def ICSTN(opt,imageFull,p):
def conv2Layer(opt,feat,outDim):
weight,bias = createVariable(opt,[7,7,int(feat.shape[-1]),outDim],stddev=opt.stdGP)
conv = tf.nn.conv2d(feat,weight,strides=[1,1,1,1],padding="VALID")+bias
return conv
def linearLayer(opt,feat,outDim):
weight,bias = createVariable(opt,[int(feat.shape[-1]),outDim],stddev=opt.stdGP)
fc = tf.matmul(feat,weight)+bias
return fc
imageWarpAll = []
for l in range(opt.warpN):
with tf.variable_scope("geometric",reuse=l>0):
pMtrx = warp.vec2mtrx(opt,p)
imageWarp = warp.transformCropImage(opt,imageFull,pMtrx)
imageWarpAll.append(imageWarp)
feat = imageWarp
with tf.variable_scope("conv1"):
feat = conv2Layer(opt,feat,6)
feat = tf.nn.relu(feat)
with tf.variable_scope("conv2"):
feat = conv2Layer(opt,feat,24)
feat = tf.nn.relu(feat)
feat = tf.reshape(feat,[opt.batchSize,-1])
with tf.variable_scope("fc3"):
feat = linearLayer(opt,feat,opt.warpDim)
dp = feat
p = warp.compose(opt,p,dp)
pMtrx = warp.vec2mtrx(opt,p)
imageWarp = warp.transformCropImage(opt,imageFull,pMtrx)
imageWarpAll.append(imageWarp)
return imageWarpAll
def perturbBG(opt, imageData):
"""
generate pereturbed image
교란된 이미지 생성하는 함수
:param opt: 옵션
:param imageData: 전체 이미지 데이터셋
:return: 이미지 데이터
"""
# batch 사이즈 만큼의 랜덤한 값을 가진 배열들을 생성하여 각 원소에 pertBG 값 곱해줌
rot = opt.pertBG * tf.random_normal([opt.batchSize])
tx = opt.pertBG * tf.random_normal([opt.batchSize])
ty = opt.pertBG * tf.random_normal([opt.batchSize])
# 배치 사이즈 만큼의 0으로만 이루어진 배열 생성
O = tf.zeros([opt.batchSize])
# 교란된 백그라운드 이미지(인물 사진) 생성을 위한 warp 매개변수 생성
# 왜곡 타입이 homography(가로, 세로간의 평행을 유지하지 않고 이미지를 왜곡시킴)일 때와
# 왜곡 타입이 affine(가로, 세로간의 평행을 유지하며 이미지를 왜곡시킴)으로 나누어 생성
pPertBG = \
tf.stack([tx, rot, O, O, ty, -rot, O, O], axis=1) if opt.warpType == "homography" else\
tf.stack([O, rot, tx, -rot, O, ty], axis=1) if opt.warpType == "affine" else None
# 벡터 형태의 warp 매개변수를 matrix로 변환
pPertBGmtrx = warp.vec2mtrx(opt, pPertBG)
# warp 파라미터에 따라 이미지 데이터셋을 변형 및 크롭하여 생성
image = warp.transformCropImage(opt, imageData, pPertBGmtrx)
return image
def perturbBG(opt, imageData):
rot = opt.pertBG * tf.random_normal([opt.batchSize])
tx = opt.pertBG * tf.random_normal([opt.batchSize])
ty = opt.pertBG * tf.random_normal([opt.batchSize])
O = tf.zeros([opt.batchSize])
pPertBG = tf.stack([tx,rot,O,O,ty,-rot,O,O],axis=1) if opt.warpType=="homography" else \
tf.stack([O,rot,tx,-rot,O,ty],axis=1) if opt.warpType=="affine" else None
pPertBGmtrx = warp.vec2mtrx(opt, pPertBG)
image = warp.transformCropImage(opt, imageData, pPertBGmtrx)
return image
print(util.toMagenta("building graph..."))
tf.reset_default_graph()
# build graph
with tf.device("/gpu:0"):
# ------ define input data ------
imageFull = tf.placeholder(tf.float32,shape=[opt.batchSize,opt.fullH,opt.fullW,3])
imageMean,imageVar = tf.nn.moments(imageFull,axes=[1,2],keep_dims=True)
imageFullNormalize = (imageFull-imageMean)/tf.sqrt(imageVar)
label = tf.placeholder(tf.int64,shape=[opt.batchSize])
PH = [imageFull,label]
# ------ generate perturbation ------
pInit = data.genPerturbations(opt)
pInitMtrx = warp.vec2mtrx(opt,pInit)
# ------ build network ------
imagePert = warp.transformCropImage(opt,imageFullNormalize,pInitMtrx)
imagePertRescale = imagePert*tf.sqrt(imageVar)+imageMean
if opt.netType=="CNN":
output = graph.fullCNN(opt,imagePert)
elif opt.netType=="STN":
imageWarpAll = graph.STN(opt,imagePert)
imageWarp = imageWarpAll[-1]
output = graph.CNN(opt,imageWarp)
imageWarpRescale = imageWarp*tf.sqrt(imageVar)+imageMean
elif opt.netType=="IC-STN":
imageWarpAll = graph.ICSTN(opt,imageFullNormalize,pInit)
imageWarp = imageWarpAll[-1]
output = graph.CNN(opt,imageWarp)
imageWarpRescale = imageWarp*tf.sqrt(imageVar)+imageMean
softmax = tf.nn.softmax(output)
labelOnehot = tf.one_hot(label,opt.labelN)