def train(img_x, img_y, patch_size, patch_step, dim_img, nb_filters, nb_conv, batch_size, nb_epoch):
"""
Function description.
Parameters
----------
parameter_01 : type
Description.
parameter_02 : type
Description.
parameter_03 : type
Description.
Returns
-------
return_01
Description.
"""
img_x = utils.nor_data(img_x)
img_y = utils.nor_data(img_y)
img_input = utils.extract_patches(img_x, patch_size, patch_step)
img_output = utils.extract_patches(img_y, patch_size, patch_step)
img_input = np.reshape(img_input, (len(img_input), 1, dim_img, dim_img))
img_output = np.reshape(img_output, (len(img_input), 1, dim_img, dim_img))
mdl = model(dim_img, nb_filters, nb_conv)
mdl.fit(img_input, img_output, batch_size=batch_size, nb_epoch=nb_epoch)
return mdl
def train(img_x, img_y, patch_size, patch_step, dim_img, nb_filters, nb_conv,
batch_size, nb_epoch):
"""
Function description.
Parameters
----------
parameter_01 : type
Description.
parameter_02 : type
Description.
parameter_03 : type
Description.
Returns
-------
return_01
Description.
"""
img_x = utils.nor_data(img_x)
img_y = utils.nor_data(img_y)
img_input = utils.extract_patches(img_x, patch_size, patch_step)
img_output = utils.extract_patches(img_y, patch_size, patch_step)
img_input = np.reshape(img_input, (len(img_input), 1, dim_img, dim_img))
img_output = np.reshape(img_output, (len(img_input), 1, dim_img, dim_img))
mdl = model(dim_img, nb_filters, nb_conv)
mdl.fit(img_input, img_output, batch_size=batch_size, nb_epoch=nb_epoch)
return mdl
def seg_train(img_x, img_y, patch_size = 32,
patch_step = 1, nb_conv = 32, size_conv = 3,
batch_size =1000, nb_epoch = 20, nb_down = 2, nb_gpu = 1):
"""
Function description.
Parameters
----------
img_x: array, 2D or 3D
Training input of the model. It is the raw image for the segmentation.
img_y: array, 2D or 3D
Training output of the model. It is the corresponding segmentation of the training input.
patch_size: int
The size of the small patches extracted from the input images. This size should be big enough to cover the
features of the segmentation object.
patch_step: int
The pixel steps between neighbour patches. Larger steps leads faster speed, but less quality. I recommend 1
unless you need quick test of the algorithm.
nb_conv: int
Number of the covolutional kernals for the first layer. This number doubles after each downsampling layer.
size_conv: int
Size of the convolutional kernals.
batch_size: int
Batch size for the training. Bigger size leads faster speed. However, it is restricted by the memory size of
the GPU. If the user got the memory error, please decrease the batch size.
nb_epoch: int
Number of the epoches for the training. It can be understand as the number of iterations during the training.
Please define this number as the actual convergence for different data.
nb_down: int
Number of the downsampling for the images in the model.
nb_gpu: int
Number of GPUs you want to use for the training.
Returns
-------
mdl
The trained CNN model for segmenation. The model can be saved for future segmentations.
"""
# if img_x.ndim == 3:
# _, ih, iw = img_x.shape
# else:
# ih, iw = img_x.shape
patch_shape = (patch_size, patch_size)
# print img_x.shape
# print img_x.max(), img_x.min()
img_x = nor_data(img_x)
img_y = nor_data(img_y)
# print img_x.shape
# print img_x.max(), img_x.min()
train_x = extract_3d(img_x, patch_shape, patch_step)
train_y = extract_3d(img_y, patch_shape, patch_step)
# print train_x.shape
# print train_x.max(), train_x.min()
train_x = np.reshape(train_x, (len(train_x), patch_size, patch_size, 1))
train_y = np.reshape(train_y, (len(train_y), patch_size, patch_size, 1))
mdl = model_choose(patch_size, patch_size, nb_conv, size_conv, nb_down, nb_gpu)
print(mdl.summary())
mdl.fit(train_x, train_y, batch_size=batch_size, epochs=nb_epoch)
return mdl
def predict(mdl, img, patch_size, patch_step, batch_size, dim_img):
"""
the cnn model for image transformation
Parameters
----------
img : array
The image need to be calculated
patch_size : (int, int)
The patches dimension
dim_img : int
The input image dimension
Returns
-------
img_rec
Description.
"""
img = np.float16(utils.nor_data(img))
img_y, img_x = img.shape
x_img = utils.extract_patches(img, patch_size, patch_step)
x_img = np.reshape(x_img, (len(x_img), 1, dim_img, dim_img))
y_img = mdl.predict(x_img, batch_size=batch_size)
del x_img
y_img = np.reshape(y_img, (len(y_img), dim_img, dim_img))
img_rec = utils.reconstruct_patches(y_img, (img_y, img_x), patch_step)
return img_rec
def predict(mdl, img, patch_size, patch_step, batch_size, dim_img):
"""
the cnn model for image transformation
Parameters
----------
img : array
The image need to be calculated
patch_size : (int, int)
The patches dimension
dim_img : int
The input image dimension
Returns
-------
img_rec
Description.
"""
img = np.float16(utils.nor_data(img))
img_y, img_x = img.shape
x_img = utils.extract_patches(img, patch_size, patch_step)
x_img = np.reshape(x_img, (len(x_img), 1, dim_img, dim_img))
y_img = mdl.predict(x_img, batch_size=batch_size)
del x_img
y_img = np.reshape(y_img, (len(y_img), dim_img, dim_img))
img_rec = utils.reconstruct_patches(y_img, (img_y, img_x), patch_step)
return img_rec
def predict(mdl, img, patch_size, patch_step, batch_size, dim_img):
"""
the cnn model for image transformation
Parameters
----------
img : array
The image need to be calculated
patch_size : (int, int)
The patches dimension
dim_img : int
The input image dimension
Returns
-------
img_rec
Description.
"""
img = np.float16(utils.nor_data(img))
img_h, img_w = img.shape
input_img = utils.extract_patches(img, patch_size, patch_step)
input_img = np.reshape(input_img, (input_img.shape[0], dim_img, dim_img, 1))
output_img = mdl.predict(input_img, batch_size=batch_size)
del input_img
output_img = np.reshape(output_img, (output_img.shape[0], dim_img, dim_img))
img_rec = utils.reconstruct_patches(output_img, (img_h, img_w), patch_step)
return img_rec
mdl.load_weights('weight_center.h5')
print('The model loading time is %s seconds'%(time.time()-start_time))
Y_score = np.zeros((50, 501))
for i in range(50):
slice_num = (i+2)*20
datapath = '/home/oxygen/YANGX/Globus/center/test_04/slice'+str(slice_num)+'/*.tiff'
# print(datapath)
fnames = glob.glob(datapath)
fnames = np.sort(fnames)
# print(fnames)
for j in range(len(fnames)):
img = dxchange.read_tiff(fnames[j])
img = -nor_data(img)
# X_evl = np.zeros((nb_evl, dim_img, dim_img))
# for k in range(nb_evl):
# X_evl[k] = img_window(img[360:1660, 440:1640], dim_img)
X_evl = extract_patches(img[360:1660, 440:1640],
(128, 128), step=64, max_patches=None, random_state=None)
X_evl = X_evl.reshape(X_evl.shape[0], 1, dim_img, dim_img)
Y_evl = mdl.predict(X_evl, batch_size=batch_size)
Y_score[i, j] = sum(np.dot(Y_evl, [0, 1]))
# print(Y_score[i])
#print('The evaluate score is:', Y_score[i])
#Y_score = sum(np.round(Y_score))/len(Y_score)
ind_max = np.argmax(Y_score[i, :])
print('The well-centered reconstruction is:', fnames[ind_max])
mdl.load_weights('weight_center.h5')
print('The model loading time is %s seconds' % (time.time() - start_time))
Y_score = np.zeros((50, 501))
for i in range(50):
slice_num = (i + 2) * 20
datapath = '/home/oxygen/YANGX/Globus/center/test_04/slice' + str(
slice_num) + '/*.tiff'
# print(datapath)
fnames = glob.glob(datapath)
fnames = np.sort(fnames)
# print(fnames)
for j in range(len(fnames)):
img = dxchange.read_tiff(fnames[j])
img = -nor_data(img)
# X_evl = np.zeros((nb_evl, dim_img, dim_img))
# for k in range(nb_evl):
# X_evl[k] = img_window(img[360:1660, 440:1640], dim_img)
X_evl = extract_patches(img[360:1660, 440:1640], (128, 128),
step=64,
max_patches=None,
random_state=None)
X_evl = X_evl.reshape(X_evl.shape[0], 1, dim_img, dim_img)
Y_evl = mdl.predict(X_evl, batch_size=batch_size)
Y_score[i, j] = sum(np.dot(Y_evl, [0, 1]))
# print(Y_score[i])
#print('The evaluate score is:', Y_score[i])
#Y_score = sum(np.round(Y_score))/len(Y_score)
def rec_dcgan_back(prj, ang, save_wpath, init_wpath=None, **kwargs):
tf.reset_default_graph()
cnn_kwargs = [
'learning_rate', 'num_steps', 'display_step', 'conv_nb', 'conv_size',
'dropout', 'weights_init', 'method', 'cost_rate'
]
kwargs_defaults = _get_tomolearn_kwargs()
for kw in cnn_kwargs:
kwargs.setdefault(kw, kwargs_defaults[kw])
if init_wpath:
kwargs['weights_init'] = True
_, nang, px, _ = prj.shape
prj = nor_data(prj)
img_input = tf.placeholder(tf.float32, prj.shape)
img_output = tf.placeholder(tf.float32, prj.shape)
pred, recon = tomo_learn(img_input,
ang,
px,
reuse=False,
conv_nb=kwargs['conv_nb'],
conv_size=kwargs['conv_size'],
dropout=kwargs['dropout'],
method=kwargs['method'])
disc_real = discriminator(img_output)
disc_fake = discriminator(pred, reuse=True)
gen_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,
labels=tf.ones_like(disc_fake))) \
+ tf.reduce_mean(tf.abs(img_output-pred))*kwargs['cost_rate']
disc_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=disc_real, labels=tf.ones_like(disc_real)))
disc_loss_fake = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=disc_fake, labels=tf.zeros_like(disc_fake)))
disc_loss = disc_loss_real + disc_loss_fake
gen_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='generator')
disc_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='discriminator')
optimizer_gen = tf.train.AdamOptimizer(
learning_rate=kwargs['learning_rate'])
optimizer_disc = tf.train.AdamOptimizer(
learning_rate=kwargs['learning_rate'])
train_gen = optimizer_gen.minimize(gen_loss, var_list=gen_vars)
train_disc = optimizer_disc.minimize(disc_loss, var_list=disc_vars)
fig, axs = plt.subplots(2, 1, figsize=(8, 16))
init = tf.global_variables_initializer()
saver = tf.train.Saver()
with tf.Session() as sess:
# Run the initializer
sess.run(init)
if kwargs['weights_init']:
if init_wpath == None:
print('Please provide the file name of initial weights.')
saver.restore(sess, init_wpath)
for step in range(1, kwargs['num_steps'] + 1):
with tf.device('/device:GPU:1'):
dl, _ = sess.run([disc_loss, train_disc],
feed_dict={
img_input: prj,
img_output: prj
})
with tf.device('/device:GPU:2'):
gl, _ = sess.run([gen_loss, train_gen],
feed_dict={
img_input: prj,
img_output: prj
})
if step % kwargs['display_step'] == 0 or step == 1:
pred, recon = sess.run(
tomo_learn(prj,
ang,
px,
reuse=True,
conv_nb=kwargs['conv_nb'],
conv_size=kwargs['conv_size'],
dropout=kwargs['dropout'],
method=kwargs['method']))
sino_plt = np.reshape(pred, (nang, px))
rec_plt = np.reshape(recon, (px, px))
#
ax = axs[0]
ax.imshow(sino_plt, vmax=1, cmap='jet')
plt.axis('off')
ax = axs[1]
ax.imshow(rec_plt, vmax=1, cmap='jet')
plt.axis('off')
plt.pause(0.1)
print("Step " + str(step) + ", Generator Loss= " +
"{:.7f}".format(gl) + ', Discriminator loss = ' +
"{:.7f}".format(dl))
plt.close(fig)
saver.save(sess, save_wpath)
return recon
def phase_dcgan(ifp, h, save_wpath, init_wpath = None, **kwargs):
ops.reset_default_graph()
tf.compat.v1.disable_eager_execution()
cnn_kwargs = ['pure_phase', 'learning_rate_g', 'learning_rate_d', 'num_steps', 'display_step', 'conv_nb',
'conv_size', 'dropout', 'weights_init','cost_rate', 'gl_tol', 'iter_plot']
kwargs_defaults = _get_phaselearn_kwargs()
for kw in cnn_kwargs:
kwargs.setdefault(kw, kwargs_defaults[kw])
if init_wpath:
kwargs['weights_init'] = True
_, px, px, _ = ifp.shape
img_input = tf.compat.v1.placeholder(tf.float32, ifp.shape)
img_output = tf.compat.v1.placeholder(tf.float32, ifp.shape)
ifp = nor_data(ifp)
pred, phase, absorption = phase_learn(ifp, h, px, reuse=False,
pure_phase= kwargs['pure_phase'],
conv_nb = kwargs['conv_nb'],
conv_size = kwargs['conv_size'],
dropout = kwargs['dropout'])
disc_real = discriminator(img_output)
disc_fake = discriminator(pred, reuse=True)
gen_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,
labels=tf.ones_like(disc_fake))) \
+ tf.reduce_mean(tf.abs(img_output - pred)) * kwargs['cost_rate']
# + tf.reduce_mean(tf.losses.mean_squared_error(img_output, pred)) * kwargs['cost_rate']
# + tf.reduce_mean(tf.losses.mean_squared_error(img_output, pred)) * kwargs['cost_rate']
disc_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_real,
labels=tf.ones_like(disc_real)))
disc_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=disc_fake,
labels=tf.zeros_like(disc_fake)))
disc_loss = disc_loss_real + disc_loss_fake
gen_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope='generator')
disc_vars = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator')
optimizer_gen = tf.compat.v1.train.AdamOptimizer(learning_rate=kwargs['learning_rate_g'])
optimizer_disc = tf.compat.v1.train.AdamOptimizer(learning_rate=kwargs['learning_rate_d'])
train_gen = optimizer_gen.minimize(gen_loss, var_list=gen_vars)
train_disc = optimizer_disc.minimize(disc_loss, var_list=disc_vars)
######################################################################
# plots for debug
if kwargs['iter_plot']:
fig, axs = plt.subplots(2, 3, figsize=(16, 8))
im0 = axs[0, 0].imshow(ifp.reshape(px, px), cmap='jet')
tx0 = axs[0, 0].set_title('Intensity')
fig.colorbar(im0, ax=axs[0, 0])
tx1 = axs[1, 0].set_title('Absolute error of the intensity for iteration 0')
im1 = axs[1, 0].imshow(ifp.reshape(px, px), cmap='jet')
fig.colorbar(im1, ax=axs[1, 0])
im2 = axs[0, 1].imshow(np.zeros((px, px)), cmap='gray')
fig.colorbar(im2, ax=axs[0, 1])
tx2 = axs[0, 1].set_title('Phase')
im3 = axs[0, 2].imshow(np.zeros((px, px)), cmap='jet')
fig.colorbar(im3, ax=axs[0, 2])
tx3 = axs[0, 2].set_title('Absorption')
xdata, g_loss = [], []
im4, = axs[1, 1].plot(xdata, g_loss, 'r-')
axs[1, 1].set_yscale('log')
tx4 = axs[1, 1].set_title('Generator loss')
plt.tight_layout()
#########################################################################
rec_tmp = tf.zeros([1, px, px, 1])
init = tf.compat.v1.global_variables_initializer()
saver = tf.compat.v1.train.Saver()
with tf.compat.v1.Session() as sess:
# Run the initializer
sess.run(init)
if kwargs['weights_init']:
if init_wpath == None:
print('Please provide the file name of initial weights.')
saver.restore(sess, init_wpath)
for step in range(1, kwargs['num_steps'] + 1):
# disc_y = np.concatenate([np.ones([1]), np.zeros([1])], axis=0)
# gen_y = np.ones([1])
# feed_dict = {img_input: prj, img_output: prj}
# _, _, gl, dl = sess.run([train_gen, train_disc, gen_loss, disc_loss], feed_dict=feed_dict)
with tf.device('/device:GPU:1'):
dl, _ = sess.run([disc_loss, train_disc],
feed_dict={img_input: ifp, img_output: ifp})
with tf.device('/device:GPU:2'):
gl, _ = sess.run([gen_loss, train_gen], feed_dict={img_input: ifp, img_output: ifp})
# print(gl, gl.shape, gl.dtype)
xdata.append(step)
g_loss.append(gl)
# print(np.array(g_loss).dtype, np.array(g_loss))
# ax = axs[0]
# ax.plot(gl)
if np.isnan(gl):
sess.run(init)
if step % kwargs['display_step'] == 0 or step == 1:
pred, phase, absorption = sess.run(phase_learn(ifp, h, px, reuse=False,
pure_phase= kwargs['pure_phase'],
conv_nb = kwargs['conv_nb'],
conv_size = kwargs['conv_size'],
dropout = kwargs['dropout']))
# if (np.isnan(recon.any())) or (recon.all()==0):
# sess.run(init)
###########################################################
ifp_plt = np.reshape(pred, (px, px))
ifp_plt = np.abs(ifp_plt - ifp.reshape((px, px)))
rec_plt = np.reshape(phase, (px, px))
abs_plt = np.reshape(absorption, (px, px))
tx1.set_text('Absolute error of the intensity for iteration {0}'.format(step))
vmax = np.max(ifp_plt)
vmin = np.min(ifp_plt)
im1.set_data(ifp_plt)
im1.set_clim(vmin, vmax)
im2.set_data(rec_plt)
vmax = np.max(rec_plt)
vmin = np.min(rec_plt)
im2.set_clim(vmin, vmax)
im3.set_data(abs_plt)
vmax = np.max(abs_plt)
vmin = np.min(abs_plt)
im3.set_clim(vmin, vmax)
axs[1, 1].plot(xdata, g_loss, 'r-')
plt.pause(0.1)
######################################################################
print("Step " + str(step) + ", Generator Loss= " + "{:.7f}".format(gl) +
', Discriminator loss = '+ "{:.7f}".format(dl))
if gl<kwargs['gl_tol']:
_, phase, aborption = sess.run(phase_learn(ifp, h, px, reuse=False,
pure_phase = kwargs['pure_phase'],
conv_nb = kwargs['conv_nb'],
conv_size = kwargs['conv_size'],
dropout = kwargs['dropout']))
break
if step > (kwargs['num_steps'] - 10):
_, phase, absorption = sess.run(phase_learn(ifp, h, px, reuse=False,
pure_phase = kwargs['pure_phase'],
conv_nb = kwargs['conv_nb'],
conv_size = kwargs['conv_size'],
dropout = kwargs['dropout']))
rec_tmp = tf.concat([rec_tmp, phase], axis=0)
# print(rec_tmp.shape)
plt.close(fig)
saver.save(sess, save_wpath)
# _, recon = sess.run(tomo_learn(prj, ang, px, reuse=True, conv_nb = kwargs['conv_nb'],
# conv_size = kwargs['conv_size'],
# dropout = kwargs['dropout'],
# method = kwargs['method']))
if rec_tmp.shape[0] >1:
phase = tf.reduce_mean(rec_tmp, axis=0).eval()
# print(recon.shape)
return phase, absorption
nb_epoch = 12
# number of convolutional filters to use
nb_filters = 32
# size of pooling area for max pooling
nb_pool = 2
# convolution kernel size
nb_conv = 3
fname = '../../test/test_data/1038.tiff'
ind_uncenter1 = range(1038, 1047)
ind_uncenter2 = range(1049, 1057)
uncenter1 = dxchange.read_tiff_stack(fname, ind=ind_uncenter1, digit=4)
uncenter2 = dxchange.read_tiff_stack(fname, ind=ind_uncenter2, digit=4)
uncenter = np.concatenate((uncenter1, uncenter2), axis=0)
uncenter = nor_data(uncenter)
print (uncenter.shape)
uncenter = img_window(uncenter[:, 360:1460, 440:1440], 200)
print (uncenter.shape)
uncenter_patches = extract_3d(uncenter, patch_size, 1)
np.random.shuffle(uncenter_patches)
print (uncenter_patches.shape)
# print uncenter_patches.shape
center_img = dxchange.read_tiff('../../test/test_data/1048.tiff')
center_img = nor_data(center_img)
print (center_img.shape)
center_img = img_window(center_img[360:1460, 440:1440], 400)
center_patches = extract_3d(center_img, patch_size, 1)
np.random.shuffle(center_patches)
print (center_patches.shape)
# plt.imshow(center_img, cmap='gray', interpolation= None)
def seg_predict(img, wpath, patch_size = 32, patch_step = 1,
nb_conv=32, size_conv=3,
batch_size=1000, nb_down=2, nb_gpu = 1):
"""
Function description
Parameters
----------
img : array
The images need to be segmented.
wpath: string
The path where the trained weights of the model can be read.
patch_size: int
The size of the small patches extracted from the input images. This size should be big enough to cover the
features of the segmentation object.
patch_step: int
The pixel steps between neighbour patches. Larger steps leads faster speed, but less quality. I recommend 1
unless you need quick test of the algorithm.
nb_conv: int
Number of the covolutional kernals for the first layer. This number doubles after each downsampling layer.
size_conv: int
Size of the convolutional kernals.
batch_size: int
Batch size for the training. Bigger size leads faster speed. However, it is restricted by the memory size of
the GPU. If the user got the memory error, please decrease the batch size.
nb_epoch: int
Number of the epoches for the training. It can be understand as the number of iterations during the training.
Please define this number as the actual convergence for different data.
nb_down: int
Number of the downsampling for the images in the model.
nb_gpu: int
Number of GPUs you want to use for the training.
Returns
-------
save the segmented images to the spath.
"""
patch_shape = (patch_size, patch_size)
img = np.float32(nor_data(img))
mdl = model_choose(patch_size, patch_size, nb_conv, size_conv, nb_down, nb_gpu)
# print(mdl.summary())
mdl.load_weights(wpath)
if img.ndim == 2:
ih, iw = img.shape
predict_x = extract_3d(img, patch_shape, patch_step)
predict_x = np.reshape(predict_x, (predict_x.shape[0], patch_size, patch_size, 1))
predict_y = mdl.predict(predict_x, batch_size=batch_size)
predict_y = np.reshape(predict_y, (predict_y.shape[0],patch_size, patch_size))
predict_y = reconstruct_patches(predict_y, (ih, iw), patch_step)
return predict_y
else:
pn, ih, iw = img.shape
images = np.empty(pn, dtype=object) # create empty array
for i in range(pn):
print('Processing the %s th image' % i)
tstart = time.time()
predict_x = img[i]
predict_x = extract_3d(predict_x, patch_shape, patch_step)
predict_x = np.reshape(predict_x, (len(predict_x), patch_size, patch_size, 1))
predict_y = mdl.predict(predict_x, batch_size=batch_size)
predict_y = np.reshape(predict_y, (len(predict_y), patch_size, patch_size))
predict_y = reconstruct_patches(predict_y, (ih, iw), patch_step)
images[i]=predict_y
return images
def seg_train(img_x, img_y, **kwargs):
"""
Function description.
Parameters
----------
img_x: array, 2D or 3D
Training input of the model. It is the raw image for the segmentation.
img_y: array, 2D or 3D
Training output of the model. It is the corresponding segmentation of the training input.
patch_size: int
The size of the small patches extracted from the input images. This size should be big enough to cover the
features of the segmentation object.
patch_step: int
The pixel steps between neighbour patches. Larger steps leads faster speed, but less quality. I recommend 1
unless you need quick test of the algorithm.
conv_nb: int
Number of the covolutional kernals for the first layer. This number doubles after each downsampling layer.
conv_size: int
Size of the convolutional kernals.
batch_size: int
Batch size for the training. Bigger size leads faster speed. However, it is restricted by the memory size of
the GPU. If the user got the memory error, please decrease the batch size.
nb_epoch: int
Number of the epoches for the training. It can be understand as the number of iterations during the training.
Please define this number as the actual convergence for different data.
model_layers: int
Number of the downsampling for the images in the model.
gpu_nb: int
Number of GPUs you want to use for the training.
Returns
-------
mdl
The trained CNN model for segmenation. The model can be saved for future segmentations.
"""
seg_kwargs = [
'patch_size', 'patch_step', 'conv_nb', 'conv_size', 'batch_size',
'epoch_nb', 'model_layers', 'gpu_nb'
]
kwargs_defaults = _get_seg_kwargs()
for kw in seg_kwargs:
kwargs.setdefault(kw, kwargs_defaults[kw])
patch_shape = (kwargs['patch_size'], kwargs['patch_size'])
# print img_x.shape
# print img_x.max(), img_x.min()
img_x = nor_data(img_x)
img_y = nor_data(img_y)
# print img_x.shape
# print img_x.max(), img_x.min()
train_x = extract_3d(img_x, patch_shape, kwargs['patch_step'])
train_y = extract_3d(img_y, patch_shape, kwargs['patch_step'])
# print train_x.shape
# print train_x.max(), train_x.min()
train_x = np.reshape(
train_x, (len(train_x), kwargs['patch_size'], kwargs['patch_size'], 1))
train_y = np.reshape(
train_y, (len(train_y), kwargs['patch_size'], kwargs['patch_size'], 1))
mdl = model_choose(kwargs['patch_size'], kwargs['patch_size'],
kwargs['conv_nb'], kwargs['conv_size'],
kwargs['model_layers'], kwargs['gpu_nb'])
print(mdl.summary())
mdl.fit(train_x,
train_y,
batch_size=kwargs['batch_size'],
epochs=kwargs['nb_epoch'])
return mdl
def seg_predict(img, wpath, spath, **kwargs):
"""
Function description
Parameters
----------
img : array
The images need to be segmented.
wpath: string
The path where the trained weights of the model can be read.
spath: string
The path to save the segmented images.
patch_size: int
The size of the small patches extracted from the input images. This size should be big enough to cover the
features of the segmentation object.
patch_step: int
The pixel steps between neighbour patches. Larger steps leads faster speed, but less quality. I recommend 1
unless you need quick test of the algorithm.
conv_nb: int
Number of the covolutional kernals for the first layer. This number doubles after each downsampling layer.
conv_size: int
Size of the convolutional kernals.
batch_size: int
Batch size for the training. Bigger size leads faster speed. However, it is restricted by the memory size of
the GPU. If the user got the memory error, please decrease the batch size.
nb_epoch: int
Number of the epoches for the training. It can be understand as the number of iterations during the training.
Please define this number as the actual convergence for different data.
model_layers: int
Number of the downsampling for the images in the model.
gpu_nb: int
Number of GPUs you want to use for the training.
Returns
-------
save the segmented images to the spath.
"""
seg_kwargs = [
'patch_size', 'patch_step', 'conv_nb', 'conv_size', 'batch_size',
'epoch_nb', 'model_layers', 'gpu_nb'
]
kwargs_defaults = _get_seg_kwargs()
for kw in seg_kwargs:
kwargs.setdefault(kw, kwargs_defaults[kw])
patch_shape = (kwargs['patch_size'], kwargs['patch_size'])
img = np.float32(nor_data(img))
mdl = model_choose(kwargs['patch_size'], kwargs['patch_size'],
kwargs['conv_nb'], kwargs['conv_size'],
kwargs['model_layers'], kwargs['gpu_nb'])
# print(mdl.summary())
mdl.load_weights(wpath)
if img.ndim == 2:
ih, iw = img.shape
predict_y = pred_single(mdl, img, ih, iw, patch_shape,
kwargs['patch_step'], kwargs['patch_size'],
kwargs['batch_size'])
fname = spath + 'seg'
dxchange.write_tiff(predict_y, fname, dtype='float32')
else:
pn, ih, iw = img.shape
for i in range(pn):
print('Processing the %s th image' % i)
tstart = time.time()
predict_x = img[i]
predict_y = pred_single(mdl, predict_x, ih, iw, patch_shape,
kwargs['patch_step'], kwargs['patch_size'],
kwargs['batch_size'])
predict_y = np.float32(predict_y)
fname = spath + 'seg' + "-%03d" % (i)
dxchange.write_tiff(predict_y, fname, dtype='float32')
print('The prediction runs for %s seconds' %
(time.time() - tstart))
nb_evl = 100
start_time = time.time()
fnames = glob.glob('../../test/test_data/*.tiff')
fnames = np.sort(fnames)
mdl = model(dim_img, nb_filters, nb_conv, nb_classes)
mdl.load_weights('classify_training_weights.h5')
print('The model loading time is %s seconds' % (time.time() - start_time))
start_time = time.time()
Y_score = np.zeros((len(fnames)))
for i in range(len(fnames)):
img = dxchange.read_tiff(fnames[i])
img = nor_data(img)
X_evl = np.zeros((nb_evl, dim_img, dim_img))
for j in range(nb_evl):
X_evl[j] = img_window(img[360:1460, 440:1440], dim_img)
X_evl = X_evl.reshape(X_evl.shape[0], 1, dim_img, dim_img)
Y_evl = mdl.predict(X_evl, batch_size=batch_size)
Y_score[i] = sum(np.dot(Y_evl, [0, 1]))
#print('The evaluate score is:', Y_score[i])
#Y_score = sum(np.round(Y_score))/len(Y_score)
ind_max = np.argmax(Y_score)
print('The well-centered reconstruction is:', fnames[ind_max])
print('The prediction runs for %s seconds' % (time.time() - start_time))
plt.plot(Y_score)
plt.show()