Photo to Mood


In [1]:
# グラフが文章中に表示されるようにするおまじない
%matplotlib inline

# autoreload module
%load_ext autoreload
%autoreload 2

# load local package
import sys
import os
current_path = os.getcwd()
sys.path.append(os.path.join(current_path, "../../"))  # load project root

Load the Data

image_url、moodを左端に設定したファイルから学習データを読み込みます。 なお、今回値はRekognitionのスコアであり、全項目同じ範囲の値のため正規化は行いません。

In [2]:
def read_data(path, ignore_columns):
    import os
    import numpy as np

    header = []
    y = None
    X = None
    # read data part
    with open(data_file, "rb") as f:
        header = f.readline().decode("utf-8").replace("\r", "").replace("\n", "").split("\t")
        columns = [c for c in range(len(header)) if c not in ignore_columns]
        header = [h for i, h in enumerate(header) if i not in [0] + ignore_columns]
        data = np.genfromtxt(f, invalid_raise=False, usecols=columns)        
        y = data[:, 0]
        X = data[:, 1:]
    return header, y, X

data_file = os.path.join(current_path, "../../data/photo_to_mood.txt")
header, moods, labels = read_data(data_file, [1])


['Air Sport', 'Air Vehicle', 'Animal', 'Animal Sport', 'Animal Tending', 'Archery', 'Arthropod', 'Baby', 'Ballooning', 'Bike', 'Bird', 'Black', 'Blue', 'Boat', 'Boating', 'Boxing', 'Brown', 'Burning', 'Celebration', 'Climbing', 'Club Sport', 'Combat Sport', 'Croquet', 'Cyan', 'Dog', 'Drill', 'Escalator', 'Explosion', 'Fabric', 'Face', 'Field', 'Flower', 'Flower Scene', 'Food', 'Football', 'Forest', 'Girl', 'Graduation', 'Graphics', 'Gray', 'Green', 'Group of People', 'Hands', 'Head and Shoulders', 'Hot Air Balloon', 'Human', 'Indoors', 'Insect', 'Invertebrate', 'Land Vehicle', 'Landmark', 'Long Jump', 'Mammal', 'Man Made Scene', 'Meat Eater', 'Mixed Color', 'Natural Activity', 'Nature Scene', 'Object', 'Orchard', 'Outdoors', 'Parkour', 'People Activity', 'People View', 'Person', 'Person View', 'Placental Mammal', 'Pole Vault', 'Reptile', 'Rice', 'Room', 'Sailboat', 'Sailing', 'Scene', 'Scuba', 'Ship', 'Shoes', 'Shot Put', 'Skateboarding', 'Skiing', 'Snooker', 'Snow Sport', 'Spelunking', 'Sports', 'Sports Activity', 'Surf Casting', 'Surface Water Sport', 'Surfboard', 'Team Field Sport', 'Team Sport', 'Track and Field', 'Urban Structure', 'Vehicle', 'Vertebrate', 'Village', 'Water Scene', 'Water Sport', 'Water Vehicle', 'Wedding', 'White', 'Whitehouse', 'Wild Fire', 'Winter Scene', 'Winter Sport', 'Yellow', 'Zoo']
(39, 106)

Create the Model

今回扱うのは画像の分類問題になります。そこで、分類問題でよく使われるSupport Vector Machineを利用します。

In [3]:
def select_features(feature_count, X, y, header):
    from sklearn.feature_selection import SelectKBest
    from sklearn.feature_selection import f_classif

    selector = SelectKBest(f_classif, k=feature_count).fit(X, y)
    selected = selector.get_support()
    get_headers = lambda s: [i_h[1] for i_h in enumerate(header) if s[i_h[0]]]
    kbests = sorted(zip(get_headers(selected), selector.scores_[selected]), key=lambda h_s: h_s[1], reverse=True)
    return kbests

scores = select_features(10, labels, moods, header)
selected_features = [header.index(s[0]) for s in scores]

[('People Activity', 4.3214484180914772), ('Vertebrate', 4.0271861127495239), ('Man Made Scene', 3.7372071779276408), ('Invertebrate', 3.7069930848274142), ('Insect', 3.7011581869709165), ('Placental Mammal', 3.49674539734167), ('Outdoors', 3.4732509537584231), ('Boat', 3.4000764811821109), ('Human', 3.376279867097784), ('Mammal', 3.1989521198018531)]
[62, 93, 53, 48, 47, 66, 60, 13, 45, 52]

Training the Model

パラメーターはGrid Searchで探索します。

In [6]:
def create_model(X, y, selected=()):

    from sklearn.cross_validation import train_test_split
    from sklearn.grid_search import GridSearchCV
    from sklearn.metrics import classification_report
    from sklearn import svm
    X_c = X if len(selected) == 0 else X[:, selected]

    x_train, x_test, y_train, y_test = train_test_split(X_c, y, test_size=0.2, random_state=42)
    candidates = [{'kernel': ['linear'], 'C': [1, 10, 100]}]

    clf = GridSearchCV(svm.SVC(C=1), candidates, cv=2, scoring="f1_weighted"), y_train)

    for params, mean_score, scores in sorted(clf.grid_scores_, key=lambda s: s[1], reverse=True):
        print("%0.3f (+/-%0.03f) for %r" % (mean_score, scores.std() / 2, params))

    model = clf.best_estimator_
    y_predict = model.predict(x_test)
    print(classification_report(y_test, y_predict))
    return model

model = create_model(labels, moods, selected_features)

0.584 (+/-0.016) for {'C': 1, 'kernel': 'linear'}
0.567 (+/-0.001) for {'C': 10, 'kernel': 'linear'}
0.567 (+/-0.001) for {'C': 100, 'kernel': 'linear'}
             precision    recall  f1-score   support

        0.0       0.33      1.00      0.50         2
        1.0       0.00      0.00      0.00         1
        2.0       0.50      0.50      0.50         2
        3.0       0.00      0.00      0.00         3

avg / total       0.21      0.38      0.25         8

SVC(C=1, cache_size=200, class_weight=None, coef0=0.0, degree=3, gamma=0.0,
  kernel='linear', max_iter=-1, probability=False, random_state=None,
  shrinking=True, tol=0.001, verbose=False)
D:\Develop\env\Python\Miniconda3\envs\onpasha\lib\site-packages\sklearn\metrics\ UndefinedMetricWarning: Precision and F-score are ill-defined and being set to 0.0 in labels with no predicted samples.
  'precision', 'predicted', average, warn_for)

Store the Model


In [7]:
def save(model):
    from sklearn.externals import joblib
    joblib.dump(model, "./machine.pkl") 

print([header[s] for s in sorted(selected_features)])

['Boat', 'Human', 'Insect', 'Invertebrate', 'Mammal', 'Man Made Scene', 'Outdoors', 'People Activity', 'Placental Mammal', 'Vertebrate']

In [ ]: