keras

tensorflowのラッパーであるkerasを用いてセマンティックセグメンテーションをおこなう。

学習環境

項目	説明
OS	ubuntu 16.04 LTS
GPU	ELSA GeForce GTX 1070 8GB S.A.C

使用データ・前処理

データ種別	サイズ	枚数
訓練データ	384*480	10
訓練ラベル(3クラス)	384*480	10

アーキテクチャ

	サイズ	Center align
入力層	384×480	10枚
エンコーダ
Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation MaxPooling2D	6433 - relu 6433 - relu 2*2
Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation MaxPooling2D	12833 - relu 12833 - relu 2*2
Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation MaxPooling2D	25633 - relu 25633 - relu 25633 - relu 2*2
Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation MaxPooling2D	51233 - relu 51233 - relu 51233 - relu 2*2
Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation MaxPooling2D	51233 - relu 51233 - relu 51233 - relu 2*2
デコーダ
UpSampling2D Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation	22 51233 - relu 51233 - relu 5123*3 - relu
UpSampling2D Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation	22 51233 - relu 51233 - relu 2563*3 - relu
UpSampling2D Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation	22 25633 - relu 25633 - relu 1283*3 - relu
UpSampling2D Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation	22 12833 - relu 643*3 - relu
UpSampling2D Convolution2D BatchNormalization Activation Convolution2D BatchNormalization Activation	22 6433 - relu 31*1 - relu
識別層	softmax
出力層

学習規則


損失関数	クロスエントロピー
最適化手法	adadelta

実装

train.py

from __future__ import absolute_import
from __future__ import print_function
import os
import pylab as pl
import matplotlib.cm as cm
import itertools
import numpy as np
import theano.tensor as T
np.random.seed(1337) # for reproducibility

from keras.datasets import mnist
from keras.layers.noise import GaussianNoise
import keras.models as models
from keras.layers.core import Layer, Dense, Dropout, Activation, Flatten, Reshape, Merge, Permute
# from keras.layers import Merge
from keras.layers.convolutional import Convolution2D, MaxPooling2D, UpSampling2D, ZeroPadding2D
from keras.layers.normalization import BatchNormalization
from keras.utils import np_utils
from keras.regularizers import ActivityRegularizer
from keras.utils.visualize_util import plot
from keras import backend as K
K.set_image_dim_ordering('th')

import cv2
import numpy as np
import matplotlib.pyplot as plt

path = '/path/to/data_dir/'
img_h = 384
img_w = 480
data_shape = 384*480
pool_size = 2
kernel = 3
n_labels = 3

def normalized(rgb):
    #return rgb/255.0
    norm=np.zeros((rgb.shape[0], rgb.shape[1], 3),np.float32)

    b=rgb[:,:,0]
    g=rgb[:,:,1]
    r=rgb[:,:,2]

    norm[:,:,0]=cv2.equalizeHist(b)
    norm[:,:,1]=cv2.equalizeHist(g)
    norm[:,:,2]=cv2.equalizeHist(r)

    return norm

def binarylab(labels):
    x = np.zeros([384,480,3])
    for i in range(384):
        for j in range(480):
            x[i,j,labels[i][j]]=1
    return x

def prep_data():
    train_data = []
    train_label = []
    import os
    with open(path+'train_10.txt') as f:
        txt = f.readlines()
        txt = [line.split(' ') for line in txt]
    for i in range(len(txt)):
        train_data.append(np.rollaxis(normalized(cv2.imread(txt[i][0])),2))
        train_label.append(binarylab(cv2.imread(txt[i][1][:-1])[:,:,0]))
        print('.',end='')
    return np.array(train_data), np.array(train_label)

train_data, train_label = prep_data()
train_label = np.reshape(train_label,(10,data_shape,n_labels))

class_weighting= [0.1826, 0.1417,  6.2478]

class UnPooling2D(Layer):
    """A 2D Repeat layer"""
    def __init__(self, poolsize=(2, 2)):
        super(UnPooling2D, self).__init__()
        self.poolsize = poolsize

    @property
    def output_shape(self):
        input_shape = self.input_shape
        return (input_shape[0], input_shape[1],
                self.poolsize[0] * input_shape[2],
                self.poolsize[1] * input_shape[3])

    def get_output(self, train):
        X = self.get_input(train)
        s1 = self.poolsize[0]
        s2 = self.poolsize[1]
        output = X.repeat(s1, axis=2).repeat(s2, axis=3)
        return output

    def get_config(self):
        return {"name":self.__class__.__name__,
            "poolsize":self.poolsize}

encoding_layers = [
    Convolution2D(64, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(64, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    MaxPooling2D(pool_size=(pool_size, pool_size)),

    Convolution2D(128, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(128, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    MaxPooling2D(pool_size=(pool_size, pool_size)),

    Convolution2D(256, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(256, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(256, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    MaxPooling2D(pool_size=(pool_size, pool_size)),

    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    MaxPooling2D(pool_size=(pool_size, pool_size)),

    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    MaxPooling2D(pool_size=(pool_size, pool_size)),
]

decoding_layers = [
    UpSampling2D(),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),

    UpSampling2D(),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(512, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(256, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),

    UpSampling2D(),
    Convolution2D(256, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(256, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(128, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),

    UpSampling2D(),
    Convolution2D(128, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(64, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),

    UpSampling2D(),
    Convolution2D(64, kernel, kernel, border_mode='same'),
    BatchNormalization(),
    Activation('relu'),
    Convolution2D(n_labels, 1, 1, border_mode='valid'),
    BatchNormalization(),
]
autoencoder = models.Sequential()
autoencoder.add(Layer(input_shape=(n_labels, img_h, img_w)))
autoencoder.encoding_layers = encoding_layers
autoencoder.decoding_layers = decoding_layers
for l in autoencoder.encoding_layers:
    autoencoder.add(l)
for l in autoencoder.decoding_layers:
    autoencoder.add(l)
autoencoder.add(Reshape((n_labels,data_shape), input_shape=(n_labels,img_h,img_w)))
autoencoder.add(Permute((2, 1)))
autoencoder.add(Activation('softmax'))
autoencoder.compile(loss="categorical_crossentropy", optimizer='adadelta')

plot(autoencoder, to_file="autoencoder.png", show_shapes=True)

nb_epoch = 1
batch_size = 2

history = autoencoder.fit(train_data, train_label, batch_size=batch_size, nb_epoch=nb_epoch, show_accuracy=True, verbose=1, class_weight=class_weighting)

autoencoder.save_weights('/path/to/weight_file')
plt.plot(history.history['loss'], linewidth=3, label='train')

結果