def _create_pair(path, augmentation_config_encoded, data_type, model_type):
"""
Create a contrastive learning pair for nt-xent loss.
"""
augmentation_config = json.loads(augmentation_config_encoded.numpy())
path = path.numpy()
data_type = data_type.numpy().decode()
model_type = model_type.numpy().decode()
y = read_audio(path, data_type)
y1 = randomAudioAugmentation(y, augmentation_config)
y1 = fp_to_pcm(enhance(y1))
y2 = randomAudioAugmentation(y, augmentation_config)
y2 = fp_to_pcm(enhance(y2))
if model_type == "vggvox":
y1_stft = np.expand_dims(create_spectrogram(y1)[:, :300], axis=-1)
y2_stft = np.expand_dims(create_spectrogram(y2)[:, :300], axis=-1)
elif model_type == "lstm":
y1_stft = create_spectrogram(y1)[:, :300].T
y2_stft = create_spectrogram(y2)[:, :300].T
elif model_type == "transformer":
y1_stft = np.expand_dims(create_spectrogram(y1)[:, :300].T, axis=-1)
y2_stft = np.expand_dims(create_spectrogram(y2)[:, :300].T, axis=-1)
return np.array([y1_stft, y2_stft]).astype(np.float32)
def _create_triplet(filenames, data_type, model_type):
"""
Create a triplet for the triplet loss.
"""
filenames = filenames.numpy()
data_type = data_type.numpy().decode()
model_type = model_type.numpy().decode()
filename_p = filenames[0]
filename_n = filenames[1]
y = read_audio(filename_p, data_type)
a = crop(y, 4 * 16_000)
a = fp_to_pcm(enhance(a))
p = crop(y, 4 * 16_000)
p = fp_to_pcm(enhance(p))
y = read_audio(filename_n, data_type)
n = crop(y, 4 * 16_000)
n = fp_to_pcm(enhance(n))
if model_type == "vggvox":
a_stft = np.expand_dims(create_spectrogram(a)[:, :300], axis=-1)
p_stft = np.expand_dims(create_spectrogram(p)[:, :300], axis=-1)
n_stft = np.expand_dims(create_spectrogram(n)[:, :300], axis=-1)
elif model_type == "lstm":
a_stft = create_spectrogram(a)[:, :300].T
p_stft = create_spectrogram(p)[:, :300].T
n_stft = create_spectrogram(n)[:, :300].T
elif model_type == "transformer":
a_stft = np.expand_dims(create_spectrogram(a)[:, :300].T, axis=-1)
p_stft = np.expand_dims(create_spectrogram(p)[:, :300].T, axis=-1)
n_stft = np.expand_dims(create_spectrogram(n)[:, :300].T, axis=-1)
return np.array([a_stft, p_stft, n_stft]).astype(np.float32)
def _create_example(filename, num_labels, split, data_type, model_type,
dataset):
"""
Create a single example and label for classification using CCE loss.
"""
filename = filename.numpy().decode()
data_type = data_type.numpy().decode()
num_labels = num_labels.numpy()
model_type = model_type.numpy().decode()
split = split.numpy().decode()
datset = dataset.numpy().decode()
if dataset == "voxceleb":
speaker = filename.split("/")[-3]
label = int(speaker[3:]) - 1 # there is no label 0
elif dataset == "music":
label = int(filename.split("/").split("_")[1])
elif dataset == "birdsong":
label = int(filename.split("/")[-3])
one_hot_label = np.zeros(num_labels)
one_hot_label[label] = 1
y = read_audio(filename, data_type)
# comment this if you do not want to learn with random crops
if len(y) < 4 * 16_000:
y = np.tile(y, 2)
y = crop(y, 4 * 16_000)
y = fp_to_pcm(enhance(y))
stft = create_spectrogram(y)
if split == "train" or split == "validation":
if model_type == "vggvox":
stft = np.expand_dims(stft[:, :300], axis=-1)
elif model_type == "lstm":
stft = stft[:, :300].T
elif model_type == "transformer":
stft = np.expand_dims(stft[:, :300].T, axis=-1)
elif split == "test":
if model_type == "vggvox":
stft = np.expand_dims(stft, axis=-1)
elif model_type == "lstm":
stft = stft.T
stft = [
stft[i * 300:(i + 1) * 300]
for i in range(stft.shape[0] // 300)
]
stft = np.array(stft)
elif model_type == "transformer":
stft = stft.T
stft = [
stft[i * 300:(i + 1) * 300]
for i in range(stft.shape[0] // 300)
]
stft = np.expand_dims(np.array(stft), axis=-1)
return stft, one_hot_label
#open all images in source and store all filenames for search afterwards path = '/home/zhangzimou/Desktop/code_lnk/database/FVC2002/DB1_B' size, = np.shape(os.listdir(path)) filename = [1]*size #ten different fingerprints. eight images for same fingerprint count = 0 for file in os.listdir(path): filename[count] = file count += 1 #search all filenames for same images and acheive correct probability matchedPoints_max = np.zeros([size,1]) score = np.zeros([size,1]) correct = 0 wrong = 0 i = 0 img1 = cv2.imread(path+'/'+filename[i],0) imgE1,imgfore1=enhance(img1) imgB1=basic.binarize(imgE1) imgT1=pre.thinning(imgB1) for j in range(size): print "Image %d & Image %d" % (i,j) #input img2 = cv2.imread(path+'/'+filename[j],0) #decide if they same fingerprints due to index of files if i/8 == j/8: print "They are from same fingerprint" else: print "They are from different fingerprint" #preprocess for images imgE2,imgfore2=enhance(img2)
def data_generator_multi(images,
masks,
config,
shuffle=True,
augment=False,
batch_size=None,
tp_value=8):
batch_size = config.BATCH_SIZE if batch_size is None else batch_size
iterator = num_iterator(masks.shape[0], batch_size, shuffle=shuffle)
while True:
ind_batch = next(iterator)
image_batch = images[ind_batch].copy().astype('float32')
mask_batch = masks[ind_batch].copy()
if augment:
tp = np.random.randint(tp_value)
if tp == 1:
if np.random.rand() > 0.5:
image_batch = image_batch[:, :, ::-1]
mask_batch = mask_batch[:, :, ::-1]
else:
image_batch = image_batch[:, ::-1]
mask_batch = mask_batch[:, ::-1]
elif tp == 2:
image_batch, mask_batch = affine_transform_batch(
image_batch,
mask_batch,
rotation_range=config.ROTATION_RANGE,
width_shift_range=config.WIDTH_SHIFT_RANGE,
height_shift_range=config.HEIGHT_SHIFT_RANGE,
shear_range=config.SHEAR_RANGE,
zoom_range=config.ZOOM_RANGE)
elif tp == 3:
image_batch, mask_batch = elastic_transform_batch(
image_batch,
mask_batch,
alpha=config.ALPHA,
sigma=config.SIGMA)
elif tp == 4:
tries = np.random.randint(20, 100)
max_fails = 5
image_batch, mask_batch = insert(image_batch, mask_batch,
tries, max_fails)
elif tp == 5:
image_batch = enhance(image_batch,
mask_batch,
max_tries=10,
max_enhance_ratio=0.8)
elif tp == 6:
image_batch = level_noise(image_batch,
mask_batch,
max_level_ratio=4,
max_noise_ratio=0.05)
elif tp == 7:
image_batch, mask_batch = stretch(image_batch,
mask_batch,
max_ratio=2)
'''
image_max = np.max(image_batch, axis=(1,2,3), keepdims=True)
image_max[image_max==0]=1
amax = np.random.randint(200, 256) if augment else 255
image_batch = image_batch/image_max*amax
'''
yield [image_batch.astype('float32'), mask_batch]
from _match import minutiaeMatch FVC0='/home/zhangzimou/Desktop/code_lnk/database/FVC2000/' FVC2='/home/zhangzimou/Desktop/code_lnk/database/FVC2002/' FVC4='/home/zhangzimou/Desktop/code_lnk/database/FVC2004/' path=FVC2+'DB1_B/' blockSize=8 img=cv2.imread(path+'108_3.tif',0) imgE,imgfore=enhance(img) imgB=basic.binarize(imgE) imgT=pre.thinning(imgB) #plt.figure() #imshow(img) plt.figure() plt.imshow(imgE,cmap='gray') plt.figure() plt.imshow(imgT,cmap='gray') ending,bifur,theta1,theta2=minutiaeExtract(imgT,imgfore) plt.plot(ending.T[1],ending.T[0],'b.',bifur.T[1],bifur.T[0],'r.') plt.quiver(ending.T[1],ending.T[0],np.cos(theta1),np.sin(-theta1),color='b',width=0.003) plt.quiver(bifur.T[1],bifur.T[0],np.cos(theta2),np.sin(-theta2),color='r',width=0.003)
import preprocess as pre
from minutiaeExtract import minutiaeExtract
from basic import block_view
FVC0 = '/home/zhangzimou/Desktop/code_lnk/database/FVC2000/'
FVC2 = '/home/zhangzimou/Desktop/code_lnk/database/FVC2002/'
FVC4 = '/home/zhangzimou/Desktop/code_lnk/database/FVC2004/'
path = FVC4 + 'DB4_B/'
blockSize = 8
img = cv2.imread(path + '102_1.tif', 0)
img_seg, imgfore = pre.segmentation(img)
img_en = enhance(img)
imgB = basic.binarize(img_en)
imgT = pre.thinning(imgB)
plt.figure()
imshow(img_seg)
plt.figure()
plt.imshow(imgB, cmap='gray')
plt.figure()
plt.imshow(imgT, cmap='gray')
ending, bifur, theta1, theta2 = minutiaeExtract(imgT, imgfore)
plt.plot(ending.T[1], ending.T[0], 'b.', bifur.T[1], bifur.T[0], 'r.')
plt.quiver(ending.T[1],
ending.T[0],
np.cos(theta1),
from preprocess import enhance import preprocess as pre from minutiaeExtract import minutiaeExtract import minutiaeExtract as extract FVC0 = '/home/zhangzimou/Desktop/code_lnk/database/FVC2000/' FVC2 = '/home/zhangzimou/Desktop/code_lnk/database/FVC2002/' FVC4 = '/home/zhangzimou/Desktop/code_lnk/database/FVC2004/' path = FVC4 + 'DB1_B/' start = time.clock() img = cv2.imread(path + '105_1.tif', 0) image, imgfore = pre.segmentation(img) image = enhance(img) core_index, delta_index = extract.singular(image, imgfore) blockSize = 8 #theta=pre.calcDirection(image,blockSize,method='block-wise') # #P =[ theta[2:,1:-1], theta[2:,2:], theta[1:-1,2:], theta[:-2,2:], theta[:-2,1:-1],theta[:-2,:-2], theta[1:-1,:-2], theta[2:,:-2]] # # # #N,M=image.shape #N1,M1=theta.shape #delta=np.zeros((N1-2,M1-2)) #for i in range(8): # if i==7: # de=P[0]-P[7]