def noise(image, mrStruct, new_flow):
k = np.squeeze(image[0])
print(k.shape)
#input_4 = tf.convert_to_tensor(k, dtype=tf.float32)
#input_4 = input_[32:160,...]
#input_4 = tf.image.resize_image_with_crop_or_pad(input_4,128,96)
#input_4 = (input_4 - tf.reduce_min(input_4))/(tf.reduce_max(input_4) - tf.reduce_min(input_4))
#input_4 = tf.expand_dims(input_4, dim=0)
#input_4 = tf.expand_dims(input_4, dim=4)
#with graph.as_default():
input_4 = tf.convert_to_tensor(k, dtype=tf.float32)
input_4 = tf.expand_dims(input_4, dim=0)
#input_4 = tf.expand_dims(input_4, dim=4)
flag = tf.placeholder(tf.bool)
input_2p = tf.placeholder(tf.float32, shape=[1, 160,96,None, 2])
#input_2p = tf.placeholder(tf.float32, shape=[1, 160,130,None, 2])
logits_2 = Unet(x = input_2p, training=flag).model
#logits_2 = Unet2(x = input_2p, training=flag).model
llogits_2 = tf.nn.softmax(logits_2)
#sesss = tf.Session(graph=graph)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sesss:
sesss.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
#checkpoint1 = tf.train.get_checkpoint_state("E:/HB/scripts_and_stuff/aliasing/new_noise")
#sesss.run(tf.local_variables_initializer())
s = sesss.run([input_4])
print(s[0].shape)
saver.restore(sesss, tf.train.get_checkpoint_state("./new_noise").model_checkpoint_path)
h , seg3 = sesss.run([logits_2, llogits_2], feed_dict={input_2p:s[0], flag: True})
sesss.close()
tf.keras.backend.clear_session()
mask2 = np.squeeze(seg3[...,1])
mask2 = mask2>0.5
mask2 = np.logical_not(mask2)
mask2 = mask2.astype(np.float32)
mask2 = np.expand_dims(mask2,axis=3)
mask2 = np.expand_dims(mask2,axis=4)
final_flow = np.multiply(new_flow, mask2)
#mrStruct1 = flows['mrStruct']
mrStruct['dataAy'][0,0] = final_flow
#mag = mags['mrStruct']['dataAy'][0,0]
#mag = mag[h1:HH-h1,w1:WW-w1,...]
#mrStruct = mags['mrStruct']
#mrStruct['dataAy'][0,0] = mag
print(final_flow.shape)
io.savemat('MK/vel_struct.mat',{'mrStruct':mrStruct})
#io.savemat('MK/mag_struct.mat',{'mrStruct':mrStruct})
io.savemat('MK/noise_mask.mat',{'noise':mask2})
#del mrStruct1
return image, mrStruct
def noise(image, new_flow):
k = np.squeeze(image[0])
print(k.shape)
input_4 = tf.convert_to_tensor(k, dtype=tf.float32)
input_4 = tf.expand_dims(input_4, dim=0)
#input_4 = tf.expand_dims(input_4, dim=4)
flag = tf.placeholder(tf.bool)
input_2p = tf.placeholder(tf.float32, shape=[1, 160, 96, None, 2])
#input_2p = tf.placeholder(tf.float32, shape=[1, 160,130,None, 2])
logits_2 = Unet(x=input_2p, training=flag).model
#logits_2 = Unet2(x = input_2p, training=flag).model
llogits_2 = tf.nn.softmax(logits_2)
#sesss = tf.Session(graph=graph)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
#config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
#checkpoint1 = tf.train.get_checkpoint_state("E:/HB/scripts_and_stuff/aliasing/new_noise")
#sesss.run(tf.local_variables_initializer())
s = sess.run([input_4])
print(s[0].shape)
saver.restore(
sess,
tf.train.get_checkpoint_state(
"/opt/codes/python-ismrmrd-server/new_noise").
model_checkpoint_path)
h, seg3 = sess.run([logits_2, llogits_2],
feed_dict={
input_2p: s[0],
flag: True
})
sess.close()
tf.keras.backend.clear_session()
mask2 = np.squeeze(seg3[..., 1])
mask2 = mask2 > 0.5
#mask2 = np.logical_not(mask2)
mask2 = mask2.astype(np.float32)
if new_flow.shape[1] < 96:
W = new_flow.shape[1]
#h1 = np.abs((HH - 160))/2
w1 = np.abs((W - 96)) / 2
w1 = w1.astype(np.int16)
mask2 = mask2[:, w1:96 - w1, :]
mask2 = np.expand_dims(mask2, axis=3)
mask2 = np.expand_dims(mask2, axis=4)
final_flow = np.multiply(new_flow, mask2)
print(final_flow.shape)
return image, final_flow
def segment(image):
k = np.squeeze(image[0])
#mrStruct = io.loadmat('MK/mag_struct.mat')
print(k.shape)
k = k[..., 0]
input_5 = tf.convert_to_tensor(k, dtype=tf.float32)
input_5 = tf.expand_dims(input_5, dim=0)
input_5 = tf.expand_dims(input_5, dim=4)
flag = tf.placeholder(tf.bool)
input_3p = tf.placeholder(tf.float32, shape=[1, 160, 96, None, 1])
#input_3p = tf.placeholder(tf.float32, shape=[1, 160,130,None, 1])
logits_3 = Unet(x=input_3p, training=flag).model
#logits_2 = Unet2(x = input_2p, training=flag).model
llogits_3 = tf.nn.softmax(logits_3)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(tf.global_variables())
#checkpoint1 = tf.train.get_checkpoint_state("E:/HB/scripts_and_stuff/aliasing/new_noise")
#sesss.run(tf.local_variables_initializer())
s = sess.run([input_5])
print(s[0].shape)
saver.restore(
sess,
tf.train.get_checkpoint_state(
"/opt/codes/python-ismrmrd-server/pre-trained").
model_checkpoint_path)
h, seg4 = sess.run([logits_3, llogits_3],
feed_dict={
input_3p: s[0],
flag: True
})
sess.close()
mask3 = np.squeeze(seg4[..., 1])
mask3 = mask3 > 0.5
#mrStruct = mags['mrStruct']
#mrStruct['mrStruct']['dataAy'][0,0] = mask3
#mrStruct['mrStruct']['dim4'] = 'unused'
#mrStruct['mrStruct']['dim5'] = 'unused'
return mask3
def eddy(mag, flow):
imaspeed = np.power(np.mean(np.square(flow), axis=3), 0.5)
minV = np.amin(0.7 * np.reshape(mag, (-1, 1)))
maxV = np.amax(0.7 * np.reshape(mag, (-1, 1)))
tmp = mag > maxV
not_tmp = np.logical_not(tmp)
mag = np.multiply(mag, not_tmp) + np.multiply(tmp, maxV)
tmp = mag < minV
not_tmp = np.logical_not(tmp)
mag = np.multiply(mag, not_tmp) + np.multiply(tmp, minV)
mag = (mag - minV) / (maxV - minV)
pcmra = np.multiply(imaspeed, mag)
pcmra = np.squeeze(np.mean(np.square(pcmra), axis=3))
std_flow = np.std(imaspeed, axis=3)
#del flow
#del mag
print(pcmra.shape)
print(std_flow.shape)
pcmra = np.expand_dims(pcmra, axis=3)
std_flow = np.expand_dims(std_flow, axis=3)
X = np.concatenate((pcmra, std_flow), axis=3)
X = X.astype(np.float32)
del pcmra
del std_flow
for j in range(2):
X[..., j] = (X[..., j] - np.amin(X[..., j])) / (np.amax(X[..., j]) -
np.amin(X[..., j]))
#X = X[10:138,...]
print(X.shape)
input_ = tf.convert_to_tensor(X, dtype=tf.float32)
input_ = tf.transpose(input_, perm=[2, 0, 1, 3])
input_ = tf.image.resize_image_with_crop_or_pad(input_, 160, 96)
#input_ = tf.image.resize_image_with_crop_or_pad(input_, 160, 130)
input_ = tf.transpose(input_, perm=[1, 2, 0, 3])
input_ = tf.squeeze(input_)
#input_ = (input_ - tf.reduce_min(input_))/(tf.reduce_max(input_) - tf.reduce_min(input_))
input_ = tf.expand_dims(input_, dim=0)
#input_ = tf.expand_dims(input_, dim=4)
input_p = tf.placeholder(tf.float32, shape=[None, 160, 96, None, 2])
#input_p = tf.placeholder(tf.float32, shape=[None, 160,130,None, 2])
flag = tf.placeholder(tf.bool)
logits_1 = Unet(x=input_p, training=flag).model
saver1 = tf.train.Saver()
llogits_1 = tf.nn.softmax(logits_1)
checkpoint1 = tf.train.get_checkpoint_state(
"/opt/codes/python-ismrmrd-server/new_eddy")
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
#config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
image = sess.run([input_])
saver1.restore(sess, checkpoint1.model_checkpoint_path)
h, seg1 = sess.run([logits_1, llogits_1],
feed_dict={
input_p: image[0],
flag: True
})
#h , seg2 = sess.run([logits_1, llogits_1], feed_dict={input_1p: image[1], flag: True})
#h , seg3 = sess.run([logits_1, llogits_1], feed_dict={input_1p: image[2], flag: True})
sess.close()
tf.reset_default_graph()
tf.keras.backend.clear_session()
mask1 = seg1[..., 1]
mask1 = mask1 > 0.5
print(len(image))
#U = image[0]
#plt.show()
mask1 = np.squeeze(mask1)
print(mask1.shape)
HH = flow.shape[0]
WW = flow.shape[1]
h1 = np.abs((HH - 160)) / 2
w1 = np.abs((WW - 96)) / 2
print(h1)
print(w1)
h1 = np.int(h1)
w1 = np.int(w1)
if flow.shape[1] >= 96:
flow = flow[h1:HH - h1, w1:WW - w1, ...]
elif flow.shape[1] < 96:
#flows = np.zeros([flow[0].shape, 96, flow])
flow = flow[h1:HH - h1, ...]
mask1 = mask1[:, w1:96 - w1, :]
print(mask1.shape)
#print(w1)
#print(WW)
[H, W, D] = mask1.shape
xx = np.arange(W)
xxx = np.tile(xx, [H, 1])
print(xxx.shape)
yy = np.arange(H)
yy = yy[::-1]
yy = np.expand_dims(yy, axis=1)
#yy = np.fliplr([yy])[0]
#print(yy)
yyy = np.tile(yy, (1, W))
print(np.amin(flow))
print(flow.shape)
#yyy = np.fliplr(yyy)
yyy = yyy + 1
xxx = xxx + 1
print(yyy)
print(xxx)
#new_flow = np.zeros(flow.shape)
print(flow.shape)
for i in range(flow.shape[2]):
imaFlow = np.squeeze(flow[..., i, :, :])
#print(imaFlow.shape)
statMask = mask1[..., i]
statMask = statMask.astype(np.int16)
imaFlow1 = np.zeros(imaFlow.shape)
#statMask = statMask.astype(np.float32)
for k in range(imaFlow.shape[2]):
tt = flow.shape[4]
phi, alpha, beta = lsqf2(imaFlow[..., k, tt - 1], statMask, xxx,
yyy)
phi = phi.astype(np.float32)
alpha = alpha.astype(np.float32)
beta = beta.astype(np.float32)
fitPlane = phi + alpha * xxx + beta * yyy
zeroMask = imaFlow[:, :, k, tt - 1] != 0
zeroPlane = np.multiply(fitPlane, zeroMask)
#print(zeroPlane.shape)
zeroPlane = np.expand_dims(zeroPlane, axis=2)
factor3D = np.tile(zeroPlane, (1, 1, tt))
imaFlow1[:, :, k, :] = np.squeeze(imaFlow[:, :, k, :]) - factor3D
#print(phi)
#print(alpha)
#print(beta)
#print(Md)
flow[:, :, i, :, :] = imaFlow1
del imaFlow
del imaFlow1
print("Finished flow loop in Eddy Current")
del mag
#io.savemat('new_vel.mat',{'data':flow})
#io.savemat('eddy_mask.mat',{'data':mask1})
return image, flow
def Unet_test():
image_batch_placeholder = tf.placeholder(tf.float32, shape=[None, 160,96,None, 1])
#label_batch_placeholder = tf.placeholder(tf.float32, shape=[None, 160, 96,None])
#labels_pixels = tf.reshape(label_batch_placeholder, [-1, 1]) # if_training_placeholder = tf.placeholder(tf.bool, shape=[])
training_flag = tf.placeholder(tf.bool)
image_batch, venc, phase = feed_data()
#label_batch_dense = tf.arg_max(label_batch, dimension = 1)
# if_training = tf.Variable(False, name='if_training', trainable=False)
logits = Unet(x = image_batch_placeholder, training=training_flag).model
#logits = logits>1
#logs = cost_dice(logits,label_batch_placeholder)
llogits = tf.nn.softmax(logits)
#logits = tf.argmax(logits,axis=3)
#logits = tf.reshape(logits,(-1, 2))
#logits_batch = tf.to_int64(tf.arg_max(logits, dimension = 3))
#logi = tf.nn.softmax(logits)
#N,S,W,C = logits.get_shape()
#GH = tf.reshape(logits,[-1, 1])
#GT = tf.nn.softmax(GT)
#GH = tf.nn.softmax(GT, axis=1)
#GH = tf.reshape(GT,[H.value,W.value])
#GH = tf.argmax(GH, axis = 1)
#GH = tf.reshape(GH,[S.value,W.value])
#probs = tf.slice(logits,[0, 0, 0, 1], [-1, -1, -1, -1])
#probs = tf.squeeze(probs, axis = -1)
#y = tf.reshape(label_batch_placeholder,[-1,2])
#y = tf.argmax(y, axis = 1)
#y = tf.reshape(y,[S.value,W.value])
#H = tf.cast(GH,tf.float32)
#y = tf.cast(y,tf.float32)
#inter = tf.reduce_sum(H * y)
#u = tf.constant(1e-5,dtype=tf.int64)
#union = 0.00001+ tf.reduce_sum(H) + tf.reduce_sum(y)
#logits = tl.activation.pixel_wise_softmax(logits)
#log = tf.reshape(logits,[-1,2])
#log = tf.argmax(log,axis = 1)
#log = tf.reshape(log,[S.value,W.value])
#correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels_pixels, 1))
#accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
checkpoint = tf.train.get_checkpoint_state('./alias')#"/media/haben/D4CC01B2CC018FC2/alis_og_crop") #good_alis #new_alis_z #new_alis_y
saver = tf.train.Saver()
#gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
# logits_batch = tf.to_int64(tf.arg_max(logits, dimension = 1))
d = 0
config = tf.ConfigProto(log_device_placement=False)
all_trainable_vars = tf.reduce_sum([tf.reduce_prod(v.shape) for v in tf.trainable_variables()])
config.gpu_options.allow_growth=True
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
#tf.logging.info("Restoring full model from checkpoint file %s",checkpoint.model_checkpoint_path)
saver.restore(sess, checkpoint.model_checkpoint_path)
#accuracy_accu = 0
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord, sess = sess)
ave = []
ma = []
tru = []
use = []
gtruth = []
low = []
low_true = []
low_loss = []
new_data = []
phases = sess.run(phase)
print(phases)
for i in (range(int(phases))):
image_out, Venc = sess.run([image_batch, venc])
_, llogit = sess.run([logits, llogits], feed_dict={image_batch_placeholder: image_out,
#label_batch_placeholder: truth,
training_flag: True})
infer_out = llogit[...,1]
print(i)
#lo = np.squeeze(soft[...,0])
#print(logis.shape)
data = np.squeeze(infer_out)
#print(Venc)
#print(ys.shape)
#gt = np.squeeze(Y)
mask = data.transpose()
#use.append(mask>0.2)
#gt = gt.transpose()
h = data>0.2
#print(np.max(alias))
#alias = np.squeeze(alias)
#data = data>0.2
#data = data[...,1]
#im = np.squeeze(image_out)
#im = im[...,1]
#plt.imshow(im)
#plt.show()
#plt.imshow(data)
#plt.show()
#new_alis = alias
"""
for i in range(alias.shape[2]):
for x in range(alias.shape[0]):
for y in range(alias.shape[1]):
if h[x,y,i] == 1:
value = alias[x,y,i]
new = value - (np.sign(value) * Venc*2/100)
new_alis[x,y,i] = new
else:
continue
new_data.append(new_alis)
"""
#mask2 = alias>=0
#mask1 = mask2 * h
#new_alis[mask1] = new_alis[mask1] - (Venc*2/100)
#mask3 = np.invert(mask2)
#mask4 = mask3 * h
#new_alis[mask4] = new_alis[mask4] + (Venc*2/100)
new_data.append(h)
#print(np.unique(new_alis))
#ys = (ys)
#print(W.shape)
#plt.pause(0.1)
#plt.imshow(alias[...,2])
#plt.show()
#plt.imshow(mask1[...,2])
#plt.show()
#plt.imshow(mask4[...,2])
#plt.show()
#plt.pause(0.1)
io.savemat('./new_vel.mat',{'data':new_data})
#tf.train.write_graph(sess.graph_def, 'graph/', 'my_graph.pb', as_text=False)
coord.request_stop()
coord.join(threads)
sess.close()
tf.keras.backend.clear_session()
tf.reset_default_graph()
return new_data
input_1 = tf.image.resize_image_with_crop_or_pad(input_1, 128, 96)
h1 = (h - 128)
input_2 = input_[h1:h, ...]
input_2 = tf.image.resize_image_with_crop_or_pad(input_2, 128, 96)
input_1 = (input_1 - tf.reduce_min(input_1)) / (tf.reduce_max(input_1) -
tf.reduce_min(input_1))
input_1 = tf.expand_dims(input_1, dim=0)
input_1 = tf.expand_dims(input_1, dim=4)
input_1p = tf.placeholder(tf.float32, shape=[1, 128, 96, None, 1])
input_2 = (input_2 - tf.reduce_min(input_2)) / (tf.reduce_max(input_2) -
tf.reduce_min(input_2))
input_2 = tf.expand_dims(input_2, dim=0)
input_2 = tf.expand_dims(input_2, dim=4)
flag = tf.placeholder(tf.bool)
logits_1 = Unet(x=input_1p, training=flag).model
saver1 = tf.train.Saver()
#logits_1 = Unet(x = input_1p, training=flag).model
llogits_1 = tf.nn.softmax(logits_1)
checkpoint1 = tf.train.get_checkpoint_state("./pre-trained")
checkpoint1 = tf.train.get_checkpoint_state("./pre-trained")
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
def eddy(path):
flows = io.loadmat(os.path.join(os.path.sep, path, 'vel_struct.mat'))
mags = io.loadmat(os.path.join(os.path.sep, path, 'mag_struct.mat'))
flow = flows['mrStruct']
mag = mags['mrStruct']
flow = flow['dataAy']
mag = mag['dataAy']
flow = flow[0, 0]
mag = mag[0, 0]
imaspeed = np.power(np.mean(np.square(flow), axis=3), 0.5)
minV = np.amin(0.7 * np.reshape(mag, (-1, 1)))
maxV = np.amax(0.7 * np.reshape(mag, (-1, 1)))
tmp = mag > maxV
not_tmp = np.logical_not(tmp)
mag = np.multiply(mag, not_tmp) + np.multiply(tmp, maxV)
tmp = mag < minV
not_tmp = np.logical_not(tmp)
mag = np.multiply(mag, not_tmp) + np.multiply(tmp, minV)
mag = (mag - minV) / (maxV - minV)
pcmra = np.multiply(imaspeed, mag)
pcmra = np.squeeze(np.mean(np.square(pcmra), axis=3))
std_flow = np.std(imaspeed, axis=3)
del flow
del mag
#print(list(f.keys()))
#print(list(g.keys()))
#tv = f['test']
#la = g['test']
#print(list(tv.keys()))
#flow = f[tv['data'][0,0]].value
#labe = g[la['truth'][0,0]].value
#mask = g[la['mask'][0,0]].value
#print(flow.dtype)
#print(labe.dtype)
#print(mask.dtype)
#print(la)
#te = k['test']
#print(list(la.keys()))
#flow = f[tv['data'][0,0]].value
#labe = g[la['data'][0,0]].value
#test = k[te['data'][0,0]].value
#print( flow.shape)
#print(tv.shape)
#for i in range(1):
#flow = f[tv['data'][i,0]].value
#labe = g[la['truth'][i,0]].value
#mask = g[la['mask'][i,0]].value
print(pcmra.shape)
print(std_flow.shape)
pcmra = np.expand_dims(pcmra, axis=3)
std_flow = np.expand_dims(std_flow, axis=3)
X = np.concatenate((pcmra, std_flow), axis=3)
X = X.astype(np.float32)
del pcmra
del std_flow
#Y = labe.transpose()
#Z = mask.transpose()
#dim1 = X.shape[0]
#dim2 = X.shape[1]
#h1 = round((dim1-160)/2)
#w2 = round((dim2-80)/2)
#X = X.astype(np.float32)
#Y = Y.astype(np.float32)
#X = X[:140,:,:]
#Y = Y[:140,:,:]
#Z = Z[:140,:,:]
for j in range(2):
X[..., j] = (X[..., j] - np.amin(X[..., j])) / (np.amax(X[..., j]) -
np.amin(X[..., j]))
#X = X[10:138,...]
print(X.shape)
#images.append(X)
#labels.append(Y)
#masks.append(Z)
#print(len(images))
#print(len(labels))
#print(images[0].shape)
#print(labels[0].shape)
#print(masks[0].shape)
#print(images[0].dtype)
#print(labels[0].dtype)
#print(np.unique(labels[0]))
#print(np.unique(masks[0]))
"""
def _bytes_feature(value):
return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))
def _int64_feature(value):
return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))
#plt.imshow(train_image[0])
#plt.show()
s = list(range(len(images)))
print(s)
#random.shuffle(s)
test_filename = 'test_eddy.tfrecords'
writer = tf.python_io.TFRecordWriter(test_filename)
for i in s:
image_raw = images[i].tostring()
label_raw = labels[i].tostring()
#mask_raw = masks[i].tostring()
height = images[i].shape[0]
width = images[i].shape[1]
depth = images[i].shape[2]
features = {'test/image': _bytes_feature(image_raw),
'test/label': _bytes_feature(label_raw),
#'test/mask': _bytes_feature(mask_raw),
'test/height': _int64_feature(height),
'test/depth':_int64_feature(depth),
'test/width': _int64_feature(width)}
examples = tf.train.Example(features = tf.train.Features(feature = features))
writer.write(examples.SerializeToString())
writer.close()
"""
input_ = tf.convert_to_tensor(X, dtype=tf.float32)
input_ = tf.transpose(input_, perm=[2, 0, 1, 3])
input_ = tf.image.resize_image_with_crop_or_pad(input_, 160, 96)
#input_ = tf.image.resize_image_with_crop_or_pad(input_, 160, 130)
input_ = tf.transpose(input_, perm=[1, 2, 0, 3])
input_ = tf.squeeze(input_)
#input_ = (input_ - tf.reduce_min(input_))/(tf.reduce_max(input_) - tf.reduce_min(input_))
input_ = tf.expand_dims(input_, dim=0)
#input_ = tf.expand_dims(input_, dim=4)
input_p = tf.placeholder(tf.float32, shape=[None, 160, 96, None, 2])
#input_p = tf.placeholder(tf.float32, shape=[None, 160,130,None, 2])
flag = tf.placeholder(tf.bool)
logits_1 = Unet(x=input_p, training=flag).model
saver1 = tf.train.Saver()
llogits_1 = tf.nn.softmax(logits_1)
checkpoint1 = tf.train.get_checkpoint_state("./new_eddy")
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
image = sess.run([input_])
saver1.restore(sess, checkpoint1.model_checkpoint_path)
h, seg1 = sess.run([logits_1, llogits_1],
feed_dict={
input_p: image[0],
flag: True
})
#h , seg2 = sess.run([logits_1, llogits_1], feed_dict={input_1p: image[1], flag: True})
#h , seg3 = sess.run([logits_1, llogits_1], feed_dict={input_1p: image[2], flag: True})
sess.close()
tf.keras.backend.clear_session()
mask1 = seg1[..., 1]
mask1 = mask1 > 0.5
print(len(image))
#U = image[0]
#plt.imshow(np.squeeze(U[...,16,0]), cmap=plt.get_cmap('gray'))
#plt.imshow(np.squeeze(mask1[...,16]), alpha=0.3)
#plt.show()
mask1 = np.squeeze(mask1)
print(mask1.shape)
[H, W, D] = mask1.shape
xx = np.arange(W)
xxx = np.tile(xx, [H, 1])
print(xxx.shape)
yy = np.arange(H)
yy = yy[::-1]
yy = np.expand_dims(yy, axis=1)
#yy = np.fliplr([yy])[0]
#print(yy)
yyy = np.tile(yy, (1, W))
flow = flows['mrStruct']['dataAy'][0, 0]
HH = flow.shape[0]
WW = flow.shape[1]
h1 = (HH - 160) / 2
w1 = (WW - 96) / 2
print(h1)
print(w1)
h1 = np.int(h1)
w1 = np.int(w1)
flow = flow[h1:HH - h1, w1:WW - w1, ...]
#plt.imshow(np.squeeze(flow[...,16,0,5]), cmap=plt.get_cmap('gray'))
#plt.imshow(np.squeeze(mask1[...,16]), alpha=0.3)
#plt.show()
print(np.amin(flow))
print(flow.shape)
#yyy = np.fliplr(yyy)
yyy = yyy + 1
xxx = xxx + 1
print(yyy)
print(xxx)
#new_flow = np.zeros(flow.shape)
print(flow.shape)
for i in range(flow.shape[2]):
imaFlow = np.squeeze(flow[..., i, :, :])
#print(imaFlow.shape)
statMask = mask1[..., i]
statMask = statMask.astype(int)
imaFlow1 = np.zeros(imaFlow.shape)
#statMask = statMask.astype(np.float32)
for k in range(imaFlow.shape[2]):
tt = flow.shape[4]
phi, alpha, beta = lsqf2(imaFlow[..., k, tt - 1], statMask, xxx,
yyy)
phi = phi.astype(np.float32)
alpha = alpha.astype(np.float32)
beta = beta.astype(np.float32)
fitPlane = phi + alpha * xxx + beta * yyy
zeroMask = imaFlow[:, :, k, tt - 1] != 0
zeroPlane = np.multiply(fitPlane, zeroMask)
#print(zeroPlane.shape)
zeroPlane = np.expand_dims(zeroPlane, axis=2)
factor3D = np.tile(zeroPlane, (1, 1, tt))
imaFlow1[:, :, k, :] = np.squeeze(imaFlow[:, :, k, :]) - factor3D
#print(phi)
#print(alpha)
#print(beta)
#print(Md)
flow[:, :, i, :, :] = imaFlow1
del imaFlow
del imaFlow1
print("Finished flow loop in Eddy Current")
#time.sleep(10)
#print(flow.shape)
mrStruct = flows['mrStruct']
mag = mags['mrStruct']['dataAy'][0, 0]
mag = mag[h1:HH - h1, w1:WW - w1, ...]
mrStruct1 = mags['mrStruct']
mrStruct1['dataAy'][0, 0] = mag
io.savemat('MK/mag_struct.mat', {'mrStruct': mrStruct1})
del mag
del flows
io.savemat('MK/new_vel.mat', {'data': flow})
io.savemat('MK/eddy_mask.mat', {'data': mask1})
return image, mrStruct, flow
examples = tf.train.Example(features = tf.train.Features(feature = features))
writer.write(examples.SerializeToString())
writer.close()
"""
input_ = tf.convert_to_tensor(X, dtype=tf.float16)
input_ = tf.transpose(input_, perm=[2,0,1,3])
input_ = tf.image.resize_image_with_crop_or_pad(input_, 160, 96)
input_ = tf.transpose(input_, perm=[1,2,0,3])
#input_ = (input_ - tf.reduce_min(input_))/(tf.reduce_max(input_) - tf.reduce_min(input_))
#input_ = tf.expand_dims(input_, dim=0)
#input_ = tf.expand_dims(input_, dim=4)
input_p = tf.placeholder(tf.float32, shape=[1, 160,96,None, 2])
flag = tf.placeholder(tf.bool)
logits_1 = Unet(x = input_p, training=flag).model
saver1 = tf.train.Saver()
llogits_1 = tf.nn.softmax(logits_1)
checkpoint1 = tf.train.get_checkpoint_state("./new_eddy")
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.per_process_gpu_memory_fraction = 0.4
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
image = sess.run([input_])
saver1.restore(sess, checkpoint1.model_checkpoint_path)
h , seg1 = sess.run([logits_1, llogits_1], feed_dict={input_p: image, flag: True})
#h , seg2 = sess.run([logits_1, llogits_1], feed_dict={input_1p: image[1], flag: True})
#h , seg3 = sess.run([logits_1, llogits_1], feed_dict={input_1p: image[2], flag: True})