###前提・実現したいこと
chainerをGPU環境で動かして、DCGANのプログラムを実行しようとしたときに「read image error occured
画像名.jpg」とエラーが発生しました。
エラーの解決方法を教えていただけないでしょうか
OS Windows10
CUDA 8.0
GPU GTX 950x
python 3.5 64bit
###発生している問題・エラーメッセージ
read image error occured ~.jpg
###該当のソースコード
python
1import pickle
2import numpy as np
3from PIL import Image
4import os
5import io as cStringIO
6import math
7import pylab
8
9
10import chainer
11from chainer import computational_graph
12from chainer import cuda
13from chainer import optimizers
14from chainer import serializers
15from chainer import Variable
16from chainer.utils import type_check
17from chainer import function
18
19import chainer.functions as F
20import chainer.links as L
21
22
23import numpy
24import io
25
26image_dir = 'C:\images'
27out_image_dir = 'C:\out_images'
28out_model_dir = 'C:\out_models'
29
30
31nz = 100 # # of dim for Z
32batchsize=100
33n_epoch=10000
34n_train=200000
35image_save_interval = 50000
36
37# read all images
38
39fs = os.listdir(image_dir)
40print (len(fs))
41dataset = []
42for fn in fs:
43 f = open('%s/%s'%(image_dir,fn), 'rb')
44 img_bin = f.read()
45 dataset.append(img_bin)
46 f.close()
47print (len(dataset))
48
49class ELU(function.Function):
50
51 """Exponential Linear Unit."""
52 # https://github.com/muupan/chainer-elu
53
54 def __init__(self, alpha=1.0):
55 self.alpha = numpy.float32(alpha)
56
57 def check_type_forward(self, in_types):
58 type_check.expect(in_types.size() == 1)
59 x_type, = in_types
60
61 type_check.expect(
62 x_type.dtype == numpy.float32,
63 )
64
65 def forward_cpu(self, x):
66 y = x[0].copy()
67 neg_indices = x[0] < 0
68 y[neg_indices] = self.alpha * (numpy.exp(y[neg_indices]) - 1)
69 return y,
70
71 def forward_gpu(self, x):
72 y = cuda.elementwise(
73 'T x, T alpha', 'T y',
74 'y = x >= 0 ? x : alpha * (exp(x) - 1)', 'elu_fwd')(
75 x[0], self.alpha)
76 return y,
77
78 def backward_cpu(self, x, gy):
79 gx = gy[0].copy()
80 neg_indices = x[0] < 0
81 gx[neg_indices] *= self.alpha * numpy.exp(x[0][neg_indices])
82 return gx,
83
84 def backward_gpu(self, x, gy):
85 gx = cuda.elementwise(
86 'T x, T gy, T alpha', 'T gx',
87 'gx = x >= 0 ? gy : gy * alpha * exp(x)', 'elu_bwd')(
88 x[0], gy[0], self.alpha)
89 return gx,
90
91
92def elu(x, alpha=1.0):
93 """Exponential Linear Unit function."""
94 # https://github.com/muupan/chainer-elu
95 return ELU(alpha=alpha)(x)
96
97
98
99
100class Generator(chainer.Chain):
101 def __init__(self):
102 super(Generator, self).__init__(
103 l0z = L.Linear(nz, 6*6*512, wscale=0.02*math.sqrt(nz)),
104 dc1 = L.Deconvolution2D(512, 256, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*512)),
105 dc2 = L.Deconvolution2D(256, 128, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*256)),
106 dc3 = L.Deconvolution2D(128, 64, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*128)),
107 dc4 = L.Deconvolution2D(64, 3, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*64)),
108 bn0l = L.BatchNormalization(6*6*512),
109 bn0 = L.BatchNormalization(512),
110 bn1 = L.BatchNormalization(256),
111 bn2 = L.BatchNormalization(128),
112 bn3 = L.BatchNormalization(64),
113 )
114
115 def __call__(self, z, test=False):
116 h = F.reshape(F.relu(self.bn0l(self.l0z(z), test=test)), (z.data.shape[0], 512, 6, 6))
117 h = F.relu(self.bn1(self.dc1(h), test=test))
118 h = F.relu(self.bn2(self.dc2(h), test=test))
119 h = F.relu(self.bn3(self.dc3(h), test=test))
120 x = (self.dc4(h))
121 return x
122
123
124
125class Discriminator(chainer.Chain):
126 def __init__(self):
127 super(Discriminator, self).__init__(
128 c0 = L.Convolution2D(3, 64, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*3)),
129 c1 = L.Convolution2D(64, 128, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*64)),
130 c2 = L.Convolution2D(128, 256, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*128)),
131 c3 = L.Convolution2D(256, 512, 4, stride=2, pad=1, wscale=0.02*math.sqrt(4*4*256)),
132 l4l = L.Linear(6*6*512, 2, wscale=0.02*math.sqrt(6*6*512)),
133 bn0 = L.BatchNormalization(64),
134 bn1 = L.BatchNormalization(128),
135 bn2 = L.BatchNormalization(256),
136 bn3 = L.BatchNormalization(512),
137 )
138
139 def __call__(self, x, test=False):
140 h = elu(self.c0(x)) # no bn because images from generator will katayotteru?
141 h = elu(self.bn1(self.c1(h), test=test))
142 h = elu(self.bn2(self.c2(h), test=test))
143 h = elu(self.bn3(self.c3(h), test=test))
144 l = self.l4l(h)
145 return l
146
147
148
149
150def clip_img(x):
151 return np.float32(-1 if x<-1 else (1 if x>1 else x))
152
153
154def train_dcgan_labeled(gen, dis, epoch0=0):
155 o_gen = optimizers.Adam(alpha=0.0002, beta1=0.5)
156 o_dis = optimizers.Adam(alpha=0.0002, beta1=0.5)
157 o_gen.setup(gen)
158 o_dis.setup(dis)
159 o_gen.add_hook(chainer.optimizer.WeightDecay(0.00001))
160 o_dis.add_hook(chainer.optimizer.WeightDecay(0.00001))
161
162 zvis = (xp.random.uniform(-1, 1, (100, nz), dtype=np.float32))
163
164 for epoch in range(epoch0,n_epoch):
165 perm = np.random.permutation(n_train)
166 sum_l_dis = np.float32(0)
167 sum_l_gen = np.float32(0)
168
169 for i in range(0, n_train, batchsize):
170 # discriminator
171 # 0: from dataset
172 # 1: from noise
173
174 print ("load image start ", i)
175 x2 = np.zeros((batchsize, 3, 96, 96), dtype=np.float32)
176 for j in range(batchsize):
177 try:
178 rnd = np.random.randint(len(dataset))
179 rnd2 = np.random.randint(2)
180
181 img = np.asarray(Image.open(cStringIO(dataset[rnd])).convert('RGB')).astype(np.float32).transpose(2, 0, 1)
182 if rnd2==0:
183 x2[j,:,:,:] = (img[:,:,::-1]-128.0)/128.0
184 else:
185 x2[j,:,:,:] = (img[:,:,:]-128.0)/128.0
186 except:
187 print ('read image error occured', fs[rnd])
188 #print "load image done"
189
190 # train generator
191 z = Variable(xp.random.uniform(-1, 1, (batchsize, nz), dtype=np.float32))
192 x = gen(z)
193 yl = dis(x)
194 L_gen = F.softmax_cross_entropy(yl, Variable(xp.zeros(batchsize, dtype=np.int32)))
195 L_dis = F.softmax_cross_entropy(yl, Variable(xp.ones(batchsize, dtype=np.int32)))
196
197 # train discriminator
198
199 x2 = Variable(cuda.to_gpu(x2))
200 yl2 = dis(x2)
201 L_dis += F.softmax_cross_entropy(yl2, Variable(xp.zeros(batchsize, dtype=np.int32)))
202
203 #print "forward done"
204
205 o_gen.zero_grads()
206 L_gen.backward()
207 o_gen.update()
208
209 o_dis.zero_grads()
210 L_dis.backward()
211 o_dis.update()
212
213 sum_l_gen += L_gen.data.get()
214 sum_l_dis += L_dis.data.get()
215
216 #print "backward done"
217
218 if i%image_save_interval==0:
219 pylab.rcParams['figure.figsize'] = (16.0,16.0)
220 pylab.clf()
221 vissize = 100
222 z = zvis
223 z[50:,:] = (xp.random.uniform(-1, 1, (50, nz), dtype=np.float32))
224 z = Variable(z)
225 x = gen(z, test=True)
226 x = x.data.get()
227 for i_ in range(100):
228 tmp = ((np.vectorize(clip_img)(x[i_,:,:,:])+1)/2).transpose(1,2,0)
229 pylab.subplot(10,10,i_+1)
230 pylab.imshow(tmp)
231 pylab.axis('off')
232 pylab.savefig('%s/vis_%d_%d.png'%(out_image_dir, epoch,i))
233
234 serializers.save_hdf5("%s/dcgan_model_dis_%d.h5"%(out_model_dir, epoch),dis)
235 serializers.save_hdf5("%s/dcgan_model_gen_%d.h5"%(out_model_dir, epoch),gen)
236 serializers.save_hdf5("%s/dcgan_state_dis_%d.h5"%(out_model_dir, epoch),o_dis)
237 serializers.save_hdf5("%s/dcgan_state_gen_%d.h5"%(out_model_dir, epoch),o_gen)
238 print ('epoch end', epoch, sum_l_gen/n_train, sum_l_dis/n_train)
239
240
241
242xp = cuda.cupy
243cuda.get_device(0).use()
244
245gen = Generator()
246dis = Discriminator()
247gen.to_gpu()
248dis.to_gpu()
249
250
251try:
252 os.mkdir(out_image_dir)
253 os.mkdir(out_model_dir)
254except:
255 pass
256
257train_dcgan_labeled(gen, dis)
###試したこと
画像サイズに問題があると考え、96×96を用意できているか確認しました
また、read image error のソースコードの部分を見て、ネット上で調べてみましたが、わかりませんでした。
###補足情報(言語/FW/ツール等のバージョンなど)
より詳細な情報
Q&A
ベストアンサー
バッドをするには、ログインかつ
こちらの条件を満たす必要があります。
2017/05/12 02:20
2017/05/12 03:48 編集
2017/05/12 17:00
2017/05/13 00:52
2017/05/13 03:25