「名刺の項目予測」モデル実装例

import os
import pandas as pd

_data_path = os.path.join('..','data','image','sansan')
_train_path = os.path.join(_data_path,'train.csv')

df = pd.read_csv(_train_path)
df.head()

from PIL import Image

_img_path = os.path.join(_data_path,'train_images')

img = Image.open(os.path.join(_img_path, '2842.png'))

img_cropped = img.crop((491, 455, 796, 485))
img_cropped

img = Image.open(os.path.join(_img_path, '95.png'))
img_cropped = img.crop((320, 498, 865, 521))
img_cropped

print('画像ファイルの数:', len(df['file_name'].unique()))
print('領域の総数:', len(df))

print('全データ:', len(df))
print(df[['company_name','full_name','position_name', 'address', 'phone_number', 'fax', 'mobile', 'email','url']].agg(sum))

print('全データ:', len(df))
print(df[['company_name','full_name','position_name', 'address', 'phone_number', 'fax', 'mobile', 'email','url']].sum(axis=1).value_counts())

%matplotlib inline
df_img_file = df.groupby('file_name').apply(len)
df_img_file.hist()

img_cropped = img_cropped.convert('L') # convertメソッドによりグレースケール化
print('元の大きさ:', img_cropped.size)
img_resized = img_cropped.resize((216, 72)) # (216, 72)にリサイズする
print('リサイズ後の大きさ:', img_resized.size)

img_cropped

img_resized

import numpy as np
from skimage.feature import hog

img_array = np.array(img_resized)  # 画像データを数値データに変換
img_feat = hog(img_array,
               orientations = 12,
               pixels_per_cell = (12, 12),
               cells_per_block = (1, 1),
               block_norm='L1-sqrt')
print(img_feat.shape)

import time

def load_data(file_name, img_dir, img_shape, orientations, pixels_per_cell, cells_per_block, label=True):
    classes = ['company_name', 'full_name', 'position_name', 'address', 'phone_number', 'fax', 'mobile', 'email', 'url']
    df = pd.read_csv(file_name)
    n = len(df)
    if label:
        Y = np.zeros((n, len(classes)))
    print('loading...')
    s = time.clock()
    for i, row in df.iterrows():
        f, l, t, r, b = row['file_name'], row['left'], row['top'], row['right'], row['bottom']
        img = Image.open(os.path.join(img_dir, f)).crop((l,t,r,b)) # 項目領域画像を切り出す
        if img.size[0]<img.size[1]:                                # 縦長の画像に関しては90度回転して横長の画像に統一する
            img = img.transpose(Image.ROTATE_90)
        
        # preprocess
        img = img.convert('L')
        img_array = np.array(img.resize(img_shape))/255.       # img_shapeに従った大きさにそろえる
        
        
        # feature extraction
        img_feat = hog(img_array,
                       orientations = orientations,
                       pixels_per_cell = pixels_per_cell,
                       cells_per_block = cells_per_block, 
                       block_norm='L1-sqrt')
        if i == 0:
            feature_dim = len(img_feat)
            print('feature dim:', feature_dim)
            X = np.zeros((n, feature_dim))
        if (i+1)%5000==0:
            print((i+1), '/', n)
        
        X[i,:] = np.array([img_feat])
        if label:
            y = list(row[classes])
            Y[i,:] = np.array(y)
    
    print('Done. Took', time.clock()-s, 'seconds.')
    if label:
        return X, Y
    else:
        return X

img_shape = (216,72)
orientations = 12
pixels_per_cell = (12, 12)
cells_per_block = (1, 1)
X_train, Y_train = load_data(_train_path, _img_path, img_shape, orientations, pixels_per_cell, cells_per_block)
print('学習データの数:', X_train.shape[0])
print('特徴ベクトルの次元数:', X_train.shape[1])

from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.neural_network import MLPClassifier

class MultiLabelClassifier():
    def __init__(self, n_out, model_name=None):
        self.n_out = n_out
        if model_name == 'LR':
            print('LR')
        elif model_name == 'GB':
            print('GB')
        elif model_name == 'RF':
            print('RF')
        elif model_name == 'MLP':
            print('MLP')
        else:
            print('LR')
        model_list = []
        for l in range(self.n_out):
            if model_name == 'LR':
                model_list.append(LogisticRegression(C=0.1))
            elif model_name == 'GB':
                model_list.append(GradientBoostingClassifier())
            elif model_name == 'RF':
                model_list.append(RandomForestClassifier())
            elif model_name == 'MLP':
                model_list.append(MLPClassifier(alpha=0.05, early_stopping=True))
            else:
                model_list.append(LogisticRegression())
        self.models = model_list
        
    def fit(self, X, Y):
        i = 0
        start_overall = time.clock()
        for model in self.models:
            start = time.clock()
            print('training model No.%s...'%(i+1))
            model.fit(X, Y[:,i])
            print('Done. Took', time.clock()-start, 'seconds.')
            i += 1
        print('Done. Took', time.clock()-start_overall, 'seconds.')
    
    def predict(self, X):
        i = 0
        predictions = np.zeros((len(X), self.n_out))
        start = time.clock()
        print('predicting...')
        for model in self.models:
            predictions[:,i] = model.predict(X)
            print(str(i+1),'/',str(self.n_out))
            i += 1
        print('Done. Took', time.clock()-start, 'seconds.')
        
        return predictions

model_lr = MultiLabelClassifier(n_out = 9, model_name='LR')  # 今回は9項目あるため, クラス数は9個に設定
model_lr.fit(X_train, Y_train)

model_mlp = MultiLabelClassifier(n_out = 9, model_name='MLP')  # 今回は9項目あるため, クラス数は9個に設定
model_mlp.fit(X_train, Y_train)

predictions_lr_train = model_lr.predict(X_train)
predictions_mlp_train = model_mlp.predict(X_train)

def mae(y, yhat):
    return np.mean(np.abs(y - yhat))

print('Logistic Regression:', mae(Y_train, predictions_lr_train))
print('MLP:', mae(Y_train, predictions_mlp_train))

_test_path = os.path.join(_data_path, 'test.csv')
_test_img_path = os.path.join(_data_path, 'test_images')
_ans_path = os.path.join(_data_path, 'ans.csv')
_fin_path = os.path.join(_data_path, 'fin.csv')
_tmp_path = os.path.join(_data_path, 'temp.csv')

X_test = load_data(_test_path, _test_img_path, img_shape, orientations, pixels_per_cell, cells_per_block, label=False)

predictions_lr_test = model_lr.predict(X_test)
predictions_mlp_test = model_mlp.predict(X_test)

ans = pd.read_csv(_ans_path,header=None,index_col=0)
ans_array = np.array(ans)

print('Logistic Regression:', mae(ans_array, predictions_lr_test))
print('MLP:', mae(ans_array, predictions_mlp_test))

predictions_ensemble_test = ((predictions_lr_test+predictions_mlp_test)>=1).astype(np.int)

mae(ans_array, predictions_ensemble_test)