Logo der codecentric AG, einem in Deutschland führenden IT-Consulting Unternehmen

Logo der codecentric AG einem in Deutschland führenden IT-Consulting Unternehmen

Cloud-Native

Data & Machine Learning

Generative KI

IT-Security

Produktentwicklung

Weitere Leistungen...

Leistungen

Arbeitsweise

Projekte

Berlin

Bielefeld

Dortmund

Erfurt

Frankfurt am Main

Hamburg

Karlsruhe

Leipzig

München

Münster

Nürnberg

Solingen

Stuttgart

Standorte

Blog

Softwerker

Fachmedien

Webinare

Events

Wissens-Hub

Unsere Kultur

Kultur & Karriere Podcast

Deine Entwicklung

Jobs

Karriere

Über uns

Aktuelles

Kontakt

International contact

Alle Standorte

Du hast Fragen oder möchtest dich mit mir austauschen? Gern – trag einfach deine Kontaktdaten ein.

Contact Us | Sub-Navi-Kontakt oder Zero Search Hits

Kontakt aufnehmen

Marketing & Community Management

Konntest du keine Einträge zu deinem Thema finden?  

IT-Consultant / Machine Learning Engineer

Machine Learning Specialist

Informatik Aktuell

Data Science

Beyond CNNs: Wie Vision Transformers die Zukunft der Bildverarbeitung verändern

Artikel von Denis Stalz-John und Nils Uhrberg bei Informatik Aktuell

<p>I have been working as a data scientist at codecentric for several years now. Thus, my language of choice is Python and I am using it in several projects on a daily basis. Last year, I got pretty excited about the announcement of the new versions of the Apple M1 chip because it offered a much higher performance. Usually, I don’t need to run long trainings of neural networks on my laptop. But for small experiments, and of course debugging, my hope was to save a lot of time. In December last year I was privileged enough to choose a new business laptop and so I took the opportunity to get a Macbook Pro 16 with M1Pro Silicon. The whole installation started smoothly until I wanted to run my Python projects. Then I ran into …</p><h2>Problems</h2><p>Apple’s M1 chip is built on the ARM architecture in contrast to x64 chips used in prior Macbook versions. On the one hand, the advantage of this is that Apple became independent from Intel so they could design their own chip. The disadvantage, on the other hand, is that all software either needs to be emulated (with Rosetta2) or recompiled with an architecture-specific compiler for arm64 (M1) instead of x86 like before … Unfortunately, a recompilation is unlikely to work out of the box and code changes must be applied. Although Python is a scripting language, this also holds true for it because the interpreter is written in C. Furthermore, a lot of major packages like Numpy and Pandas are using C/C++ extensions for getting better performance, too. In short: With <code>pip install</code> I was not able to get a running environment with all necessary packages installed. There is the alternative way of using Miniforge which is a variant of Conda. But this approach lacks the possibility of reproducible environments and therefore was not an option.</p><h2>Goal</h2><p>In a Python project with multiple people involved it is crucial that the software environment is consistent across different platforms and systems. A way to achieve this is using a package manager like Poetry (<a href="https://python-poetry.org/" target="_blank">https://python-poetry.org/</a>  ) for Python. It stores all package dependencies and their exact versions in files tracked with Git, which makes it possible to rerun the installation everywhere generating the same environment. Besides the reproducibility of the environment, another mandatory feature is the possibility to debug code in an easy way. Everyone who develops on a regular basis knows how much time it can save to investigate the variables and behavior step by step in an interactive manner. And since this  is already possible with IDEs like PyCharm or VsCode, I didn’t want to miss out on this feature when changing  to a new architecture.</p><h2>What didn’t work</h2><p>The first thing I tried was installing Miniforge and running <code>poetry install</code>, a way of installing the specified dependencies into a virtualenv. Every time a package couldn’t be installed via pip (which Poetry is using in the background), I tried to install it with <code>conda install</code>. This soon became very complicated and due to package version restrictions I abandoned this approach. The second attempt I gave a try was running the terminal in an x64 emulation with Rosetta2. The idea was to use this to install only Python x64 packages. Unfortunately, I didn’t find out how to set the compiler correctly and for me it was intransparent what compilers and which version of Homebrew was used. Thus, not seeing any way to succeed here, I searched for a different way.</p><h2>Final approach</h2><p>The final approach I tried and which led to success was using Docker and running the required environments as containers. Each container is also emulated as x64, which makes it possible to install every package as before on prior Macbooks. All commands and steps are provided in the Git repository: <a href="https://github.com/JohnDenis/py-poetry-m1" target="_blank">https://github.com/JohnDenis/py-poetry-m1</a> . For abbreviation purposes I will only use Makefile targets for the description. They will be replaced with the corresponding command in the <strong>Makefile</strong>. You find it at the end of this post or in the Github repository. Additionally, Docker-Desktop must be installed to get everything running.</p><ol><li>As a first step, you need to run a container from your desired Python base image (here 3.8) <code>make run_raw</code></li><li>Now you are in a shell of the container and can install whatever is necessary. Additionally, the project directory is mounted in the container, which allows for changes via Poetry to be stored directly in the correct <strong>pyproject.toml</strong> and <strong>poetry.lock</strong> files. For the first installation a script is provided which is triggered by <code>make install</code>.</li><li>After the initial installation, your goal is to persist the current container as an image to use it for every run of your code. Open a second terminal without stopping the container. Run <code>make commit_raw</code> in the second terminal.</li><li>Now it’s possible to use the <strong>m1-built:latest</strong> image in combination with your favorite IDE to run and debug your Python scripts. Or you can run them from plain shell by running a container with a specific command.</li></ol><p><img width="1024" height="444" src="https://media.graphassets.com/XKBPekrCQvuyfryze0Vl" alt="Workflow of how docker images and containers are created." /></p><h2>Updating the environment</h2><p>In some projects it occurs  rarely,in others it occurs more often: the need to change packages of your environment. To save time, you don’t want to start from scratch and reinstall all packages with every change coming. This is why this solution presents a way of changing the built Docker image:</p><ol><li>Start a Docker container of <strong>m1-built:latest</strong> by running <code>make run_built</code></li><li>Interact with the Python environment as desired (e.g. <code>poetry add</code>, <code>poetry update</code>)</li><li>Open a second terminal without stopping the current container. Run make <code>commit_raw</code> in the second terminal.</li><li>Now you have a new version stored at the <strong>m1-built:latest</strong> tag.</li></ol><h2>Flexibility of the solution</h2><p>This solution does not only work for Poetry-based environments, but also for all environments where packages are installed via pip. This means, the version of your python environment can be selected in the <strong>docker-compose.yml</strong> (e.g. 3.8, 3.9, etc) and everything else can be applied via command line when connected to the container.</p><h2>Debugging with PyCharm</h2><p>With the Docker image <strong>m1-built:latest</strong> being committed, PyCharm offers a convenient way to run and debug your project scripts.</p><ol><li>Add a new interpreter (Project Settings -> Python Interpreter)</li><li>Add a <strong>Path mapping</strong> from your project root to <code>/opt/project</code></li><li>Add Run configuration for main.py file</li></ol><p>With these steps you create a run configuration which is working in the same way as normal environments.</p><p><img loading="lazy" src="https://media.graphassets.com/k2o5j4ATSWWUWGrIqYWj" alt="python on m1 chip" width="900" height="655" /><img loading="lazy" src="https://media.graphassets.com/Qzjc8dfKSbuaNXLdjHVS" alt="python on m1 chip docker interpreter" width="900" height="655" /> <img loading="lazy" src="https://media.graphassets.com/C7PWYS2cQ3SYWhxVeasK" alt="docker config" width="900" height="605" /></p><h2>Remarks on TensorFlow</h2><p>The only package where the original Pypi packages are not working is TensorFlow. The reason seems to be  that the AVX speed-up options cannot be emulated. Unfortunately, the workaround isn’t as clean as the plain solution, but currently I don’t know a different one:</p><ol><li>Follow the instructions above and create your environment in a container.</li><li>Compile or download a version of TensorFlow where the AVX instructions are deactivated.</li><li>Replace the original installations with <code>pip install /path/to/tensorflow_wheel</code>.</li><li>Go ahead and commit the container.</li></ol><h2>Outlook</h2><p>As long as not all Python packages are compatible with the Apple M1 silicon, this solution gives you a great way to run any Python environment on your Apple computer. And even if it’s possible to install all packages via pip, I would recommend sticking to this approach because it makes the software stack encapsulated and reproducible which saves a lot of time and nerves in the long run. One additional enhancement can be building the Docker image in a CI-pipeline. Thus, not every team member needs to conduct the described steps, but you can use the image from your private Pypi repository immediately.</p>

Python on an M1 chip: Running smoothly using Docker

<p>In this blog post, we want to give you an insight into the product development of <em>tactics.ai</em>. Our initial idea was a data-driven football analysis tool that applies machine learning techniques to analyze the strengths and weaknesses of opponents and predict player and team behavior. During our development process we had to shift our primary focus away from our initial idea. Here, you will read about the key learnings, our vision and the value of customer feedback.</p><h2>Matchplan</h2><p>Being football fans since our early childhood, we have followed different teams across several competitions and observed numerous changes over the years. From systemic changes such as the implementation of new technologies (most recently: VAR) to tactical trends on the field, like FC Barcelona’s “Tiki Taka” or Jürgen Klopp’s fast pace and high pressure style, football never stops changing. As experts in Data Science with focus on Machine Learning we thought about using deep-learning methods to formulate game strategies based on video of opponent tendencies and exploitable weaknesses. Furthermore, we would be able to predict strengths and weaknesses of individual players and forecast strategic and tactical orientation of the opponent. This knowledge would be substantial for both coaches and analysts.</p><p>While in basketball play-calls can be very predictable and players <a href="https://grantland.com/features/the-toronto-raptors-sportvu-cameras-nba-analytical-revolution/" target="_blank">already have implemented technological analysis</a>  into their daily game preparations, in football the data analysis and data-driven coaching has just picked up. In general, football clubs are investing more and more in digitalization, such as implementing fan portals, offering cashless payment options in stadiums or using new technologies, like GPS-tracking. Still, we found that most tools focus on the fitness performance of the individual player and do not use the required position and event data that is already collected by the DFL (Deutsche Fußball Liga). Furthermore, many clubs are still dependent on tools that require a large amount of manual input. This leads to analysts frequently working long hours and being pressed for time in between games, e.g. in times of double-fixture weeks. Having <em>tactics.ai</em> as a tool to comprehensively scout the other team in short order would provide meaningful advantages to the analysts and provide new possibilities for game preparation.<br /> This is how <em>tactics.ai</em> was born.</p><p>Our vision with <em>tactics.ai</em> is to improve the performance of football clubs by offering them in-depth game analysis and automatically draft plays and movements of their own and the opponents’ players. <em>tactics.ai</em> quantitatively analyzes and compares fine-grain movement patterns and transfers this data into valuable information. This knowledge is both quickly accessible and easy to grasp. That not only reduces the hours analysts have to spend retrieving scenes manually but also provides clubs with new competitive advantages.</p><h2>Starting XI</h2><p>The most important step at the beginning of our process was the establishment of a suitable contact network. Using a lean startup approach, where every product decision is based on customer feedback, the actual customer plays a central role. In particular, we were looking for both coaches and analysts that could tell us about their daily routines, their pain points and what features would make their work easier. If we would not be able to get in touch with either of them, the project would most certainly not be successful, because developing a product without a participating user often results in a software not meeting the requirements – therefore a software that will not be used.<br /> Another challenge was the accessibility of data. In order to provide potential customers with a demo of our application, we needed data to feed our algorithm. From our research we knew that the availability of data for the first two leagues in Germany is very good. Besides the scouting feed video of every game, the position and event data is recorded by a centralized organization and is distributed to every club of those leagues. This means we only needed to get in touch with one club.</p><p>After various attempts through company networks, friends & family and direct approaches of football clubs, we were finally able to establish professional partnerships with whom we decided to develop the product. As a result of this partnership we were legitimized to access all recorded data, besides the video footage. Namely, the position and event data streams. The position data contains the movement of all players and the ball for the whole game. Every other event, like fouls, shots on target or yellow cards are located in the event data.</p><p>Based on these two pillars – customer partnership and data accessibility – we were well equipped to move on with our product development.</p><h2>Kickoff</h2><p>To start off, we conducted several interviews with coaches and analysts to understand the behavior and needs of our users. During the interview process, we quickly learned that the features we had in mind – to predict the behavior of the opponent or automatically expose weaknesses and strengths – were already two steps ahead. We found that our users were facing other difficulties and that they needed different features than we had expected.</p><p>In the following we describe the two main challenges, one of the coach and one of the analyst.</p><p><strong>Coach:</strong> To teach football players how they should behave in specific situations it is common to show them various video snippets while the coach is giving an explanation simultaneously. To emphasize the insights some coaches are working with tactic boards where they can visualize the tactical behavior. Thus, in the case which was described to us the coach had to pause the video, go to the magnetic board which was standing a few meters away, placing the magnets symbolizing the players and ball the same way they are in the video, and finally starting to explain his points. This means a lot of time is used for unnecessary things and especially in a setting where you have to keep the concentration high to understand most of the tactical discussion.</p><p><a href="https://media.graphassets.com/ekUmEclETMOvQ6abyvUz"><img width="700" height="467" src="https://media.graphassets.com/ekUmEclETMOvQ6abyvUz" alt="Shows the view between the video footage and the tactics board" /></a> </p><p><strong>Analyst:</strong> A key in game analysis is the inspection of video footage. Talking about game preparation; an analyst usually watches a couple of games of the upcoming opponent. Either he watches the whole game gaining arbitrary insights during the process or he wants to find out specific behaviors of the opponent. Concerning the latter it is often the case that he wants to watch only the scenes where he could gain insights into the criteria he is investigating. E.g. every corner kick of one team. Right now the analyst we were talking to had a two step approach. At first he was watching the whole game, simultaneously labeling every scene regarding its content. Thus, afterwards he was able to filter the scenes regarding to his needs. The task of simultaneously labeling and watching is very time consuming which means doing that automatically would have the benefit of saving time and having this information accessible for every game.</p><p>Based on these observations, we decided to start with features that would tackle the above described challenges and thereby bring immediate value to the user. If we had decided otherwise, we would have taken high risks due to several factors.</p><ol><li>We are experts in the IT industry, not of football tactics. This means, even if we would have implemented fancy AI algorithms it is unclear whether they would have created insights for the coaches or analysts right away.</li><li>From our interviews we knew that professional analysts and coaches are constantly working long hours and have time pressure in between games in case of double-fixture weeks. Hence, if we could not reduce their daily workload it seemed unlikely that they would spend much time in helping us developing advanced AI algorithms.</li><li>Every coach has more or less his own way of analyzing games. But the commonality is to find relevant scenes and showing them to the team. Thus, with an AI focused product we might have tackled only a small subset of our target group.</li></ol><p>In summary, the user experience of our application turned out to play an important role in our product development. If we are able to provide the customer with an extraordinary user experience that is easy to grasp and fast to master, in order to optimize current workflows, we could substitute the existing tools and become their main video analyzing tool. From there, we would get even more customer feedback and could further develop our product towards our original vision.</p><h2>Halftime analysis</h2><p>Today, our tactical assistant <em>tactics.ai</em> is a tool that already solves the previous named problems of analysts and coaches. It automatically labels scenes, allows you to filter scenes for specific use cases, lets you create and manage playlists with relevant scenes, integrates a visual tactic board next to the real video footage, and thus enables the coach to directly instruct the team.</p><p>With this in mind the greatest challenge has been to take a step back from our vision and have a look at the broader picture. You don’t want to build a product which you yourself like, but a product which is liked and used by your target customer group. Thus, we implemented a customer driven development and in our opinion this is the key that determines how likely your product is to become a success or a failure.</p>

Why user-oriented development is so important – the story of tactics.ai

<p>Although everyone has an intuitive way of understanding what AI means, the term is somehow difficult to grasp in its whole complexity.<br /> One example of a definition is given by <a href="https://www.sciencedirect.com/science/article/pii/S0007681318301393" target="_blank">Kaplan, A. and M. Haenlein </a> (2019), who characterize AI as “a system’s ability to correctly interpret external data, to learn from such data, and to use those learnings to achieve specific goals and tasks through flexible adaptation.”</p><p>These days, one question arises: is there an objective way to measure the capabilities of an intelligent machine?  Or: is it even possible to draw a line between AI systems and standard computer systems?</p><h2>One of the first attempts: the Turing Test</h2><p>One of the first attempts to answer these questions was done by <a href="https://en.wikipedia.org/wiki/Alan_Turing" target="_blank">Alan Turing</a> , a famous mathematician and cryptologist during World War II.<br /> In 1950, he designed an experiment called Turing Test, where an interrogator is chatting with two other clients (see figure below).</p><p><a href="https://blog.codecentric.de/files/2019/06/turing_test-1.png"><img src="https://blog.codecentric.de/files/2019/06/turing_test-1-700x495.png" alt="Turing Test" width="700" height="495" /></a> </p><p>Client A is controlled by a computer program, whereas client B is a human. The task of the interrogator is now to determine only from the answers received and chat behavior of the clients which one is which.<br /> If he is not able to distinguish them reliably, the computer program passes the Turing Test successfully and is then considered AI.</p><p>The assumption of Alan Turing for this test was that only a machine with human-like intelligence could fool a human in this complex task.</p><h2>Since its creation …</h2><p>Since its creation, the test has become the most popular way in determining intelligent behavior in machines.<br /> Nevertheless, scientists question the model in its sufficiency to measure real intelligence and argue about the reliability of certain programs passing this test.</p><p>A discussion like this came up when Google presented the Google Assistant in 2018. Recordings were shown where the digital assistant <a href="https://www.youtube.com/watch?v=JvbHu_bVa_g" target="_blank">made phone calls</a> , for example, to schedule appointments at the salon or a restaurant for their clients.<br /> Compared to other digital assistants like Siri or Alexa, the new Google Assistant had a much greater capability of interacting and understanding difficult nuances in human speech. To everyone’s surprise, in the examples above, the Google Assistant would not even be recognized as a „false conversationalist“ by the interlocutors answering the phone.<br /> This indicates a success for the machine regarding the conditions of the Turing Test.</p><h2>So what is to criticize?</h2><p>So what is to criticize? Although the presentation by Google was very impressive, the outcome is not verified by the original Turing Test design. This is because the interlocutor of the phone call would take the role of the interrogator for the Turing Test, but they are not aware of this status at the time of the phone call. Therefore, they would not pay attention to indications for whether or not the caller might be a machine.<br /> It is unlikely for the receptionists to expect a computer program to call during their busy working time. They can easily miss or ignore things such as robotic voice patterns, anomalies in the sound, or unusual pauses in the conversation.</p><h2>In the future …</h2><p>In the future of determining AI, the Turing Test may hold great possibilities of measuring general intelligence, especially in interaction. However, these days, the scientific community focuses more on objective ways to measure AI in very specific settings, such as computer vision tasks, voice recognition and so forth.</p><p>Personally, I don’t consider the development of real, complex human-like intelligent machines possible in the upcoming years. The current techniques are just too restricted to their area of design. Alan Turing predicted machines to pass his test with certainty by the year 2000, but time proved him wrong and we are still waiting for this event to happen.</p>

How to define AI? – Using the Turing Test to measure human-like intelligence in machines

In this article, we'll explore how to use the Poetry package manager to manage the dependencies of a machine learning project that makes use of the M1 GPU for TensorFlow training. We'll cover the motivation for using Poetry in this context, and we'll provide a solution that makes use of Poetry's extras feature to manage the project's dependencies in a consistent and reproducible way.

## Motivation

There are several reasons why using Poetry to manage the dependencies of a machine learning project that uses the M1 GPU for TensorFlow training can be advantageous. First, Poetry makes it easy to manage the Python software environment in a consistent way across different platforms and systems. This is especially important in larger projects, where reproducibility and consistency are critical.
Secondly, using the M1 GPU for TensorFlow training requires the installation of different Python packages than those required for x64/x86 systems. Manually managing these dependencies can be tedious and error-prone, especially in a CI/CD pipeline where the software environment needs to be reproduced consistently.
Finally, using only Pip or Conda as a software manager may not be sufficient for projects that require the consistent reproduction of the software environment. In these cases, Poetry can provide a more robust and flexible solution.

## Solution

To manage the dependencies of a machine learning project that uses the M1 GPU for tensorflow training, we can make use of Poetry's extras feature. This allows us to specify optional dependencies for our project, which can be installed using the --extras flag when running poetry install.
First, we'll create an optional extra called m1 that includes the necessary packages for using the M1 GPU for TensorFlow training. This extra can be specified in the pyproject.toml file like this:

```toml
[tool.poetry.extras]
tensorflow-m1 =["tensorflow-macos==2.9.2", "tensorflow-metal==0.5.0"]
```

Next, we can create an optional extra for x64/x86 systems, which includes the necessary packages for that platform. This extra can be specified in the same `pyproject.toml` file like this:

```toml
[tool.poetry.extras]
tensorflow-x64 = ["tensorflow==2.9.2"]
```

Now, when installing our project on an M1 system, we can use the *--extras/ -E* flag to tell Poetry to also install the `tensorflow-m1` extra:

```bash
poetry install -E tensorflow-m1
```

Similarly, when installing our project on an x64/x86 system, we can use the *--extras/ -E*  flag to tell Poetry to also install the `tensorflow-x64` extra:

```bash
poetry install -E tensorflow-x64
```

This allows us to manage our project's dependencies in a way that is consistent across different platforms.
Finally, after Poetry install has finished installing our project's dependencies, we can use Conda to install any additional packages that require compiled binaries. This allows us to manage our project's dependencies in a way that is both consistent and flexible.
 
## Recommendation

If you're considering using the M1 GPU for TensorFlow training, we highly recommend making use of it. The M1 GPU is a lot faster in calculations with TensorFlow than a CPU, which can significantly accelerate the training process. By using the M1 GPU, you can potentially reduce the time it takes to train your machine learning models by an order of magnitude or more, making it an extremely valuable resource for any machine learning project.

In summary, the M1 GPU is a powerful hardware for TensorFlow training that can greatly accelerate the process of training machine learning models. By making use of Poetry and its extras feature, you can manage the dependencies of your project in a consistent and reproducible way, ensuring that your machine learning project is both reliable and efficient.
 
You can find an example in
<a href="https://github.com/dstalzjohn/py-poetry-m1" target="_blank"> this repository</a>.

How to combine Poetry, TensorFlow, and the power of the Apple M1 GPU

Lass uns austauschen!

Möchtest du dich über dieses Thema austauschen?

Auf unserem Blog findest du viele Artikel zu Tech-Themen von unseren Mitarbeitenden. Vielleicht ist auch hier ein Thema für dich dabei. Schau gerne mal rein!

Hier gibt's noch mehr spannende Infos von uns für euch

Vielen Dank für deine Anfrage. Ich melde mich bei dir.

Vielen Dank für deine Anmeldung. Eine E-Mail zur Bestätigung ist zu dir unterwegs.

Diese Seite ist eine Vorschau. \nKlicken Sie hier\n, um den Vorschaumodus zu verlassen.

Zu deiner Suche haben wir leider keine passenden Blogartikel gefunden. 

Du hast noch Fragen zu diesem Thema? Dann sprich mich einfach an.

Du hast Fragen oder möchtest dich mit mir austauschen? Gern - trag einfach deine Kontaktdaten ein.

Die \nDatenschutzerklärung\n der codecentric AG habe ich zur Kenntnis genommen und bestätige das mit der Absendung des Formulars.*

Entdecke spannende weiterführende Themen und lass dich von der codecentric Welt inspirieren.

Ich möchte eure Inhalte abonnieren. Diese Einwilligung kann ich jederzeit widerrufen bzw. den Blog Newsletter abbestellen. Die Datenschutzerklärung habe ich gelesen. Ich bin damit einverstanden, dass die codecentric AG, Hochstr. 11, 42697 Solingen, mich über eigene Dienstleistungen, branchenrelevante Events, Webinare und Publikationen der codecentric AG über die von mir angegebene E-Mail-Adresse informiert. Meine Einwilligung ist freiwillig. Ich kann meine Einwilligung jederzeit mit Wirkung für die Zukunft widerrufen (z.B. über den hierfür vorgesehenen Abmeldelink in unserem Newsletter, per E-Mail an datenschutz@codecentric.de oder per Post an die oben genannte Adresse der codecentric AG) erklären. Weitere Hinweise zur Verarbeitung meiner personenbezogenen Daten und zu meinen Rechten als betroffene Person finde ich hier: \nhttps://www.codecentric.de/datenschutzerklaerung\n

Bitte wähle deine/n gewünschte/n Standort/e aus.

Wir möchten dich als Mensch kennenlernen. Erzähle uns, wer du bist, worauf du Wert legst, wie du gerne arbeitest und was du von uns als Arbeitgeber erwartest.

Bitte füge deine Bewerbungsunterlagen ​(mind. Lebenslauf, gerne auch mehr)​​ ​als eine ​PDF-​Datei ​hinzu.

Es gilt die Datenschutzerklärung der codecentric AG, abrufbar unter \nDatenschutzbestimmungen\n. Mit dem Absenden des Formulars bestätige ich, dass ich diese zur Kenntnis genommen habe.

Job application form field placeholder 10

Keine passende Rolle gefunden? Bewirb dich bitte trotzdem initiativ! Jenseits von starren Schubladen zählen für uns vielmehr Motivation, Kompetenz und der Cultural Fit.

Job application form field placeholder 11

Dies ist ein Pflichtfeld, bitte fülle es aus.

Beyond CNNs: Wie Vision Transformers die Zukunft der Bildverarbeitung verändern

Aus dem Inhalt:

Lass uns austauschen!

Hier gibt's noch mehr spannende Infos von uns für euch

Gemeinsam bessere Projekte umsetzen.

Wir helfen deinem Unternehmen.

Unsere Leistungen

Hilf uns, noch besser zu werden.

Zu den Jobangeboten