model8.py

# モデル8
#
# 各データ数に応じてモデルの特化
# ref: https://elix-tech.github.io/ja/2016/06/02/kaggle-facial-keypoints-ja.html
#

import time
from datetime import datetime
from load_data import load2d
from saver import save_arch, save_history, load_arch
from utils import reshape2d_by_image_dim_ordering
from plotter import plot_hist, plot_model_arch
import pickle

import numpy as np
from collections import OrderedDict
# This module will be removed in 0.20.
from sklearn.cross_validation import train_test_split

from data_generator import FlippedImageDataGenerator
from keras.models import Sequential, model_from_json
from keras.layers import Convolution2D, MaxPooling2D, Flatten, Dense, Activation, Dropout
from keras.optimizers import SGD
from keras.callbacks import ModelCheckpoint, LearningRateScheduler, EarlyStopping

# 変数
model_name = 'model8'
nb_epoch = 5000
validation_split = 0.2
lr = 0.01
start = 0.03
stop = 0.001
learning_rates = np.linspace(start, stop, nb_epoch)
patience = 100 # EarlyStoppingでn回連続でエラーの最小値が更新されなかったらストップさせる
momentum = 0.9
nesterov = True
loss_method = 'mean_squared_error'


# 定数
SPECIALIST_SETTINGS = [
    dict(
        id='1',
        columns=(
            'left_eye_center_x', 'left_eye_center_y',
            'right_eye_center_x', 'right_eye_center_y',
            ),
        flip_indices=((0, 2), (1, 3)),
        ),

    dict(
        id='2',
        columns=(
            'nose_tip_x', 'nose_tip_y',
            ),
        flip_indices=(),
        ),

    dict(
        id='3',
        columns=(
            'mouth_left_corner_x', 'mouth_left_corner_y',
            'mouth_right_corner_x', 'mouth_right_corner_y',
            'mouth_center_top_lip_x', 'mouth_center_top_lip_y',
            ),
        flip_indices=((0, 2), (1, 3)),
        ),

    dict(
        id='4',
        columns=(
            'mouth_center_bottom_lip_x',
            'mouth_center_bottom_lip_y',
            ),
        flip_indices=(),
        ),

    dict(
        id='5',
        columns=(
            'left_eye_inner_corner_x', 'left_eye_inner_corner_y',
            'right_eye_inner_corner_x', 'right_eye_inner_corner_y',
            'left_eye_outer_corner_x', 'left_eye_outer_corner_y',
            'right_eye_outer_corner_x', 'right_eye_outer_corner_y',
            ),
        flip_indices=((0, 2), (1, 3), (4, 6), (5, 7)),
        ),

    dict(
        id='6',
        columns=(
            'left_eyebrow_inner_end_x', 'left_eyebrow_inner_end_y',
            'right_eyebrow_inner_end_x', 'right_eyebrow_inner_end_y',
            'left_eyebrow_outer_end_x', 'left_eyebrow_outer_end_y',
            'right_eyebrow_outer_end_x', 'right_eyebrow_outer_end_y',
            ),
        flip_indices=((0, 2), (1, 3), (4, 6), (5, 7)),
        ),
    ]



# モデル定義
specialists = OrderedDict()

for setting in SPECIALIST_SETTINGS:
    # 特定のカラムのデータの取得
    cols = setting['columns']
    X, y = load2d(cols=cols)
    X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=validation_split, random_state=42)

    # output layerだけカラムに合わせて修正
    model = load_arch('model/model7-arch-5000.json') # アーキテクチャのみを取り出す
    model.layers.pop() # 出力層を取り除く
    model.outputs = [model.layers[-1].output]
    model.layers[-1].outbound_nodes = []
    model.add(Dense(len(cols))) # 新しい出力層を追加

    # モデルを保存しておく
    arch_path = 'model/' + model_name + '-' + setting['id'] + '-arch-' + str(nb_epoch) + '.json'
    save_arch(model, arch_path) # モデルを保存しておく

    # トレーニングの準備
    sgd = SGD(lr=start, momentum=momentum, nesterov=nesterov)
    model.compile(loss=loss_method, optimizer=sgd)
    #plot(model, to_file="model_{}.png".format(cols[0]), show_shapes=True)

    flipgen = FlippedImageDataGenerator()
    flipgen.flip_indices = setting['flip_indices']
    early_stop = EarlyStopping(patience=patience)
    learning_rates = np.linspace(start, stop, nb_epoch)
    change_lr = LearningRateScheduler(lambda epoch: float(learning_rates[epoch]))
    weights_path = 'model/' + model_name + '-' + setting['id'] + '-weights-' + str(nb_epoch) + '.hdf5'
    checkpoint_collback = ModelCheckpoint(filepath = weights_path,
                                          monitor='val_loss',
                                          save_best_only=True,
                                          mode='auto')

    print("Training model for columns {} for {} epochs".format(cols, nb_epoch))


    # トレーニング実施
    hist   = model.fit_generator(flipgen.flow(X_train, y_train),
                                 samples_per_epoch=X_train.shape[0],
                                 nb_epoch=nb_epoch,
                                 validation_data=(X_val, y_val),
                                 callbacks=[checkpoint_collback, change_lr, early_stop])
    save_history(hist, model_name + '-' + setting['id'])