def mie_loss(self, params, image, dim):
'''Returns the residual between the image and our Mie model.'''
p = params.valuesdict()
h = LMHologram(coordinates=coordinates(dim))
h.particle.r_p = [p['x'] + dim[0] // 2, p['y'] + dim[1] // 2, p['z']]
h.particle.a_p = p['a_p']
h.particle.n_p = p['n_p']
h.instrument.wavelength = p['lamb']
h.instrument.magnification = p['mpp']
h.instrument.n_m = p['n_m']
hologram = h.hologram().reshape(dim)
return (hologram - image) / self.noise
def makedata(config={}):
'''Make Training Data'''
# set up pipeline for hologram calculation
shape = config['shape']
holo = LMHologram(coordinates=coordinates(shape))
holo.instrument.properties = config['instrument']
# create directories and filenames
directory = os.path.expanduser(config['directory'])
imgtype = config['imgtype']
nframes = config['nframes']
start = 0
tempnum = nframes
for dir in ('images', 'labels', 'params'):
path = os.path.join(directory, dir)
if not os.path.exists(path):
os.makedirs(path)
already_files = len(os.listdir(path))
if already_files < tempnum: #if there are fewer than the number of files desired
tempnum = already_files
if not config['overwrite']:
start = tempnum
if start >= nframes:
return
with open(directory + '/config.json', 'w') as f:
json.dump(config, f)
filetxtname = os.path.join(directory, 'filenames.txt')
imgname = os.path.join(directory, 'images', 'image{:04d}.' + imgtype)
jsonname = os.path.join(directory, 'params', 'image{:04d}.json')
yoloname = os.path.join(directory, 'labels', 'image{:04d}.txt')
filetxt = open(filetxtname, 'w')
for n in range(start, nframes): # for each frame ...
print(imgname.format(n))
sample = make_sample(config) # ... get params for particles
# ... calculate hologram
frame = np.random.normal(0, config['noise'], shape)
if len(sample) > 0:
holo.particle = sample
frame += holo.hologram().reshape(shape)
else:
frame += 1.
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
# ... and save the results
cv2.imwrite(imgname.format(n), frame)
with open(jsonname.format(n), 'w') as fp:
fp.write(format_json(sample, config))
with open(yoloname.format(n), 'w') as fp:
fp.write(format_yolo(sample, config))
filetxt.write(imgname.format(n) + '\n')
#print('finished image {}'.format(n+1))
return
def feature_extent(sphere, config, nfringes=20):
'''Radius of holographic feature in pixels'''
s = Sphere()
s.a_p = sphere.a_p
s.n_p = sphere.n_p
s.z_p = sphere.z_p
h = LMHologram(coordinates=np.arange(300))
h.instrument.properties = config['instrument']
h.particle = sphere
# roughly estimate radii of zero crossings
b = h.hologram() - 1.
ndx = np.where(np.diff(np.sign(b)))[0] + 1
return float(ndx[nfringes])
def feature_extent(sphere, config, nfringes=20, maxrange=300):
'''Radius of holographic feature in pixels'''
h = LMHologram(coordinates=np.arange(maxrange))
h.instrument.properties = config['instrument']
h.particle.a_p = sphere.a_p
h.particle.n_p = sphere.n_p
h.particle.z_p = sphere.z_p
# roughly estimate radii of zero crossings
b = h.hologram() - 1.
ndx = np.where(np.diff(np.sign(b)))[0] + 1
if len(ndx) <= nfringes:
return maxrange
else:
return float(ndx[nfringes])
def mtd(configfile='mtd.json'):
'''Make Training Data'''
# read configuration
with open(configfile, 'r') as f:
config = json.load(f)
# set up pipeline for hologram calculation
shape = config['shape']
holo = LMHologram(coordinates=coordinates(shape))
holo.instrument.properties = config['instrument']
# create directories and filenames
directory = os.path.expanduser(config['directory'])
imgtype = config['imgtype']
for dir in ('images_labels', 'params'):
if not os.path.exists(os.path.join(directory, dir)):
os.makedirs(os.path.join(directory, dir))
shutil.copy2(configfile, directory)
filetxtname = os.path.join(directory, 'filenames.txt')
imgname = os.path.join(
directory, 'images_labels', 'image{:04d}.' + imgtype)
jsonname = os.path.join(directory, 'params', 'image{:04d}.json')
yoloname = os.path.join(directory, 'images_labels', 'image{:04d}.txt')
filetxt = open(filetxtname, 'w')
for n in range(config['nframes']): # for each frame ...
print(imgname.format(n))
sample = make_sample(config) # ... get params for particles
# ... calculate hologram
frame = np.random.normal(0, config['noise'], shape)
if len(sample) > 0:
holo.particle = sample
frame += holo.hologram().reshape(shape)
else:
frame += 1.
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
# ... and save the results
cv2.imwrite(imgname.format(n), frame)
with open(jsonname.format(n), 'w') as fp:
fp.write(format_json(sample, config))
with open(yoloname.format(n), 'w') as fp:
fp.write(format_yolo(sample, config))
filetxt.write(imgname.format(n) + '\n')
def scale_float(s, config):
shape = config['shape']
ext = feature_extent(s, config)
#introduce 1% noise to ext
ext = np.random.normal(ext, 0.01*ext)
extsize = ext*2
shapesize = shape[0]
scale = float(extsize)/float(shapesize)
newshape = [int(extsize)]*2
holo = LMHologram(coordinates=coordinates(newshape))
holo.instrument.properties = config['instrument']
# ... calculate hologram
frame = np.random.normal(0, config['noise'], newshape)
holo.particle = s
holo.particle.x_p += (scale-1)*100.
holo.particle.y_p += (scale-1)*100.
frame += holo.hologram().reshape(newshape)
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
#reshape
frame = cv2.resize(frame, tuple(shape))
return frame, scale
def scale_int(s, config):
shape = config['shape']
ext = feature_extent(s, config)
#introduce 1% noise to ext
ext = np.random.normal(ext, 0.01*ext)
extsize = ext*2
shapesize = shape[0]
if extsize <= shapesize:
scale = 1
else:
scale = int(np.floor(extsize/shapesize) + 1)
newshape = [i * scale for i in shape]
holo = LMHologram(coordinates=coordinates(newshape))
holo.instrument.properties = config['instrument']
# ... calculate hologram
frame = np.random.normal(0, config['noise'], newshape)
holo.particle = s
holo.particle.x_p += (scale-1)*100
holo.particle.y_p += (scale-1)*100
frame += holo.hologram().reshape(newshape)
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
#decimate
frame = frame[::scale, ::scale]
return frame, scale
-edit train_config.json with appropriate params
(make sure you have available disk space)
-Run this file with python or nohup
(dataset generation + training will take at least few hours)
'''
configfile = 'keras_train_config.json'
with open(configfile, 'r') as f:
config = json.load(f)
#always only one particle per stamp
config['particle']['nspheres'] = [1, 2]
'''Make Training Data'''
# set up pipeline for hologram calculation
shape = config['shape']
holo = LMHologram(coordinates=coordinates(shape))
holo.instrument.properties = config['instrument']
imgtype = config['imgtype']
#Parameter Info
particle = config['particle']
zmin, zmax = particle['z_p']
amin, amax = particle['a_p']
nmin, nmax = particle['n_p']
def makedata(settype='train/', nframes=10):
# create directories and filenames
directory = os.path.expanduser(config['directory']) + settype
for dir in ('images', 'params'):
if not os.path.exists(os.path.join(directory, dir)):
def example():
'''
Make a "noisy" hologram. Then fit the noisy hologram. Plot the results.
'''
## Make Noisy Hologram.
# Create hologram to be fitted.
from time import time
x, y, z = 0., 0., 100.
a_p = 0.5
n_p = 1.5
n_m = 1.339
dim = [201, 201]
lamb = 0.447
mpp = 0.048
h = LMHologram(coordinates=coordinates(dim))
h.particle.r_p = [x + dim[0] // 2, y + dim[1] // 2, z]
h.particle.a_p = a_p
h.particle.n_p = n_p
h.instrument.wavelength = lamb
h.instrument.magnification = mpp
h.instrument.n_m = n_m
hologram = h.hologram().reshape(dim)
# Add noise.
std = 0.05
noise = np.random.normal(size=hologram.shape) * std
noisy_hologram = hologram + noise
# Fit the noisy hologram.
init_params = {
'x': x,
'y': y,
'z': z + 8,
'a_p': a_p - .3,
'n_p': n_p + .03,
'n_m': n_m,
'mpp': mpp,
'lamb': lamb
}
mie_fit = Mie_Fitter(init_params)
t = time()
result = mie_fit.fit(noisy_hologram)
print("Time to fit: {:.05f}".format(time() - t))
# Calculate the resulting image.
residual = result.residual.reshape(*dim)
final = hologram
# Write error report.
report_fit(result)
## Make plots.
# Plot images.
sns.set(style='white', font_scale=1.4)
plt.imshow(np.hstack([noisy_hologram, final, residual + 1]))
plt.title('Image, Fit, Residual')
plt.gray()
plt.show()
# Plot Covariance.
f, ax = plt.subplots()
#cmap = sns.diverging_palette(220, 10, as_cmap=True)
sns.set(font_scale=1.5)
plt.title('Log Covariance Matrix')
sns.heatmap(np.log(result.covar),
cmap='PuBu',
square=True,
cbar_kws={},
ax=ax)
ax.set_xticklabels(['x', 'y', 'z', r'a$_p$', r'n$_p$'])
ax.set_yticklabels([r'n$_p$', r'a$_p$', 'z', 'y', 'x'])
plt.show()
def makedata_inner(config, settype='train', nframes=None):
# create directories and filenames
directory = os.path.abspath(
os.path.expanduser(config['directory']) + settype)
if nframes is None:
nframes = config[settype]['nframes']
start = 0
tempnum = nframes
for dir in ('images', 'params'):
path = os.path.join(directory, dir)
if not os.path.exists(path):
os.makedirs(path)
already_files = len(os.listdir(path))
if already_files < tempnum: #if there are fewer than the number of files desired
tempnum = already_files
if not config['overwrite']:
start = tempnum
if start >= nframes:
return
with open(directory + '/config.json', 'w') as f:
json.dump(config, f)
filetxtname = os.path.join(directory, 'filenames.txt')
imgname = os.path.join(directory, 'images',
'image{:04d}.' + config['imgtype'])
jsonname = os.path.join(directory, 'params', 'image{:04d}.json')
filetxt = open(filetxtname, 'w')
#always only one particle per stamp
config['particle']['nspheres'] = [1, 2]
shape = config['shape']
for n in range(start, nframes): # for each frame ...
print(imgname.format(n))
sample = make_sample(config) # ... get params for particles
s = sample[0]
ext = feature_extent(s, config)
#introduce 1% noise to ext
ext = np.random.normal(ext, 0.01 * ext)
extsize = ext * 2
shapesize = shape[0]
if extsize <= shapesize:
scale = 1
else:
scale = int(np.floor(extsize / shapesize) + 1)
newshape = [i * scale for i in shape]
holo = LMHologram(coordinates=coordinates(newshape))
holo.instrument.properties = config['instrument']
# ... calculate hologram
frame = np.random.normal(0, config['noise'], newshape)
if len(sample) > 0:
holo.particle = sample[0]
holo.particle.x_p += (scale - 1) * 100
holo.particle.y_p += (scale - 1) * 100
frame += holo.hologram().reshape(newshape)
else:
frame += 1.
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
#decimate
frame = frame[::scale, ::scale]
# ... and save the results
cv2.imwrite(imgname.format(n), frame)
with open(jsonname.format(n), 'w') as fp:
fp.write(format_json(sample, config, scale))
filetxt.write(imgname.format(n) + '\n')
return
def makedata(settype='train/', nframes=10):
# create directories and filenames
directory = os.path.expanduser(config['directory'])+settype
for dir in ('images', 'params'):
if not os.path.exists(os.path.join(directory, dir)):
os.makedirs(os.path.join(directory, dir))
shutil.copy2(configfile, directory)
filetxtname = os.path.join(directory, 'filenames.txt')
imgname = os.path.join(directory, 'images', 'image{:04d}.' + imgtype)
jsonname = os.path.join(directory, 'params', 'image{:04d}.json')
filetxt = open(filetxtname, 'w')
img_list = []
scale_list = []
zlist = []
alist = []
nlist = []
for n in range(nframes): # for each frame ...
print(imgname.format(n))
sample = make_sample(config) # ... get params for particles
s = sample[0]
zlist.append(s.z_p)
alist.append(s.a_p)
nlist.append(s.n_p)
ext = feature_extent(s, config)
#introduce 1% noise to ext
ext = np.random.normal(ext, 0.01*ext)
extsize = ext*2
shapesize = shape[0]
if extsize <= shapesize:
scale = 1
else:
scale = int(np.floor(extsize/shapesize) + 1)
newshape = [i * scale for i in shape]
holo = LMHologram(coordinates=coordinates(newshape))
holo.instrument.properties = config['instrument']
# ... calculate hologram
frame = np.random.normal(0, config['noise'], newshape)
if len(sample) > 0:
holo.particle = sample[0]
holo.particle.x_p += (scale-1)*100
holo.particle.y_p += (scale-1)*100
frame += holo.hologram().reshape(newshape)
else:
frame += 1.
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
#decimate
frame = frame[::scale, ::scale]
img_list.append(frame)
scale_list.append(scale)
# ... and save the results
#do we need?
cv2.imwrite(imgname.format(n), frame)
with open(jsonname.format(n), 'w') as fp:
fp.write(format_json(sample, config, scale))
filetxt.write(imgname.format(n) + '\n')
img_list = np.array(img_list).astype('float32')
img_list *= 1./255
zlist = np.array(zlist).astype('float32')
zlist = rescale(zmin, zmax, zlist)
alist = np.array(alist).astype('float32')
alist = rescale(amin, amax, alist)
nlist = np.array(nlist).astype('float32')
nlist = rescale(nmin, nmax, nlist)
scale_list = np.array(scale_list).astype(int)
params_list = [zlist, alist, nlist]
return img_list, params_list, scale_list