58
70

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 5 years have passed since last update.

LSTMを改良してconvLSTMにする

Last updated at Posted at 2016-10-28

##はじめに
chainerでLSTMを実装する際に、備え付けのlinks/connection/lstm.pyなどを使わずに書いた場合、convLSTMへの変更が楽だと気づいた。そこでLSTMをconvLSTMへ改良してみる。
##環境
GPU GTX1070
ubuntu 14.04
chainer 1.14.0
など
##convLSTMについて
convLSTMはX. Shiらが提案しているconvolutionとLSTMを組み合わせた手法である。
https://arxiv.org/pdf/1506.04214.pdf
通常のLSTMはlinearな状態で処理されるので、画像などは位置的な情報が死んでしまう。convLSTMは画像の状態を維持したまま入力するので位置情報が保持される。linearの場合の行列演算はconvolutionに置き換えれられる。これによりLSTMの時間情報、convolutionの位置情報が同時にいかされる。

peepholeがないヴァージョンのLSTMにおける順伝播計算は以下の図のようになるだろう。
qiita_lstm01.png
入力や各ゲート等で行列積が使われる。これに対してconvLSTMは以下のようになる。
qiita_convlstm01.png
行列積がconvlutionになっただけ。これをコード上で変更する。

##基本とするLSTMコード
LSTMを最小単位で作成する。ただし、後にconvLSTMへの変更を考えて、処理するデータは画像とする。具体的にはKitti datasetの時系列画像を入力していく。

あるtの画像を入力して、目標値をt+1の画像とする。つまり、次の瞬間の画像を予測するように学習させる。

LSTM_minimum.py
#!/usr/bin/env python
import six
import cupy
#import numpy
import chainer
from chainer import computational_graph, Chain, Variable, utils, gradient_check, Function
from chainer import cuda
import chainer.links as L
import chainer.functions as F
from chainer import optimizers
from chainer import serializers

from PIL import Image

#read Kitti datas
str_dir1 = '65x65data/'
str_png = '.png'
str_03 = '500'
str_02 = '50'
str_01 = '5'

imgArrayTrain = cupy.zeros((126, 65, 65), dtype=cupy.float32)
#imgArrayTest = cupy.zeros((26, 65, 65), dtype=cupy.float32)
imgArrayOut = cupy.zeros(4225, dtype=cupy.float32)


#load to train array
for i in range(126):
    str_sum = str_dir1
    if i < 10:
        str_sum = str_dir1 + str_03 + str(i) + str_png

    elif i < 100:
        str_sum = str_dir1 + str_02 + str(i) + str_png

    else:
        str_sum = str_dir1 + str_01 + str(i) + str_png
    
    img_read = Image.open(str_sum)
    imgArrayPart = cupy.asarray(img_read).astype(dtype=cupy.float32)
    imgArrayTrain[i] = imgArrayPart

imgArrayTrain = imgArrayTrain / 255

print 'imgArrayTrain.shape =', imgArrayTrain.shape
print 'imgArrayTrain.max()', imgArrayTrain.max()
print 'imgArrayTrain.min()', imgArrayTrain.min()

#load to Test array
for i in range(26):
    str_sum = str_dir1 + str_01 + str(i + 126) + str_png

    img_read = Image.open(str_sum)
    imgArrayPart = cupy.asarray(img_read).astype(dtype=cupy.float32)
    imgArrayTest[i] = imgArrayPart

imgArrayTest = imgArrayTest / 255

#Array reshape
imgArrayTrain2 = imgArrayTrain.reshape((len(imgArrayTrain), -1)).astype(dtype=imgArrayTrain.dtype)
imgArrayTest2 = imgArrayTest.reshape((len(imgArrayTest), -1)).astype(dtype=imgArrayTest.dtype)
print 'imgArrayTrain2.shape =', imgArrayTrain2.shape
print 'imgArrayTest2.shape =', imgArrayTest2.shape

len_1data = 25
count_j = 0
count_epoch = 0

#model class
class MyLSTM(chainer.Chain):
    def __init__(self, k):
        super(MyLSTM, self).__init__(
            Wz = L.Linear(k, k),
            Wi = L.Linear(k, k),
            Wf = L.Linear(k, k),
            Wo = L.Linear(k, k),
            Rz = L.Linear(k, k),
            Ri = L.Linear(k, k),
            Rf = L.Linear(k, k),
            Ro = L.Linear(k, k),
            W = L.Linear(k, k),
        )

    def __call__(self, s):
        accum_loss = None
        k = len(imgArrayTrain2[0])
        #v, k = self.embed.W.data.shape
        h = Variable(cupy.zeros((1,k), dtype=cupy.float32))
        c = Variable(cupy.zeros((1,k), dtype=cupy.float32))
        
        for i in range(len(s) - 1): #len(s) is expected to 26
            tx = Variable(cupy.array(s[i + 1], dtype=cupy.float32).reshape(1,-1))
            x_k = Variable(cupy.array([s[i]], dtype=cupy.float32).reshape(1,-1))
            z0 = self.Wz(x_k) + self.Rz(h)
            z1 = F.tanh(z0)
            i0 = self.Wi(x_k) + self.Ri(h)
            i1 = F.sigmoid(i0)
            f0 = self.Wf(x_k) + self.Rf(h)
            f1 = F.sigmoid(f0)
            c = z1 * i1 + f1 * c
            o0 = self.Wo(x_k) + self.Ro(h)
            o1 = F.sigmoid(o0)
            h = o1 * F.tanh(c)
            loss = F.mean_squared_error(self.W(h), tx)
            accum_loss = loss if accum_loss is None else accum_loss + loss
        return accum_loss

#optimize
model = MyLSTM(len(imgArrayTrain2[0]))
cuda.get_device(0).use() #for GPU
model.to_gpu() #for GPU
optimizer = optimizers.Adam()
optimizer.setup(model)

#learning phase
for epoch in range(10):
    
    print "epoch =", epoch
    for j in range(5):
        
        s = imgArrayTrain2[j*25:(j+1)*25 + 1,:]
        model.zerograds()
        loss = model(s)
        loss.backward()    
        optimizer.update()

    count_epoch += 1
print 'learning is finished'

LSTMの行列演算は、MyLSTMクラス内のcall関数におけるループ内。ここで上図のFigure1の演算を行っている。つまり、ここを中心に変更する。
##convLSTMへの変更
まずMyLSTMクラスの初期化をL.LinearからL.Convolution2Dへ変更する。

def __init__(self):
        super(MyLSTM, self).__init__(
            Wz = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Wi = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Wf = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Wo = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Rz = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Ri = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Rf = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Ro = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
        )

そうすると、call関数のループはそのままでよい。また、convolutionの場合、Variableへ入力するnumpy(cupy)データは(batchsize, channel, heith, width)になるので要注意。このあたりも若干変更する。最終的なコードは以下。

convLSTM.py
#!/usr/bin/env python
import six
import cupy
#import numpy
import chainer
from chainer import computational_graph, Chain, Variable, utils, gradient_check, Function
from chainer import cuda
import chainer.links as L
import chainer.functions as F
from chainer import optimizers
from chainer import serializers

from PIL import Image

#read Kitti datas
str_dir1 = '65x65data/'
str_png = '.png'
str_03 = '500'
str_02 = '50'
str_01 = '5'

epochNum = 10 # the number of epoch
len_1data = 25
channelIn = 1 #channel number of input image
channelOut = 1 #channel number of output image
width = 65 #width of input image
height = 65 #height of imput image
ksize = 5
padsize = (ksize - 1) / 2

imgArrayTrain = cupy.zeros((126, height, width), dtype=cupy.float32)
imgArrayTest = cupy.zeros((26, height, width), dtype=cupy.float32)


#load to train array
for i in range(126):
    str_sum = str_dir1
    if i < 10:
        str_sum = str_dir1 + str_03 + str(i) + str_png

    elif i < 100:
        str_sum = str_dir1 + str_02 + str(i) + str_png

    else:
        str_sum = str_dir1 + str_01 + str(i) + str_png
    
    img_read = Image.open(str_sum)
    imgArrayPart = cupy.asarray(img_read).astype(dtype=cupy.float32)
    imgArrayTrain[i] = imgArrayPart

imgArrayTrain = imgArrayTrain / 255
imgArrayTrain2 = imgArrayTrain.reshape(len(imgArrayTrain), height, width).astype(dtype=cupy.float32)


#model class
class MyLSTM(chainer.Chain):
    def __init__(self):
        super(MyLSTM, self).__init__(
            Wz = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Wi = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Wf = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Wo = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Rz = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Ri = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Rf = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
            Ro = L.Convolution2D(channelIn, channelOut, ksize, stride=1, pad=padsize),
        )

    def __call__(self, s): #s is expected to cupyArray(num, height, width)
        accum_loss = None
        chan = channelIn
        hei = len(s[0])
        wid = len(s[0][0])
        h = Variable(cupy.zeros((1, chan, hei, wid), dtype=cupy.float32))
        c = Variable(cupy.zeros((1, chan, hei, wid), dtype=cupy.float32))
        
        for i in range(len(s) - 1): #len(s) is expected to 26

            tx = Variable(cupy.array(s[i + 1], dtype=cupy.float32).reshape(1, chan, hei, wid))
            x_k = Variable(cupy.array(s[i], dtype=cupy.float32).reshape(1, chan, hei, wid))
            z0 = self.Wz(x_k) + self.Rz(h)
            z1 = F.tanh(z0)
            i0 = self.Wi(x_k) + self.Ri(h)
            i1 = F.sigmoid(i0)
            f0 = self.Wf(x_k) + self.Rf(h)
            f1 = F.sigmoid(f0)
            c = z1 * i1 + f1 * c
            o0 = self.Wo(x_k) + self.Ro(h)
            o1 = F.sigmoid(o0)
            h = o1 * F.tanh(c)
            loss = F.mean_squared_error(h, tx)
            accum_loss = loss if accum_loss is None else accum_loss + loss

        return accum_loss

#optimize
model = MyLSTM()
cuda.get_device(0).use() #for GPU
model.to_gpu() #for GPU
optimizer = optimizers.Adam()
optimizer.setup(model)

#learning phase
for epoch in range(epochNum):
    
    print "epoch =", epoch
    for j in range(5):
        
        print "now j is", j
        s = imgArrayTrain[j*25:(j+1)*25 + 1,:]
        model.zerograds()
        loss = model(s)
        loss.backward()    
        optimizer.update()

print 'learning is finished'

##コードの保存場所
このコードはこちらへあげました。
https://github.com/masataka46/convLSTM_minimum

58
70
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
58
70

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?