def get_exp_image_generator():
from numpy.random import randint, uniform, normal, choice
from math import pi
import deeptrack
image_parameters_function = lambda: deeptrack.get_image_parameters(
particle_center_x_list=lambda: normal(0, 5, 1),
particle_center_y_list=lambda: normal(0, 5, 1),
particle_radius_list=lambda: uniform(2, 5, 1), #uniform(1.5, 3, 1)
particle_bessel_orders_list=lambda: [
[
randint(1, 2),
],
],
particle_intensities_list=lambda: [
[
-1 * uniform(.05, .4, 1),
],
], #[[choice([-1, 1]) * uniform(.2, .6, 1), ], ],
image_half_size=lambda: 25,
image_background_level=lambda: uniform(.15, .8), #uniform(.2, .8)
signal_to_noise_ratio=lambda: uniform(20, 100), #uniform(10, 100)
gradient_intensity=lambda: uniform(0, 0.2),
gradient_direction=lambda: uniform(-pi, pi),
ellipsoidal_orientation=lambda: uniform(-pi, pi, 1),
ellipticity=lambda: uniform(0.5, 1.5, 1)) # Changed from 1
### Define image generator
image_generator = lambda: deeptrack.get_image_generator(
image_parameters_function)
return image_generator
def get_image_generator_movies():
from numpy.random import randint, uniform, normal, choice
from math import pi
import deeptrack
image_parameters_function = lambda: deeptrack.get_image_parameters(
particle_center_x_list=lambda: [
normal(0, 2, 1),
],
particle_center_y_list=lambda: [
normal(0, 2, 1),
],
particle_radius_list=lambda: uniform(2, 3, 1),
particle_bessel_orders_list=lambda: [
[1, 2],
],
particle_intensities_list=lambda: [
[uniform(.7, .9, 1), -uniform(.2, .3, 1)],
],
image_half_size=lambda: 25,
image_background_level=lambda: uniform(.2, .5),
signal_to_noise_ratio=lambda: uniform(5, 100),
gradient_intensity=lambda: uniform(0, .8),
gradient_direction=lambda: uniform(-pi, pi))
### Define image generator
image_generator = lambda: deeptrack.get_image_generator(
image_parameters_function)
return image_generator
def predictions_on_noise_levels(
network,
SNT_levels,
nbr_images_to_evaluate=1000,
particle_center_x_list=lambda: normal(0, 1, 1),
particle_center_y_list=lambda: normal(0, 1, 1),
particle_radius_list=lambda: uniform(1.5, 3, 1),
particle_bessel_orders_list=lambda: [
[
randint(1, 3),
],
],
particle_intensities_list=lambda: [
[
choice([-1, 1]) * uniform(.2, .6, 1),
],
],
image_half_size=lambda: 25,
image_background_level=lambda: uniform(.2, .8),
signal_to_noise_ratio=lambda: uniform(10, 100),
gradient_intensity=lambda: uniform(0, 1),
gradient_direction=lambda: uniform(-pi, pi)):
"""
Function for making predictions on different noise levels
Inputs: network to be evaluated, noise levels to evaluate on and parameters of the image generator(other than noise level)
Number of images to evaluate on
Outputs: predicted positions and target positions.
"""
image_shape = network.get_layer(
index=0).get_config()['batch_input_shape'][1:]
all_predictions = []
all_targets = []
for SNT_level in SNT_levels:
### Create an image generator
image_parameters_function = lambda: deeptrack.get_image_parameters(
particle_center_x_list=particle_center_x_list,
particle_center_y_list=particle_center_y_list,
particle_radius_list=particle_radius_list,
particle_bessel_orders_list=particle_bessel_orders_list,
particle_intensities_list=particle_intensities_list,
image_half_size=image_half_size,
image_background_level=image_background_level,
signal_to_noise_ratio=lambda: SNT_level,
gradient_intensity=gradient_intensity,
gradient_direction=gradient_direction)
image_generator = lambda: deeptrack.get_image_generator(
image_parameters_function)
### Genereate images for evaluation
images, targets = deeptrack.get_images_and_targets(
image_generator, nbr_images_to_evaluate, image_shape=image_shape)
all_targets.append(targets)
predictions = network.predict(images)
all_predictions.append(predictions)
return all_predictions, all_targets
def generate_images_and_targets_from_noise_levels(
network,
SNT_levels,
nbr_images_to_evaluate=1000,
translation_distance=1,
particle_radius_list=lambda: uniform(1.5, 3, 1),
particle_bessel_orders_list=lambda: [
[
randint(1, 3),
],
],
particle_intensities_list=lambda: [
[
choice([-1, 1]) * uniform(.2, .6, 1),
],
],
image_half_size=lambda: 25,
image_background_level=lambda: uniform(.2, .8),
signal_to_noise_ratio=lambda: uniform(10, 100),
gradient_intensity=lambda: uniform(0, 1),
gradient_direction=lambda: uniform(-pi, pi)):
"""
Generates images of different noise levels and corresponding targets
"""
translation_distance = 1
particle_center_x_list = lambda: normal(0, translation_distance,
translation_distance)
particle_center_y_list = lambda: normal(0, translation_distance,
translation_distance)
image_shape = network.get_layer(
index=0).get_config()['batch_input_shape'][1:]
#all_predictions = []
all_targets = []
all_images = []
for SNT_level in SNT_levels:
### Create an image generator
image_parameters_function = lambda: deeptrack.get_image_parameters(
particle_center_x_list=particle_center_x_list,
particle_center_y_list=particle_center_y_list,
particle_radius_list=particle_radius_list,
particle_bessel_orders_list=particle_bessel_orders_list,
particle_intensities_list=particle_intensities_list,
image_half_size=image_half_size,
image_background_level=image_background_level,
signal_to_noise_ratio=lambda: SNT_level,
gradient_intensity=gradient_intensity,
gradient_direction=gradient_direction)
image_generator = lambda: deeptrack.get_image_generator(
image_parameters_function)
### Genereate images for evaluation
images, targets = deeptrack.get_images_and_targets(
image_generator, nbr_images_to_evaluate, image_shape=image_shape)
all_targets.append(targets)
all_images.append(images)
return all_targets, all_images
def evaluate_noise_levels(
network,
SNT_levels,
nbr_images_to_evaluate=1000,
particle_center_x_list = lambda : normal(0, 1, 1),
particle_center_y_list=lambda : normal(0, 1, 1),
particle_radius_list=lambda : uniform(1.5, 3, 1),
particle_bessel_orders_list=lambda : [[randint(1, 3),], ],
particle_intensities_list=lambda : [[choice([-1, 1]) * uniform(.2, .6, 1), ], ],
image_half_size=lambda : 25,
image_background_level=lambda : uniform(.2, .8),
signal_to_noise_ratio=lambda : uniform(10, 100),
gradient_intensity=lambda : uniform(0, 1),
gradient_direction=lambda : uniform(-pi, pi)):
"""
Function for evaluating performance of network for different noise levels
Inputs: network to be evaluated, noise levels to evaluate on and parameters of the image generator(other than noise level)
Number of images to evaluate on
Outputs: Loss,mae and mse for the various noise levels
"""
image_shape = network.get_layer(index=0).get_config()['batch_input_shape'][1:]
losses = []
MSEs = []
MAEs = []
for SNT_level in SNT_levels:
### Create an image generator
image_parameters_function = lambda : deeptrack.get_image_parameters(
particle_center_x_list=particle_center_x_list,
particle_center_y_list=particle_center_y_list,
particle_radius_list=particle_radius_list,
particle_bessel_orders_list=particle_bessel_orders_list,
particle_intensities_list=particle_intensities_list,
image_half_size=image_half_size,
image_background_level=image_background_level,
signal_to_noise_ratio= lambda: SNT_level,
gradient_intensity=gradient_intensity,
gradient_direction= gradient_direction)
image_generator = lambda : deeptrack.get_image_generator(image_parameters_function)
### Genereate images for evaluation
images, targets = deeptrack.get_images_and_targets(image_generator,nbr_images_to_evaluate,image_shape=image_shape)
loss,mse,mae = network.evaluate(images,targets)
### Fix units and append to final results
mse *= image_half_size()**2 #what happens if it is not a function? Silly error?
mae *= image_half_size()
losses.append(loss)
MSEs.append(mse)
MAEs.append(mae)
return losses,MSEs,MAEs
def get_default_image_generator_deeptrack(translation_distance=5,
SN_limits=[10, 100],
radius_limits=[1.5, 3]):
"""
Returns a default image genereator for deeptrack. Allows for some parameter tweaking.
"""
import deeptrack
from numpy.random import randint, uniform, normal, choice
from math import pi
image_parameters_function = lambda: deeptrack.get_image_parameters(
particle_center_x_list=lambda: normal(0, translation_distance,
translation_distance),
particle_center_y_list=lambda: normal(0, translation_distance,
translation_distance),
particle_radius_list=lambda: uniform(radius_limits[0], radius_limits[
1], 1),
particle_bessel_orders_list=lambda: [
[
randint(1, 3),
],
],
particle_intensities_list=lambda: [
[
choice([-1, 1]) * uniform(.2, .6, 1),
],
],
image_half_size=lambda: 25,
image_background_level=lambda: uniform(.2, .8),
signal_to_noise_ratio=lambda: uniform(SN_limits[0], SN_limits[1]),
gradient_intensity=lambda: uniform(0, 1),
gradient_direction=lambda: uniform(-pi, pi),
ellipsoidal_orientation=lambda: uniform(-pi, pi, 1),
ellipticity=lambda: 1)
### Define image generator
image_generator = lambda: deeptrack.get_image_generator(
image_parameters_function)
return image_generator
def get_vesicle_image_generator():
# Used for vesicles
import deeptrack
image_generator = lambda: deeptrack.get_image_generator(
get_image_parameters_optimized_vesicles())
return image_generator
def get_multiparticle_genertator():
# Used for multiparticle tracking
import deeptrack
image_generator = lambda: deeptrack.get_image_generator(
get_image_parameters_multi_particle())
return image_generator
image_parameters['Image Half-Size'] = image_half_size
image_parameters['Image Background Level'] = 0.2#uniform(.2, .4)
image_parameters['Signal to Noise Ratio'] = 5#10#20#uniform(2, 8)
image_parameters['Gradient Intensity'] = 0#uniform(0, 0.2)
image_parameters['Gradient Direction'] = uniform(-pi, pi)
image_parameters['Ellipsoid Orientation'] = uniform(-pi, pi, particle_number)
image_parameters['Ellipticity'] = 1
return image_parameters
image_parameters_function = lambda : get_image_parameters_optimized()
### Define image generator
image_generator = lambda : deeptrack.get_image_generator(image_parameters_function)
### Show some examples of generated images
#path = 'vesicle_models/LBL_vesicles_larger_r_rangerun_0_network_no' # WORKS!!!
path = 'vesicle_models/default_training_run_1_network_no' # ok, 8.7as threshold for large, run 0!?!?!
#path = 'vesicle_models/LBL_vesiclesrun_2_network_no'
#path = 'multi_particle/large_dist_vesicle_training_run0_network_no'
saved_network_file_name = path+'0.h5'#'vesicle_models/LBL_vesicles_larger_r_rangerun_0_network_no0.h5'
network_small = deeptrack.load(saved_network_file_name)
saved_network_file_name = path+'4.h5' # 4.h5
network_large = deeptrack.load(saved_network_file_name)
enhanced_thres = 7.1
small_thres_enhanced = 7.5 # use smaller for actual accurate predictionsmall_thres
small_thres = 7#7.3
large_thres = 6.7
