Hayato M. Yamanouchi

Tag: python

What Exactly Is YORU, a New Behavior Analysis Tool Powered by Object Detection?

In November 2024, a preprint was posted on bioRxiv describing YORU, a behavior analysis tool that takes a different approach from existing behavior analysis tools.
In this post, I’d like to explain what kind of tool it is.

The YORU preprint

The YORU documentation

What is YORU?

YORU is a deep learning-based tool for analyzing animal behavior.

YORUのサイトより引用

A number of deep learning-based tools, such as DeepLabCut and SLEAP, have already been released.
These tools use deep learning to track body parts of animals and estimate their posture (here I’ll refer to them as tracking tools).

By estimating posture, they reveal what posture an animal is in and use that for behavior analysis.

YORU, by contrast, recognizes behavior using an object detection algorithm.
In short, it does not track body parts; instead, it defines behavior directly from the appearance of the animal as it behaves.

When people tell an apple from a mandarin orange, they rely on cues such as shape and color.
Object detection excels at classifying objects, so it makes exactly this kind of appearance-based distinction.

Applying this to animal behavior analysis means classifying animal behavior from its appearance.
Within YORU, we call this a “behavior object.”

Whereas tracking tools represent behavior as points and lines, YORU analyzes behavior by enclosing it in a bounding box.

Advantages of defining behavior with object detection

The advantages of object detection-based behavior analysis include:

・Errors are less likely as the number of individuals increases

・Less sensitive to the orientation of the animal

・Fast analysis speed

・Can capture behaviors that are hard to capture by tracking (such as a mouse crouching)

and so on.

It makes it easy to analyze the behavior of multiple individuals—a weak point of tracking tools—at a considerably lower computational cost.
It can also easily analyze behaviors that are difficult to define from the positional information of body parts.

Consider, for example, the act of opening and closing a hand.
With a tracking tool, you would track the fingertips with a camera and define the hand as open or closed, but if the fingertips become hidden when the hand closes, tracking fails and the positional information can no longer be computed accurately.

With object detection, however, the model looks at the overall shape of the open and closed hand, so it can capture the state accurately even when part of the hand is hidden.

The same logic explains its robustness to larger numbers of individuals: with a tracking tool, as the number of individuals grows, you must assign each body part to the correct individual to capture each individual’s behavior accurately, whereas YORU analyzes only the shape of each individual and therefore requires no individual identification or assignment.

Because analysis is fast, real-time analysis is possible, allowing you to control equipment when a particular behavior occurs and thus automate behavioral experiments.

Disadvantages of defining behavior with object detection

Of course, there are disadvantages as well:

・Cannot capture a sequence of behaviors

・Cannot identify individuals

・Cannot capture unknown behaviors

・Cannot capture behaviors that are hard to distinguish by appearance

・Provides no detailed information about the defined behavior

These are precisely the strengths of tracking tools; rather than one approach being superior, the two complement each other’s advantages and disadvantages.

Experimenters therefore need to choose the method appropriate to their situation.

What makes YORU special?

Having described the advantages and disadvantages of the object detection-based behavior analysis that underlies YORU, what exactly makes YORU special?

Everything runs through a GUI

YORU lets you perform behavior analysis without any programming.

The same is true of tools like DeepLabCut and SLEAP, and it goes a long way toward making these analyses accessible to biologists.

Design

Unlike previous tools, YORU’s design doesn’t really feel like a research tool.

Built-in real-time analysis

Unusually for tools of this kind, YORU comes with a full GUI for real-time analysis.

Building a real-time analysis system isn’t that difficult if you can program to some extent, but otherwise the barrier is quite high.

With YORU, however, real-time analysis and even external triggering can all be done without any programming.

Adaptable to a wide range of experimental setups

The biggest hurdle in real-time analysis is figuring out how to combine the code that drives your own experimental apparatus with the analysis itself.

YORU adopts a plugin system, letting you choose the program that controls your external device.

In other words, you simply select the plugin that matches your own system.

You can also write your own plugins: by wrapping the control program for your own apparatus in a plugin, you can easily link it to YORU’s real-time analysis.

This makes for a remarkably easy-to-use system that no other tool offers.

Closing remarks

YORU has only just been released, and there are still rough edges in usability as well as gaps in the documentation.

That said, as these are addressed over time, I am very much looking forward to seeing how it comes to be used.

2024-11-17
How to Install SLEAP [Updated February 2024]
Here I’d like to describe how to install SLEAP, a machine learning-based tool for tracking animal body parts.

DeepLabCut is the most widely used tracking tool, and I have written about how to install it in the past, as have many other people.

For SLEAP, however, there are currently few articles available in Japanese.
SLEAP has many advantages over DeepLabCut, so I decided to write up the installation procedure to make it a viable option for more people.

SLEAP’s documentation, from installation to usage, is extremely well organized and comes with videos, so anyone comfortable with English should find it easy to get started.

The SLEAP paper is available here

The official SLEAP website is here

The SLEAP GitHub repository is here

Introduction

First of all, regarding the approach described in the official SLEAP installation guide,
```
mamba create -y -n sleap -c conda-forge -c nvidia -c sleap -c anaconda sleap=1.3.3
```
which sets up the environment, installs the packages, and handles everything else in a single command: I tried this on several computers, but in every case it stalled indefinitely during the environment setup and never completed successfully.

So here I introduce an alternative installation method, also suggested by the developers.

Installation

We will basically follow the installation instructions on the official SLEAP website.

Environment

・Windows 11 Pro
・Verified on an NVIDIA 3080
・SLEAP 1.3.3
・Python 3.7.12

Downloading the files

First, download the files from the SLEAP GitHub repository to your computer.

If you are using Git, run the following command.
```
git clone https://github.com/talmolab/sleap
```
If the git command is not available, either install git or download the files directly from GitHub.

Installing the GPU driver

Install the NVIDIA driver.
(Skip this step if it is already installed.)

Installing Anaconda

Next, download and install the Windows 10 64-bit version from the Anaconda website.
Click Free Download on the Anaconda site and scroll down; you should see a screen like the one below.
Install the Windows installer on the far left.

Once Anaconda is installed, you should find “Anaconda Prompt” in your Windows app list.
We will use it to run the commands below to create the virtual environment, install SLEAP, and launch it.

Creating the virtual environment and installing

Next, run the following commands to create the virtual environment and install SLEAP.
The environment will be named “sleap”.
```
cd sleap
conda env create -f environment.yml -n sleap
```
Note that the commands above work on computers with a GPU;
on computers without one, use the following commands instead.
```
cd sleap
conda env create -f environment_no_cuda.yml -n sleap
```
That completes the installation.

Verifying the installation

Activate the virtual environment and launch SLEAP with sleap-label.
```
conda activate sleap
sleap-label
```
You need to activate the virtual environment (the conda activate sleap command) every time you reopen Anaconda Prompt.

After activation, the prompt should change from (base) to (sleap).

If SLEAP launches, the installation was successful.

Checking version information for reporting in a paper

To check the version, activate the sleap environment and run the following command.
```
python -c "import sleap; sleap.versions()"
```
This will output
```
SLEAP: 1.3.3
TensorFlow: 2.7.0
Numpy: 1.21.5
Python: 3.7.12
OS: Windows-10-10.0.22621-SP0
```
showing the SLEAP version information and related details.

You can also check whether SLEAP can use the GPU with
```
python -c "import sleap; sleap.system_summary()"
```
which should produce output like the following when the GPU is available.
```
GPUs: 1/1 available
  Device: /physical_device:GPU:0
         Available: True
       Initialized: False
     Memory growth: None
```
In future posts, I hope to walk through how to actually use SLEAP.

Bonus

For those unsure which computer to buy, I’ve started offering PC purchase consultations on Coconala!
あなたの要望に合わせてパソコンを選び、提案しますパソコン選びに困っている方々へ！様々な目的に対応できます！
I often pick out computers and give advice on them, and many friends have told me I could make money doing PC consultations.
That inspired me to give it a try!

I’ll recommend a machine that fits both what you want to use it for and your budget.
In particular, I’ve chosen and used many computers for machine learning.

And if you’d like, I’m happy to advise you on what to look for the next time you buy a computer.

Machines I’ve picked out so far include lab analysis computers, everyday-use computers, computers for game streaming, computers for incoming university students, computers that can run CAD for architecture students, and simple stopgap machines.

I use both Mac and Windows, so I can talk about and recommend either one!

Please feel free to make use of it.
2024-01-21
[Updated December 2022] How to install YOLOX
What is YOLOX?

Among the many releases in the YOLO series, the newest one, YOLOX, came out in July 2021.

The YOLOX paper is available here (currently a preprint).
The GitHub repository is here.

YOLO is an object detection algorithm.
It is designed to run in real time, so it is extremely lightweight (inference alone can even run on a smartphone), it is easy to build systems around, and it has been adopted in many IoT devices.

“Object detection algorithm” may sound complicated, but an easy example to picture is a camera’s automatic face or eye focus.
A marker appears telling you “here is a face”; the same idea applies to other objects as well, and each one can be classified as to what it is.

There are many versions in the YOLO series: YOLO, YOLOv2, YOLOv3, YOLOv4, YOLOv5, and YOLOX.
It also runs not only on Python but on MATLAB as well, so you can adapt it to whichever platform you use.

Unlike the system used from YOLOv3 onward, YOLOX is based on the system of the original YOLO (the first published method), and it is said to be both faster and considerably more accurate.

Other blogs cover the details as well.

How to install YOLOX

Installation instructions are given on the GitHub page, but they are in English and do not cover using a virtual environment, so here I explain how to do it with an Anaconda virtual environment.

On both Windows and Mac, following the YOLOX manual exactly produced a string of errors.
The errors will likely keep changing with future versions, so look them up as they come up.

See below for how to use the conda command in the macOS Terminal.

Using Conda Commands on macOS

On Windows, I recommend using Anaconda Prompt.

Open Terminal (Anaconda Prompt), and a black window filled with text will appear.
To the left of the last line you should see (base).
At this point you are not inside a virtual environment.

YOLOX will run without a virtual environment, but since environments can conflict with other packages, I recommend creating one.

First, enter the following code to create the virtual environment.
```
conda create -n YoloX python=3.8
```
YoloX is the name of the virtual environment; you can use any other name.
You can also specify the Python version with python=3.8.

Then enter the virtual environment with the following code.
```
conda activate YoloX
```
If the (base) on the left has changed to (YoloX), you have succeeded.

Installing CUDA

First, if this is your first time installing machine learning libraries and you plan to train YOLO on a GPU, install CUDA.
(Training on a CPU takes an absurdly long time, so I do not recommend it. Note, however, that GPUs cannot currently be used on Apple Silicon machines.)
CUDA can be downloaded from the official site.

As for the CUDA version, I suggest checking which CUDA version PyTorch supports and then installing that version.
PyTorch official site

To install an older version, click Download now and then, on the following screen, scroll down to Resources and use Archive of Previous CUDA Releases.

The latest version may well work too, but for now, matching the version to PyTorch should let you run everything without trouble.

To check whether it is installed on Windows, open “Edit the system environment variables,” go to Advanced > Environment Variables, and
check whether there is a path beginning with CUDA_PATH.
The number after the V indicates the version.

Installing PyTorch

Before installing, update pip.
Enter the virtual environment in Anaconda Prompt and run the following command.
```
python -m pip install --upgrade pip
```
Go to the PyTorch official site and select the installation method that matches your environment.
One thing to watch out for is the Package field: here, choose pip.
Installing with conda will cause errors later on.

Copy and paste the command shown under “Run this Command” to install it in your virtual environment.

Once the installation finishes, use the following command to check that PyTorch is installed.
```
pip list
```
As long as torch appears in this list, you’re good to go!!!
Just to make sure everything works, launch python and run the following command.
```
import torch
print(torch.cuda.is_available())
```
If the output is True, the installation went through without problems and torch is ready to run on the GPU.
(On a Mac or on a computer without a GPU, you will get False, but as long as the import works you’re fine.)
If you get an error, go back and check the installation again.

Installing the packages required for YOLOX

Download (clone) the various YOLOX files from GitHub.
If you can use Git commands, you can clone it with the code below.
```
git clone git@github.com:Megvii-BaseDetection/YOLOX.git
```
Next, navigate into the cloned folder and install the required packages.

First, edit the contents of requirements.txt in the cloned YOLOX folder as follows.
```
# TODO: Update with exact module version
numpy
#torch>=1.7
opencv_python
loguru
tqdm
#torchvision
thop
ninja
tabulate

# verified versions
# pycocotools corresponds to https://github.com/ppwwyyxx/cocoapi
#pycocotools>=2.0.2
#onnx==1.8.1
onnxruntime==1.8.0
onnx-simplifier==0.3.5
```
It seems that packages such as pycocotools cannot be installed with this command.

Once you’ve done that, save requirements.txt and run the command below in the terminal (Anaconda prompt).
```
cd YOLOX
pip3 install -r requirements.txt
pip3 install cython pycocotools
```
The YOLOX folder you downloaded contains a file called “requirements.txt” that lists the required packages. The second command opens this file in read-only mode and installs the packages listed in it.
```
pip3 install cython pycocotools
```
This command installs cython and pycocotools.
It appears these need to be installed separately.

When you run this command, some of you may see the following error.
```
      building 'pycocotools._mask' extension
      error: Microsoft Visual C++ 14.0 or greater is required. Get it with "Microsoft C++ Build Tools": https://visualstudio.microsoft.com/visual-cpp-build-tools/
```
Apparently pycocotools requires Microsoft C++ Build Tools. (Even though there is no Mac version, the error still shows up on Mac.)
If you’re on Windows, it’s a good idea to install it just in case.

Install it, then try installing again with the code below.
```
pip3 install "git+https://github.com/philferriere/cocoapi.git#egg=pycocotools&subdirectory=PythonAPI"
```
If that still doesn’t work, install it with the following command.
```
conda install -c conda-forge pycocotools
```
In general we’ve been installing these packages with pip, but since this one simply won’t install that way, we fall back on the last resort of installing it with conda.

Next, download yolox_x.pth.
The download starts as soon as you click the link.

Move the downloaded file into the YOLOX folder you cloned from GitHub.
This gives you the model used for detection in the demo, placed wherever you like.

Add the following lines near the top of tools/demo.py so that the script can access the yolox folder.
```
#demo.pyの中身に書き込みましょう。
import sys
import os
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
```
Without this, you will get an error saying that the yolox package cannot be found.

Then run the following command in the terminal.
This lets you check whether the demo runs correctly.
```
python tools/demo.py image -n yolox-x -c yolox_x.pth --path assets/dog.jpg --conf 0.25 --nms 0.45 --tsize 640 --save_result --device gpu 
#gpuで動かさない場合には、gpuをcpuに書き換えてください。
```
Change –device gpu to –device cpu as appropriate for your setup.

You should now find the following image inside YOLOX/YOLOX_outputs/yolox_x/vis_res/20……..

Bonus

For those who aren’t sure which computer to buy, I’ve started offering computer purchase consultations on Coconala!
あなたの要望に合わせてパソコンを選び、提案しますパソコン選びに困っている方々へ！様々な目的に対応できます！
I often pick out computers and give advice on them, and many friends told me I could make money doing computer consultations.
Inspired by that, I decided to give it a try!

I’ll recommend a computer that suits what you plan to use it for and your budget.
In particular, I’ve chosen and used many computers for machine learning.

And if you’d like, I can also advise you on what to look for the next time you buy a computer.

Computers I’ve picked out so far include lab analysis machines, everyday-use machines, game-streaming machines, machines for incoming university students, machines for architecture students that can run CAD, and basic no-frills machines.

I use both Mac and Windows, so I can talk about and recommend either!

Please feel free to make use of it.
2022-11-15
My elif conditional branching isn’t working!! [Python]
Have you ever wanted to branch on multiple conditions with an if statement, tried using elif, and found that although no error was raised the results still weren’t right?
If so, the situation described below might be what’s happening.
I made this mistake myself, so I’m writing it down here for reference.

The problematic program

The problem arises when you write two conditions in the if and elif statements of a program that includes elif.
```
def trable(a, b):
    if a >= 10 & b >= 10:
        print("patern A")
    elif a >= 10 & b < 10:
        print("patern B")
    elif a < 10 & b >= 10:
        print("patern C")
    else:
        print("patern D")

trable(11, 11)
trable(11, 9)
trable(9, 11)
trable(9, 9)
```
When you run the program above, you might expect the cases to be sorted into patterns A, B, C, and D in order from the top, but what you actually get is the output below.
```
patern A
patern B
patern C
patern B
```
So what happens if we swap the order?
```
def trable(a, b):
    if a >= 10 & b >= 10:
        print("patern A")
    elif a < 10 & b < 10:
        print("patern B")
    elif a >= 10 & b >= 10:
        print("patern C")
    else:
        print("patern D")

trable(11, 11)
trable(9, 11)
trable(11, 9)
trable(9, 9)
```
If you swap the code for patterns B and C, expecting the output to come out as A, B, C, D, what you actually get is the output below.
```
patern A
patern D
patern D
patern D
```
elif itself is used as follows when you want to define multiple conditions.
```
if 条件式A:
  条件式Aが真(True)となった場合の処理
elif 条件式B:
  条件式Aが偽(False)で、条件式Bが真(True)となった場合の処理
else:
  条件式Aが偽(False)で、条件式Bも偽(False)となった場合の処理
```
However, once the conditional expressions in the if and elif statements contain two conditions, the problem described above is likely to occur.
As a result, you don’t get the output you intended.

The tricky part is that no error is raised, so you can’t tell whether things are working until you actually look at the results.

How to fix it

When you want to branch into multiple cases using compound conditions, avoid specifying multiple conditions in the elif statement.
```
def resolve(a, b):
    if a>= 10:
        if b >= 10:
            print("patern A")
        else:
            print("patern B")

    else:
        if b >= 10:
            print("patern C")
        else:
            print("patern D")

resolve(11, 11)
resolve(11, 9)
resolve(9, 11)
resolve(9, 9)
```
It’s more cumbersome, but doing it this way gives you exactly the output you expect.

Programming has unexpected pitfalls like this, so it’s a good lesson in carefully checking the code you write.

Addendum (August 2022)

It turns out the problem was that I hadn’t wrapped the conditions in parentheses.
```
def trable(a, b):
    if (a >= 10) & (b >= 10):
        print("patern A")
    elif (a < 10) & (b < 10):
        print("patern B")
    elif (a >= 10) & (b >= 10):
        print("patern C")
    else:
        print("patern D")
```
With this change, the branching worked correctly.
Alternatively, you can also fix it by writing and instead of &.
```
def trable(a, b):
    if a >= 10 and b >= 10:
        print("patern A")
    elif a < 10 and b < 10:
        print("patern B")
    elif a >= 10 and b >= 10:
        print("patern C")
    else:
        print("patern D")
```
& and and may seem equivalent, but & also acts as a bitwise AND, so in the original program
```
a >= 10 & b < 10
```
this apparently means a >= (10 & b) < 10, and
with a=9, b=9 the inequality becomes
9>=8<10, which evaluates to True.

Tricky stuff…
2022-08-25
DeepLabCut: A Deep Learning-Based Behavior Analysis Tool

(The image above links to the paper.)

Behavioral analysis is a crucial experimental approach in biology, and accurate quantification of behavior is especially indispensable for understanding the brain.

Traditionally, high-accuracy pose estimation has been achieved by attaching markers (such as sensors) to the subject (a person or a laboratory animal).
However, sensors get in the subject’s way, which can alter its behavior or restrict its movements.

One marker-free alternative is to fit a skeleton model, but developing such models is time-consuming and has to be done on a large scale.
There are also image-based pose estimation systems, but these generate enormous amounts of data, which again makes them very difficult for a single laboratory to run given the equipment required.

With DeepLabCut, anyone with a single computer can perform pose estimation very easily.
Using deep learning-based image recognition, you can track any body parts you choose without markers.

Because the labeling is learned by a deep network, you can analyze recorded videos without correcting them beforehand.
In other words, tracking works even if the lighting across the field of view is uneven or the image is somewhat distorted depending on the camera angle.

The page of the laboratory that developed DeepLabCut is here.
The GitHub repository is here.

From actually using DeepLabCut, I found that the computer used for analysis needs a GPU (graphics card) with at least 8 GB of memory.
In gaming PC terms, that means a mid- to high-end GPU, so comfortable analysis is not possible on an everyday laptop.

That said, even without such a high-spec machine, you can run the analysis on a virtual GPU using Google Colab.
I haven’t tried it yet, though…

In practice, I found the plotting accuracy to be very high.
However, analysis often fails when the background differs from that of the training data, so it is best to assume the exact setting you want to analyze and use training data recorded under conditions as close to it as possible.

The paper shows that plotting accuracy improves further when you label not only the few points you are interested in but the whole subject, including a rough skeleton, even for parts you will not use in the analysis.
In addition, if the results are unsatisfactory after training and analysis, you can remove the problematic frames or add new training data, allowing you to actively refine the model.

DeepLabCut itself is very user-friendly, and anyone can use it easily.
Why not give it a try in your own analyses?

I’d like to write about how to install and use DeepLabCut in future posts.

2021-03-25

Hayato M. Yamanouchi

PhD student, Kamikouchi Laboratory, Nagoya University / JSPS Research Fellow (DC1)
Systems Ethology (Neuroethology × Informatics)

© 2026 Hayato M. Yamanouchi

On this site

Elsewhere

Contact details are on the About page.