
_________________________________________________________________
max_pooling2d_37 (MaxPooling (None, 12, 12, 32)        0         
_________________________________________________________________
conv2d_38 (Conv2D)           (None, 12, 12, 64)        18496     
_________________________________________________________________
max_pooling2d_38 (MaxPooling (None, 6, 6, 64)          0         
_________________________________________________________________
flatten_12 (Flatten)         (None, 2304)              0         
_________________________________________________________________
dense_24 (Dense)             (None, 128)               295040    
_________________________________________________________________
dense_25 (Dense)             (None, 5)                 645       
=================================================================
Total params: 318,981
Trainable params: 318,981
Non-trainable params: 0
_________________________________________________________________

ethnicity_acc = ethnicity_model.evaluate(X_ethnicity_test, y_ethnicity_test)[1]

223/223 [==============================] - 1s 3ms/step - loss: 0.8495 - accuracy: 0.7697