deeptrack 0.5.5

A deep learning oriented microscopy image simulation package

DeepTrack 2.0

DeepTrack is a comprehensive deep learning framework for digital microscopy. We provide tools to create physical simulations of customizable optical systems, to generate and train neural network models, and to analyze experimental data.

Getting started

Installation

DeepTrack 2.0 requires at least python 3.6

To install DeepTrack 2.0, open a terminal or command prompt and run

pip install deeptrack

Basics

This section will explain core concepts about programming with DeepTrack 2.0. The section Tutorials introduces some notebooks which more directly solve problems with DeepTrack 2. Feel free to skip directly to those if you are not interested in customizing the examples.

For the following subsections we assume the following imports:

import deeptrack as dt
import numpy as np

1. Features

Features are the core computing blocks of DeepTrack 2, and operate mainly on numpy arrays. They have two special methods with which you can interact with them: resolve and update. For now, we only need to focus on resolve, which executes the feature. The first argument is the input of the feature, which does not always have to be specified. For example:

add_one = dt.Add(1)
add_one.resolve(10)
>>> 11

Here we create a feature that adds 1 to the input. We resolve that feature on the input 10, which returns 11. Simple!

2. Chaining features

Features can be chained using the + operator. If two features are chained, the output of the first is passed as the input of the second.

add_one = dt.Add(1)
multiply_two = dt.Multiply(2)
add_and_multiply = add_one + multiply_two

add_one.resolve(10)
>>> 22

Here we have two features, add_one, which adds one to the input, and multiply_two which multiplies the input with two. We chain these features, which creates a single feature which first adds 1 to the input, and then multiplies it with two. Resolving this feature on the input 10 gives (10 + 2) * 2 = 22.

3. Executing a feature multiple times

You will often find that you want to execute a single feature several times in a row. This is achieved using the ** operator. The integer on the right determines how many times to execute the feature on the left.

add_one = dt.Add(1)
add_and_multiply = add_one ** 5

add_one.resolve(10)
>>> 15

Here, the add_one feature is executed five times. Resolving this feature on the input 10 gives 10 + 1 + 1 + 1 + 1 + 1 = 15

4. Using functions as arguments

The magic of DeepTrack is that any argument used to create feature can be replaced with a function that returns a value to use. The easiet example is a function that always returns the same value:

add_one = dt.Add(lambda: 1)
add_one.resolve(10)
>>> 11

However, this is not very exciting. let's do something more interesting:

add_zero_or_one = dt.Add(lambda: np.random.randint(2)
add_zero_or_one.resolve(10)
>>> 10

Here, the value we add is randomly either 0 or 1. This time, it added zero. We can check this explicitly by running

add_zero_or_one.value.current_value
>>> 0

The first input of the Add feature is value, which we access and request its current value. Other features may have different names for their arguments, and all arguments are named arguments.

To request a new value, we call update. Until update is called, it will always use the same value.

add_zero_or_one.update()
add_zero_or_one.value.current_value
>>> 1

5. Dummy arguments

Here, we will introduce a concept that might seem useless at the time, but will make more sense in the next subsection: dummy arguments, or dummy properties. A dummy property is a argument passed to a feature that is not directly used by the feature. These can have any value and any name (that is not any of the feature's input arguments).

add_one = dt.Add(
    value=1,
    useless_argument="I do nothing"
)
add_one.resolve(10)
>>> 11

Here useless_argument is a dummy property, as it is not directly used by the Add feature

6. Dependent arguments

Now, we will show the use of dummy properties: ordinary arguments can depend on them! This means that arguments that are functions can take them as input:

add_random_integer = dt.Add(
   min_value=lambda: np.random.randint(100),
   max_value=lambda min_value: np.random.randint(min_value + 1, 101),
   value=lambda min_value, max_value: np.random.randint(min_value, max_value + 1)
)

add_random_integer.update().resolve(10)
>>> 76

Let's break this example down. We define min_value, which is a random integer between 0 and 99, which defines the minimum value to add. We also define max_value which takes min_value as an input, and returns a random integer between min_value + 1 and 100. Finally, we define value, which is the argument used by the Add feature to determine the value to add. It takes min_value and max_value as inputs, and returns a random integer between min_value and max_value.

7. Dependencies between features

A feature can dependent on the arguments of another feature. The syntax for this is simple:

add_one_or_two = dt.Add(value=np.random.randint(1, 3))
undo_add = dt.Subtract(
    value=add_one_or_two.value
)

do_nothing = add_one_or_two + undo_add
do_nothing.update().resolve(10)
>>> 10

These two arguments will now always be the same. You can of course accept it as a dummy property:

add_one_or_two = dt.Add(value=np.random.randint(1, 3))
undo_add = dt.Subtract(
    value_added=add_one_or_two.value,
    value=lambda value_added: value_added
)

It is also possible to inherit all the arguments of another feature by calling

add_one_or_two = dt.Add(value=np.random.randint(1, 3))
undo_add = dt.Subtract(
    **add_one_or_two.properties
)

Tutorials

The folder 'tutorials' contains notebooks with common applications. These may serve as a useful starting point from which to build a solution. The notebooks can be read in any order, but we provide a suggested order to introduce new concepts more naturally. This order is as follows:

1. deeptrack_introduction_tutorial gives an overview of how to use DeepTrack 2.0.
2. tracking_particle_cnn_tutorial demonstrates how to track a point particle with a convolutional neural network (CNN).
3. tracking_multiple_particles_unet_tutorial demonstrates how to track multiple particles using a U-net.
4. characterizing_aberrations_tutorial demonstrates how to add and characterize aberrations of an optical device.
5. distinguishing_particles_in_brightfield_tutorial demonstrates how to use a U-net to track and distinguish particles of different sizes in brightfield microscopy.
6. analyzing_video_tutorial demonstrates how to create videos and how to train a neural network to analyze them.

Examples

The examples folder contains notebooks which explains the different modules in more detail. Also these can be read in any order, but we provide a recommended order where more fundamental topics are introduced early. This order is as follows:

1. features_example
2. properties_example
3. scatterers_example
4. optics_example
5. aberrations_example
6. noises_example
7. augmentations_example
8. image_example
9. generators_example
10. models_example
11. losses_example
12. utils_example
13. sequences_example
14. math_example

Documentation

The detailed documentation of DeepTrack 2.0 is available at the following link: https://deeptrack-20.readthedocs.io/en/latest/

Funding

This work was supported by the ERC Starting Grant ComplexSwimmers (Grant No. 677511).