def output_valstats(self,
sess,
summary_writer,
step,
batch_x,
batch_y,
name,
store_img=True):
prediction, loss, avg_psnr = sess.run(
[self.net.recons, self.net.valid_loss, self.net.valid_avg_psnr],
feed_dict={
self.net.x: batch_x,
self.net.y: batch_y,
self.net.keep_prob: 1.,
self.net.phase: False
})
self.record_summary(summary_writer, 'valid_loss', loss, step)
self.record_summary(summary_writer, 'valid_avg_psnr', avg_psnr, step)
logging.info(
"Validation Statistics, validation loss= {:.4f}, Avg PSNR= {:.4f}".
format(loss, avg_psnr))
util.save_mat(prediction, "%s/%s.mat" % (self.prediction_path, name))
if store_img:
util.save_img(prediction[0, ...],
"%s/%s_img.tif" % (self.prediction_path, name))
def test2():
"""Test save_mat"""
Ms = [matrix(1.0, (3, 2)), spmatrix(1.0, range(3), range(3))]
names = ['M1', 'M2']
util.save_mat(Ms, names, 'test')
def bcredEst(dObj,
rObj,
num_patch=16,
patch_size=40,
pad=None,
numIter=100,
step=1,
useNoise=True,
verbose=False,
is_save=False,
save_path='bcred_intermediate_results',
xtrue=None,
xinit=None):
"""
Block Coordinate Regularization by Denoising (BCRED)
### INPUT:
dObj ~ the data fidelity term, measurement/forward model
rObj ~ the regularizer term
num_patch ~ the number of blocks assigned (Patches should not overlap with each other)
patch_size ~ the spatial size of a patch (block)
pad ~ the pad size for block-wise denoising / set to 'None' if you want to use the full denoiser
numIter ~ the total number of iterations
step ~ the step-size
verbose ~ if true print info of each iteration
is_save ~ if true save the reconstruction of each iteration
save_path ~ the save path for is_save
xtrue ~ the ground truth of the image, for tracking purpose
xinit ~ the initial value of x
### OUTPUT:
x ~ reconstruction of the algorithm
outs ~ detailed information including cost, snr, step-size and time of each iteration
"""
########### HELPER FUNCTION ###########
evaluateSnr = lambda xtrue, x: 20 * np.log10(
np.linalg.norm(xtrue.flatten('F')) / np.linalg.norm(
xtrue.flatten('F') - x.flatten('F')))
########### INITIALIZATION ###########
# initialize save foler
if is_save:
abs_save_path = os.path.abspath(save_path)
if os.path.exists(save_path):
print("Removing '{:}'".format(abs_save_path))
shutil.rmtree(abs_save_path, ignore_errors=True)
# make new path
print("Allocating '{:}'".format(abs_save_path))
os.makedirs(abs_save_path)
#initialize info data
if xtrue is not None:
xtrueSet = True
snr = []
else:
xtrueSet = False
loss = []
dist = []
timer = []
# initialize variables
if xinit is not None:
pass
else:
xinit = np.zeros(dObj.sigSize, dtype=np.float32)
x = xinit
xnext = x
x_patches = util.extract_nonoverlap_patches(x, num_patch, patch_size)
xnext_patches = x_patches
# helper variable
p, pfull = rObj.init(num_patch,
patch_size + 2 * pad) # dual variable for TV
res = dObj.res(x) # compute the residual Ax-y for xinit
########### BC-RED (EPOCH) ############
for indIter in range(numIter):
# randomize order of patches
patchInd = np.random.permutation(num_patch)
# calculate full gradient (g = Sx)
gfull_data, dcost = dObj.grad(x)
gfull_robj, pfull = rObj.red(x,
step,
pfull,
useNoise=useNoise,
extend_p=None)
gfull_tot = gfull_data + gfull_robj
# calculate the loss for showing back-compatibility of PROX-TV
obj = dcost + rObj.eval(x)
# cost[indIter] = data
loss.append(obj)
dist.append(np.linalg.norm(gfull_tot.flatten('F'))**2)
if xtrueSet:
snr.append(evaluateSnr(xtrue, x))
# set up a timer
timeStart = time.time()
## Inner Loop ##
for i in range(num_patch):
# extract patch
patch_idx = patchInd[i]
cur_patch = x_patches[patch_idx, :, :]
# get gradient of data-fit for the extracted block
g_data = dObj.gradBloc(res, patch_idx)
# denoise the block with padding & get the full gradient G
if pad is None:
g_robj, p[patch_idx, ...] = rObj.red(x,
step,
p[patch_idx, ...],
useNoise=useNoise,
extend_p=None)
g_robj_patch = util.extract_padding_patches(g_robj,
patch_idx,
extend_p=0)
else:
padded_patch = util.extract_padding_patches(x,
patch_idx,
extend_p=pad)
g_robj_patch, p[patch_idx, ...] = rObj.red(padded_patch,
step,
p[patch_idx, ...],
useNoise=useNoise,
extend_p=pad)
g_tot = g_data + g_robj_patch
# update the selected block
xnext_patches[patch_idx, :, :] = cur_patch - step * g_tot
xnext = util.putback_nonoverlap_patches(xnext_patches)
# update
res = res - step * dObj.fmultPatch(g_tot, patch_idx)
x = xnext
x_patches = xnext_patches
# end of the timer
timeEnd = time.time() - timeStart
timer.append(timeEnd)
########### LOG INFO ###########
# save & print
if is_save:
util.save_mat(
xnext, abs_save_path + '/iter_{}_mat.mat'.format(indIter + 1))
util.save_img(
xnext, abs_save_path + '/iter_{}_img.tif'.format(indIter + 1))
if verbose:
if xtrueSet:
print('[bcredEst: ' + str(indIter + 1) + '/' + str(numIter) +
']' + ' [||Gx_k||^2/||Gx_0||^2: %.5e]' %
(dist[indIter] / dist[0]) + ' [snr: %.2f]' %
(snr[indIter]))
else:
print('[bcredEst: ' + str(indIter + 1) + '/' + str(numIter) +
']' + ' [||Gx_k||^2/||Gx_0||: %.5e]' %
(dist[indIter] / dist[0]))
# summarize outs
outs = {'dist': dist / dist[0], 'snr': snr, 'time': timer}
return x, outs
def redEst(dObj,
rObj,
numIter=100,
step=1,
accelerate=False,
mode='RED',
useNoise=True,
verbose=False,
is_save=False,
save_path='red_intermediate_results',
xtrue=None,
xinit=None):
"""
Regularization by Denoising (RED)
### INPUT:
dObj ~ data fidelity term, measurement/forward model
rObj ~ regularizer term
numIter ~ total number of iterations
accelerate ~ use APGM or PGM
mode ~ RED update or PROX update
useNoise. ~ true if CNN predict noise; false if CNN predict clean image
step ~ step-size
verbose ~ if true print info of each iteration
is_save ~ if true save the reconstruction of each iteration
save_path ~ the save path for is_save
xtrue ~ the ground truth of the image, for tracking purpose
xinit ~ initialization of x (zero otherwise)
### OUTPUT:
x ~ reconstruction of the algorithm
outs ~ detailed information including cost, snr, step-size and time of each iteration
"""
########### HELPER FUNCTION ###########
evaluateSnr = lambda xtrue, x: 20 * np.log10(
np.linalg.norm(xtrue.flatten('F')) / np.linalg.norm(
xtrue.flatten('F') - x.flatten('F')))
########### INITIALIZATION ###########
# initialize save foler
if is_save:
abs_save_path = os.path.abspath(save_path)
if os.path.exists(save_path):
print("Removing '{:}'".format(abs_save_path))
shutil.rmtree(abs_save_path, ignore_errors=True)
# make new path
print("Allocating '{:}'".format(abs_save_path))
os.makedirs(abs_save_path)
#initialize info data
if xtrue is not None:
xtrueSet = True
snr = []
else:
xtrueSet = False
loss = []
dist = []
timer = []
# initialize variables
if xinit is not None:
pass
else:
xinit = np.zeros(dObj.sigSize, dtype=np.float32)
x = xinit
s = x # gradient update
t = 1. # controls acceleration
p, pfull = rObj.init(1, dObj.sigSize[0]) # dual variable for TV
p = p[0]
########### BC-RED (EPOCH) ############
for indIter in range(numIter):
timeStart = time.time()
# get gradient
g, _ = dObj.grad(s)
if mode == 'RED':
g_robj, p = rObj.red(s, step, p, useNoise=useNoise, extend_p=None)
xnext = s - step * (g + g_robj)
elif mode == 'PROX':
xnext, p = rObj.prox(np.clip(s - step * g, 0, np.inf), step,
p) # clip to [0, inf]
elif mode == 'GRAD':
xnext = s - step * g
else:
print("No such mode option")
exit()
timeEnd = time.time() - timeStart
########### LOG INFO ###########
# calculate full gradient for convergence plot
gfull, dfull = dObj.grad(x)
if mode == 'RED':
g_robj, pfull = rObj.red(x,
step,
pfull,
useNoise=useNoise,
extend_p=None)
Px = x - step * (gfull + g_robj)
# Gx
diff = np.linalg.norm(gfull.flatten('F') + g_robj.flatten('F'))**2
obj = dfull + rObj.eval(x)
elif mode == 'PROX':
Px, pfull = rObj.prox(np.clip(x - step * gfull, 0, np.inf), step,
pfull)
# x-Px
diff = np.linalg.norm(x.flatten('F') - Px.flatten('F'))**2
obj = dfull + rObj.eval(x)
elif mode == 'GRAD':
# x-Px
Px = x - step * g
diff = np.linalg.norm(x.flatten('F') - Px.flatten('F'))**2
obj = dfull
else:
print("No such mode option")
exit()
# acceleration
if accelerate:
tnext = 0.5 * (1 + np.sqrt(1 + 4 * t * t))
else:
tnext = 1
s = xnext + ((t - 1) / tnext) * (xnext - x)
# output info
# cost[indIter] = data
loss.append(obj)
dist.append(diff)
timer.append(timeEnd)
# evaluateTol(x, xnext)
if xtrueSet:
snr.append(evaluateSnr(xtrue, x))
# update
t = tnext
x = xnext
# save & print
if is_save:
util.save_mat(
xnext, abs_save_path + '/iter_{}_mat.mat'.format(indIter + 1))
util.save_img(
xnext, abs_save_path + '/iter_{}_img.tif'.format(indIter + 1))
if verbose:
if xtrueSet:
print('[redEst: ' + str(indIter + 1) + '/' + str(numIter) +
']' + ' [||Gx_k||^2/||Gx_0||^2: %.5e]' %
(dist[indIter] / dist[0]) + ' [snr: %.2f]' %
(snr[indIter]))
else:
print('[redEst: ' + str(indIter + 1) + '/' + str(numIter) +
']' + ' [||Gx_k||^2/||Gx_0||^2: %.5e]' %
(dist[indIter] / dist[0]))
# summarize outs
outs = {'dist': dist / dist[0], 'snr': snr, 'time': timer}
return x, outs
def test2():
"""Test save_mat"""
Ms = [matrix(1.0, (3,2)), spmatrix(1.0, range(3), range(3))]
names = ['M1', 'M2']
util.save_mat(Ms, names, 'test')
def train(self,
data_provider,
output_path,
valid_provider,
valid_size,
training_iters=100,
epochs=1000,
dropout=0.75,
display_step=1,
save_epoch=50,
restore=False,
write_graph=False,
prediction_path='validation'):
"""
Lauches the training process
:param data_provider: callable returning training and verification data
:param output_path: path where to store checkpoints
:param valid_provider: data provider for the validation dataset
:param valid_size: batch size for validation provider
:param training_iters: number of training mini batch iteration
:param epochs: number of epochs
:param dropout: dropout probability
:param display_step: number of steps till outputting stats
:param restore: Flag if previous model should be restored
:param write_graph: Flag if the computation graph should be written as protobuf file to the output path
:param prediction_path: path where to save predictions on each epoch
"""
# initialize the training process.
init = self._initialize(training_iters, output_path, restore,
prediction_path)
# create output path
directory = os.path.join(output_path, "final/")
if not os.path.exists(directory):
os.makedirs(directory)
save_path = os.path.join(directory, "model.cpkt")
if epochs == 0:
return save_path
with tf.Session() as sess:
if write_graph:
tf.train.write_graph(sess.graph_def, output_path, "graph.pb",
False)
sess.run(init)
if restore:
ckpt = tf.train.get_checkpoint_state(output_path)
if ckpt and ckpt.model_checkpoint_path:
self.net.restore(sess, ckpt.model_checkpoint_path)
summary_writer = tf.summary.FileWriter(output_path,
graph=sess.graph)
logging.info("Start optimization")
# select validation dataset
valid_x, valid_y = valid_provider(valid_size, fix=True)
util.save_mat(valid_y,
"%s/%s.mat" % (self.prediction_path, 'origin_y'))
util.save_mat(valid_x,
"%s/%s.mat" % (self.prediction_path, 'origin_x'))
for epoch in range(epochs):
total_loss = 0
# batch_x, batch_y = data_provider(self.batch_size)
for step in range((epoch * training_iters),
((epoch + 1) * training_iters)):
batch_x, batch_y = data_provider(self.batch_size)
# Run optimization op (backprop)
_, loss, lr, avg_psnr = sess.run(
[
self.optimizer, self.net.loss,
self.learning_rate_node, self.net.avg_psnr
],
feed_dict={
self.net.x: batch_x,
self.net.y: batch_y,
self.net.keep_prob: dropout,
self.net.phase: True
})
if step % display_step == 0:
logging.info(
"Iter {:} (before training on the batch) Minibatch loss= {:.4f}, Minibatch abs err= {:.4f}"
.format(step, loss, avg_psnr))
self.output_minibatch_stats(sess, summary_writer, step,
batch_x, batch_y)
total_loss += loss
self.record_summary(summary_writer, 'training_loss', loss,
step)
self.record_summary(summary_writer, 'training_avg_psnr',
avg_psnr, step)
# output statistics for epoch
self.output_epoch_stats(epoch, total_loss, training_iters, lr)
self.output_valstats(sess,
summary_writer,
step,
valid_x,
valid_y,
"epoch_%s" % epoch,
store_img=True)
if epoch % save_epoch == 0:
directory = os.path.join(output_path,
"{}_cpkt/".format(step))
if not os.path.exists(directory):
os.makedirs(directory)
path = os.path.join(directory, "model.cpkt".format(step))
self.net.save(sess, path)
save_path = self.net.save(sess, save_path)
logging.info("Optimization Finished!")
return save_path
def apgmEst(dObj,
rObj,
numIter=100,
step=100,
accelerate=True,
stochastic=False,
mini_batch=None,
verbose=False,
is_save=True,
save_path='result',
xtrue=None):
"""
Plug-and-Play APGM with switch for PGM and SPGM
### INPUT:
dObj ~ data fidelity term, measurement/forward model
rObj ~ regularizer term
numIter ~ total number of iterations
accelerate ~ use APGM or PGM
stochastic ~ use SPGM or not
mini_batch ~ the size of the mini_batch
step ~ step-size
verbose ~ if true print info of each iteration
is_save ~ if true save the reconstruction of each iteration
save_path ~ the save path for is_save
xtrue ~ the ground truth of the image, for tracking purpose
### OUTPUT:
x ~ reconstruction of the algorithm
outs ~ detailed information including cost, snr, step-size and time of each iteration
"""
##### HELPER FUNCTION #####
evaluateSnr = lambda xtrue, x: 20 * np.log10(
np.linalg.norm(xtrue.flatten('F')) / np.linalg.norm(
xtrue.flatten('F') - x.flatten('F')))
evaluateTol = lambda x, xnext: np.linalg.norm(
x.flatten('F') - xnext.flatten('F')) / np.linalg.norm(x.flatten('F'))
##### INITIALIZATION #####
# initialize save foler
if save_path:
abs_save_path = os.path.abspath(save_path)
if os.path.exists(save_path):
print("Removing '{:}'".format(abs_save_path))
shutil.rmtree(abs_save_path, ignore_errors=True)
# make new path
print("Allocating '{:}'".format(abs_save_path))
os.makedirs(abs_save_path)
# initialize measurement mask
if stochastic:
totNum = dObj.y.shape[0]
idx = 0
keepIdx = np.zeros(mini_batch, dtype=np.int32)
for j in range(mini_batch):
keepIdx[j] = idx
idx = idx + int(totNum / mini_batch)
#initialize info data
if xtrue is not None:
xtrueSet = True
snr = []
else:
xtrueSet = False
dist = []
timer = []
relativeChange = []
# initialize variables
xinit = np.zeros(dObj.sigSize, dtype=np.float32)
# outs = struct(xtrueSet)
x = xinit
s = x # gradient update
t = 1. # controls acceleration
p = rObj.init() # dual variable for TV
pfull = rObj.init() # dual variable for TV
##### MAIN LOOP #####
for indIter in range(numIter):
timeStart = time.time()
if stochastic:
# get gradient
g, _, keepIdx = dObj.gradStoc(s, keepIdx)
# denoise
xnext, p = rObj.prox(np.clip(s - step * g, 0, np.inf), step,
p) # clip to [0, inf]
else:
# get gradient
g, _ = dObj.grad(s)
# denoise
xnext, p = rObj.prox(np.clip(s - step * g, 0, np.inf), step,
p) # clip to [0, inf]
# calculate full gradient
if stochastic:
gfull, _ = dObj.grad(s)
Px, pfull = rObj.prox(np.clip(s - step * gfull, 0, np.inf), step,
pfull)
else:
Px = xnext
# if indIter == 0:
# outs.dist0 = np.linalg.norm(x.flatten('F') - Px.flatten('F'))^2
# acceleration
if accelerate:
tnext = 0.5 * (1 + np.sqrt(1 + 4 * t * t))
else:
tnext = 1
s = xnext + ((t - 1) / tnext) * (xnext - x)
# output info
# cost[indIter] = data
dist.append(np.linalg.norm(x.flatten('F') - Px.flatten('F'))**2)
timer.append(time.time() - timeStart)
if indIter == 0:
relativeChange.append(np.inf)
else:
relativeChange.append(evaluateTol(x, xnext))
# evaluateTol(x, xnext)
if xtrueSet:
snr.append(evaluateSnr(xtrue, x))
# update
t = tnext
x = xnext
# save & print
if is_save:
util.save_mat(
xnext, abs_save_path + '/iter_{}_mat.mat'.format(indIter + 1))
util.save_img(
xnext, abs_save_path + '/iter_{}_img.tif'.format(indIter + 1))
if verbose:
if xtrueSet:
print('[fistaEst: ' + str(indIter + 1) + '/' + str(numIter) +
']' + '[tols: %.5e]' % (relativeChange[indIter]) +
'[||x-Px||^2: %.5e]' % (dist[indIter]) + '[step: %.1e]' %
(step) + '[time: %.1f]' % (np.sum(timer)) +
'[snr: %.2f]' % (snr[indIter]))
else:
print('[fistaEst: ' + str(indIter + 1) + '/' + str(numIter) +
']' + '[tols: %.5e]' % (relativeChange[indIter]) +
'[||x-Px||^2: %.5e]' % (dist[indIter]) + '[step: %.1e]' %
(step) + '[time: %.1f]' % (np.sum(timer)))
# summarize outs
#time.sleep(0.01)
outs = {
# 'cost': cost,
'dist': dist,
'time': timer,
'relativeChange': relativeChange,
'snr': snr
}
return x, outs
def uniform_block_process(x_blocks, dist, timer, snr, global_count, # share_able variables
cpu_idx, gpu_idx, data_obj, rglr_obj, # cpu, gpu & objects
step, is_noise, pad, minibatch_size, max_global_iter, # algorithmic parameters
logging, verbose, is_save, save_every, # logging parameters
is_conv, conv_every, save_path, xtrue) -> None:
#### CPU Core Assigment ####
process = mp.current_process()
os.system("taskset -p -c %d %d" % (cpu_idx, process.pid))
#### Random Seed ####
np.random.seed(cpu_idx)
##### Main Loop ####
# initialize some variables
rglr_obj.init(gpu_idx=gpu_idx)
while global_count.value < max_global_iter:
# select block randomly at uniform
block_idx = np.random.randint(data_obj.num_blocks)
# 1st read
xtilde_blocks = read_x_list(x_blocks)
xtilde_block = xtilde_blocks[block_idx,:,:]
# get the block gradient of data-fit
g_data_block = data_obj.gradStoc_block(xtilde_block, block_idx, minibatch_size)
if pad is 'full':
xtilde = util.putback_nonoverlap_patches(xtilde_blocks)
g_rglr = rglr_obj.red(xtilde, is_noise=is_noise, extend_p=0)
g_rglr_block = util.extract_nonoverlap_patches(g_rglr,
data_obj.num_blocks, data_obj.block_size)[block_idx,:,:]
else:
g_rglr_block = rglr_obj.red(xtilde_block, is_noise=is_noise, extend_p=pad) # pad removed
# compute the overall gradient g_tot
g_tot_block = g_data_block + g_rglr_block
# update the selected block & update global memory
# upload new x to the global memory
x_blocks[block_idx] = x_blocks[block_idx]-step*g_tot_block
if logging:
xlog = util.putback_nonoverlap_patches(read_x_list(x_blocks))
with global_count.get_lock():
local_count = np.copy(global_count.value) # record local count for logging
global_count.value += 1 # update global count
# record final finishing time
if local_count < max_global_iter:
timer[local_count] = time.time() - timer[local_count]
#### Log Info ####
if logging and local_count < max_global_iter:
if is_conv and (local_count+1) % conv_every == 0:
# calculate full gradient (g = Sx)
g_full_data = data_obj.grad(xtilde_blocks)
g_full_rglr = rglr_obj.red(xtilde_blocks, is_noise=is_noise)
g_full_tot = g_full_data + g_full_rglr
dist[local_count] = np.linalg.norm(g_full_tot.flatten('F')) ** 2
if snr is not None:
snr[local_count] = evaluateSnr(xtrue, xlog)
# save & print
if is_save and (local_count+1) % save_every == 0:
util.save_mat(xlog, save_path + '/iter_{}_mat.mat'.format(local_count + 1))
util.save_img(xlog, save_path + '/iter_{}_img.tif'.format(local_count + 1))
np.set_printoptions(precision=3)
if verbose and snr is not None:
print(
f"[uniform_block_async: {local_count + 1}/{max_global_iter}] [Process: {process.name} {process.pid}] "
f"[Step-size: {step:.{4}}] [||Gx_k||^2: {dist[local_count]:.{4}}] [SNR: {snr[local_count]:.{4}}] "
f"[Time: {timer[local_count]:.{4}}]", flush=True)
elif verbose:
print(
f"[uniform_block_async: {local_count + 1}/{max_global_iter}] [Process: {process.name} {process.pid}] "
f"[Step-size: {step:.{4}}] [||Gx_k||^2: {dist[local_count]:.{4}}] "
f"[Timer: {timer[local_count]:.{4}}]", flush=True)
