def triplet_img_gen(BASE_DIR,
IMG_ROW,
IMG_COL,
kindlist,
kindimgpathlist,
triplet_num=40):
tripletlist = []
for i in range(len(kindimgpathlist)):
print(i)
for j in range(triplet_num):
negativeindex = random.choice(
[num for num in range(len(kindimgpathlist)) if num not in [i]])
pair = np.random.randint(0, len(kindimgpathlist[i]), 2)
anchorpath = kindimgpathlist[i][pair[0]]
anchor = imgarr(BASE_DIR + anchorpath)
positivepath = kindimgpathlist[i][pair[1]]
positive = imgarr(BASE_DIR + positivepath)
rnd_id = random.randint(0, len(kindimgpathlist[negativeindex]) - 1)
negativepath = kindimgpathlist[negativeindex][rnd_id]
negative = imgarr(BASE_DIR + negativepath)
anchor = cv2.resize(anchor, (IMG_ROW, IMG_COL))
positive = cv2.resize(positive, (IMG_ROW, IMG_COL))
negative = cv2.resize(negative, (IMG_ROW, IMG_COL))
triplet = [anchor, positive, negative]
tripletlist.append(triplet)
return np.array(tripletlist)
def siamese_train_gen(BASE_DIR,
IMG_ROW,
IMG_COL,
pathdata,
labeldata,
batch_size=50):
while (True):
imglist1 = []
imglist2 = []
labellist = []
for i in range(batch_size):
rndid = random.randint(0, len(pathdata) - 1)
imgpath = BASE_DIR + pathdata[pathdata.index[rndid]]
img1 = imgarr(imgpath)
if (i % 2 == 0):
kind = 0
labellist.append(0)
else:
kind = 1
labellist.append(1)
img2 = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[rndid]],
labeldata,
pathdata,
kind=kind)
img1 = cv2.resize(img1, (IMG_ROW, IMG_COL))
img2 = cv2.resize(img2, (IMG_ROW, IMG_COL))
imglist1.append(img1)
imglist2.append(img2)
yield ([np.asarray(imglist1),
np.asarray(imglist2)], np.asarray(labellist))
def triplet_train_gen(BASE_DIR,
IMG_ROW,
IMG_COL,
pathdata,
labeldata,
batch_size=50):
while (True):
anchorlist = []
positivelist = []
negativelist = []
for i in range(batch_size):
rndid = random.randint(0, len(pathdata) - 1)
imgpath = BASE_DIR + pathdata[pathdata.index[rndid]]
anchor = imgarr(imgpath)
anchor = cv2.resize(anchor, (IMG_ROW, IMG_COL))
anchorlist.append(anchor)
positive = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[rndid]],
labeldata,
pathdata,
kind=1)
negative = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[rndid]],
labeldata,
pathdata,
kind=0)
positive = cv2.resize(positive, (IMG_ROW, IMG_COL))
negative = cv2.resize(negative, (IMG_ROW, IMG_COL))
positivelist.append(positive)
negativelist.append(negative)
yield ([
np.asarray(anchorlist),
np.asarray(positivelist),
np.asarray(negativelist)
], None)
def triplet_train_gen1(BASE_DIR,
IMG_ROW,
IMG_COL,
pathdata,
labeldata,
batch_size=50):
while (True):
imglist = []
for i in range(batch_size):
rndid = random.randint(0, len(pathdata) - 1)
imgpath = BASE_DIR + pathdata[pathdata.index[rndid]]
anchor = imgarr(imgpath)
anchor = cv2.resize(anchor, (IMG_ROW, IMG_COL))
positive = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[rndid]],
labeldata,
pathdata,
kind=1)
negative = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[rndid]],
labeldata,
pathdata,
kind=0)
positive = cv2.resize(positive, (IMG_ROW, IMG_COL))
negative = cv2.resize(negative, (IMG_ROW, IMG_COL))
imglist.append([anchor, positive, negative])
imglist = np.asarray(imglist)
yield ([imglist[:, 0], imglist[:, 1], imglist[:, 2]])
def triplet_valid_gen(BASE_DIR, IMG_ROW, IMG_COL, pathdata, labeldata):
anchorlist = []
positivelist = []
negativelist = []
for i in range(len(pathdata)):
imgpath = BASE_DIR + pathdata[pathdata.index[i]]
print(imgpath)
print(os.path.exists(imgpath))
anchor = imgarr(imgpath)
anchor = cv2.resize(anchor, (IMG_ROW, IMG_COL))
positive = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[i]],
labeldata,
pathdata,
kind=1)
negative = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[i]],
labeldata,
pathdata,
kind=0)
positive = cv2.resize(positive, (IMG_ROW, IMG_COL))
negative = cv2.resize(negative, (IMG_ROW, IMG_COL))
anchorlist.append(anchor)
positivelist.append(positive)
negativelist.append(negative)
return ([
np.asarray(anchorlist),
np.asarray(positivelist),
np.asarray(negativelist)
], None)
def img_all_list(BASE_DIR, IMG_ROW, IMG_COL, imgpathlist):
imglist = []
for path in imgpathlist:
imgpath = BASE_DIR + path
img = imgarr(imgpath)
img = cv2.resize(img, (IMG_ROW, IMG_COL))
imglist.append(img)
return np.asarray(imglist)
def extract_imgarr(BASE_DIR, pathlist, IMG_ROW, IMG_COL):
pathpair = np.random.randint(0, len(pathlist), 2)
imglist = []
for i in range(2):
path = BASE_DIR + pathlist[pathpair[i]]
img = imgarr(path)
img = cv2.resize(img, (IMG_ROW, IMG_COL))
imglist.append(img)
return imglist
def triplet_img_gen1(BASE_DIR,
IMG_ROW,
IMG_COL,
kindlist,
kindimgpathlist,
times_kind=40,
triplet_num=10,
batch_size=10):
while (True):
anchorlist = []
positivelist = []
negativelist = []
for i in range(batch_size):
print('batch_size:' + str(i))
for j in range(times_kind):
print('kind_num:' + str(j))
for k in range(triplet_num):
rndid = random.randint(0, len(kindimgpathlist) - 1)
negativeindex = random.choice([
num for num in range(len(kindimgpathlist))
if num not in [rndid]
])
pair = np.random.randint(0, len(kindimgpathlist[rndid]), 2)
anchorpath = kindimgpathlist[rndid][pair[0]]
positivepath = kindimgpathlist[rndid][pair[1]]
rndid1 = random.randint(
0,
len(kindimgpathlist[negativeindex]) - 1)
negativepath = kindimgpathlist[negativeindex][rndid1]
anchor = imgarr(BASE_DIR + anchorpath)
positive = imgarr(BASE_DIR + positivepath)
negative = imgarr(BASE_DIR + negativepath)
anchor = cv2.resize(anchor, (IMG_ROW, IMG_COL))
positive = cv2.resize(positive, (IMG_ROW, IMG_COL))
negative = cv2.resize(negative, (IMG_ROW, IMG_COL))
anchorlist.append(anchor)
positivelist.append(positive)
negativelist.append(negative)
yield ([
np.asarray(anchorlist),
np.asarray(positivelist),
np.asarray(negativelist)
], None)
def pretrain_valid_gen(BASE_DIR, IMG_ROW, IMG_COL, pathdata, labeldata):
imglist = []
labellist = []
for i in range(len(pathdata)):
imgpath = BASE_DIR + pathdata[pathdata.index[i]]
img = imgarr(imgpath)
img = cv2.resize(img, (IMG_ROW, IMG_COL))
imglist.append(img)
if (labeldata[labeldata.index[i]] == 'new_whale'):
labellist.append(0)
else:
labellist.append(1)
return (np.asarray(imglist), np.asarray(labellist))
def siamese_img_gen(BASE_DIR,
IMG_ROW,
IMG_COL,
kindimgpathlist,
contrast_times=10,
batch_size=50):
while (True):
imglist1 = []
imglist2 = []
labellist = []
for i in range(batch_size):
for j in range(contrast_times):
rndid = random.randint(0, len(kindimgpathlist) - 1)
if (i % 2 == 0):
pair = np.random.randint(0, len(kindimgpathlist[rndid]), 2)
imgpath1 = kindimgpathlist[rndid][pair[0]]
imgpath2 = kindimgpathlist[rndid][pair[1]]
labellist.append(0)
else:
rndid1 = random.randint(0, len(kindimgpathlist[rndid]) - 1)
imgpath1 = kindimgpathlist[rndid][rndid1]
index1 = random.choice([
num for num in range(len(kindimgpathlist))
if num not in [rndid]
])
rndid2 = random.randint(0,
len(kindimgpathlist[index1]) - 1)
imgpath2 = kindimgpathlist[index1][rndid2]
labellist.append(1)
img1 = imgarr(BASE_DIR + imgpath1)
img2 = imgarr(BASE_DIR + imgpath2)
img1 = cv2.resize(img1, (IMG_ROW, IMG_COL))
img2 = cv2.resize(img2, (IMG_ROW, IMG_COL))
imglist1.append(img1)
imglist2.append(img2)
yield ([np.asarray(imglist1),
np.asarray(imglist2)], np.asarray(labellist))
def pretrain_train_gen(BASE_DIR,
IMG_ROW,
IMG_COL,
pathdata,
labeldata,
batch_size=50):
while (True):
imglist = []
labellist = []
for i in range(batch_size):
rnd_id = random.randint(0, len(pathdata) - 1)
imgpath = BASE_DIR + pathdata[pathdata.index[rnd_id]]
img = imgarr(imgpath)
img = cv2.resize(img, (IMG_ROW, IMG_COL))
imglist.append(img)
if (labeldata[labeldata.index[rnd_id]] == 'new_whale'):
labellist.append(0)
else:
labellist.append(1)
yield (np.asarray(imglist), np.asarray(labellist))
def triplet_valid_gen1(BASE_DIR, IMG_ROW, IMG_COL, pathdata, labeldata):
imglist = []
for i in range(len(pathdata)):
imgpath = BASE_DIR + pathdata[pathdata.index[i]]
anchor = imgarr(imgpath)
anchor = cv2.resize(anchor, (IMG_ROW, IMG_COL))
positive = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[i]],
labeldata,
pathdata,
kind=1)
negative = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[i]],
labeldata,
pathdata,
kind=0)
positive = cv2.resize(positive, (IMG_ROW, IMG_COL))
negative = cv2.resize(negative, (IMG_ROW, IMG_COL))
imglist.append([anchor, positive, negative])
imglist = np.asarray(imglist)
return ([imglist[:, 0], imglist[:, 1], imglist[:, 2]])
def siamese_valid_gen(BASE_DIR, IMG_ROW, IMG_COL, pathdata, labeldata):
imglist1 = []
imglist2 = []
labellist = []
for i in range(len(pathdata)):
imgpath = BASE_DIR + pathdata[pathdata.index[i]]
img1 = imgarr(imgpath)
if (i % 2 == 0):
kind = 0
labellist.append(0)
else:
kind = 1
labellist.append(1)
img2 = siamese_fetch_img(BASE_DIR,
labeldata[labeldata.index[i]],
labeldata,
pathdata,
kind=kind)
img1 = cv2.resize(img1, (IMG_ROW, IMG_COL))
img2 = cv2.resize(img2, (IMG_ROW, IMG_COL))
imglist1.append(img1)
imglist2.append(img2)
return ([np.asarray(imglist1),
np.asarray(imglist2)], np.asarray(labellist))