I am trying to get Hailo’s Dataflow Compiler (DFC) working on a Windows 11 machine using Docker Desktop. This has been one of the most frustrating developer experiences I’ve had.
I am basically trying the steps shown in this tutorial, with the main difference that I am not installing everything on my host system, but rather in a Docker container.
The Steps I Took
Created a Dockerfile based on Ubuntu 22.04.
Downloaded the Dataflow Compiler wheel and installed it via pip.
Cloned and installed Hailo Model Zoo from the GitHub repository.
Installed the Hailo runtime.
Conflicting Dependencies
Now, I’ve hit a hard dependency conflict:
DFC requires numpy exactly at 1.23.3.
Model Zoo requires numpy >1.24 and <2.2.
This makes it impossible to satisfy both dependencies simultaneously. Yet, Hailo’s own tutorials rely on Model Zoo commands to run the DFC. How is this even supposed to work? Did anyone at Hailo actually test this?
Docker is Supposed to Solve This – But Hailo’s Implementation Breaks It
I use Docker to keep environments isolated and reproducible, but Hailo’s tooling completely ignores Docker best practices:
The official Docker image does not start from a simple Dockerfile – it requires a shell script to initialize.
That shell script does not run on Windows, making the image not portable.
This defeats the whole purpose of Docker, which is meant to be self-contained and cross-platform.
Yes, I could try WSL, but that led to even more issues, so I decided to just build my own Docker image. And now I’m running into these dependency conflicts.
Installation is Broken
Hailo requires hailort_4.20.0_amd64.deb, but this cannot be installed during a Docker build.
The Model Zoo and DFC do not play nicely together due to the numpy conflict.
The lack of clear documentation leaves users guessing how anyone is even getting this to work.
How Are People Making This Work?
If anyone here has actually gotten DFC and Model Zoo to work together: what’s the trick?
Are you using an older version of Model Zoo?
Are you using some hidden workaround that isn’t documented?
Is there some other approach that avoids this mess?
Hailo: Just Provide a Proper Docker Image
Instead of forcing developers to go through this nightmare, why not just provide a prebuilt Docker image with DFC, Model Zoo, and all required dependencies included?
This is standard practice for AI/ML toolchains, and it’s ridiculous that we have to manually debug dependency conflicts for tools that Hailo itself built.
I get that Hailo’s main business is providing AI solutions for enterprises, and that the AI Kit for Raspberry Pi is a newer product for them. But if they want this to work—if they want the hobbyists, prototypers, and the broader dev community to embrace it—they need to seriously improve the usability and developer experience of their toolchain. Right now, it’s just way too frustrating to get started.
At the very least, Hailo should acknowledge and document these issues instead of leaving developers to waste time figuring them out.
I agree though, the documentation is not acceptable for a consumer product. I am able to navigate it as I deal with worse in my day job, but the ras pi is targeted to beginners. The point of the ai kit / hat was to get newbies interested in computer vision.
Yeah, even for building prototypes. How much time will be spent trying to figure this piece out? These projects have many parts. Most of them can be developed very rapidly. But when encountering this piece, the project comes to a screeching halt.
I could imagine a lot of companies interested in building prototypes or proof of concepts using raspberry pi and AI kit. But these things need to be developed fast.
Have you considered to use WSL? I have Windows11 with ubuntu 22.04 WSL, it works like a charm.
I use to convert .onnx into .hef files, then move the .hef in my embedded hardware.
The most difficult thing is to do work with custom models
Yes, I have tried it. But my project involves many other parts (e.g. label studio web server, yolo training step, etc.). I use docker-compose to run each of these services. Everything is managed in code and version controlled in my repo. I have an orchestrater service that coordinates moving data from one service to the next.
Why should I say, “because of this one step, I need to change my whole architecture and orchestration “? I should be able to just run the compilation step in a docker container, just like every other service.
I understand, that could be very frustrating.
In my opinion Hailo is born for embedded systems and practically all devices are based on linux, so, probably windows users are a few, that’s why there’s no so much support
I am totally surprised by that.
Yes, most execution environments are Linux. But most development environments are not. I work with many professional developers, and they are mostly using Mac or Windows, combined with Docker. They use their computers for more than just development. They belong to organizations that dictate what OS they are allowed to work on, because these organizations have IT departments and central management of hardware and software.
Even hobbyists are unlikely to be using their computers for only development.
I mean, the whole concept of user Docker is to deal with this issue that the execution environment will often be different from the developer’s host system. It is built so that execution environments are portable, and can be defined as code and built into CI-CD pipelines.
Frustrating is just the tip of the iceberg. Frustration I could live with.
Thank you for sharing your concerns about our Windows support. I understand your frustration, and I appreciate you taking the time to provide this feedback.
To clarify, DFC currently operates on WSL (by installing from whl), though we haven’t officially tested or announced full Windows compatibility yet. You’re helping us identify areas where we can improve!
Regarding our AI Suite Docker, It can’t be completely OS independent because it includes HailoRT for hardware inference. Our software has been optimized primarily for Linux environments, where users experience a smooth setup process with minimal configuration required.
About the numpy versions mentioned - we haven’t encountered this issue , I re tested everything and it works well, if you can provide more info i can look at it again.
We’ll be releasing comprehensive guides in our community soon to improve the development experience with our SDK, including better compilation guides and usage instructions. Hailo’s software ecosystem is built with developer experience in mind, making it easy to get started and integrate into existing workflows.
Hi @omria
Thanks for your input. I’m glad to hear there is work going on, for this topic.
To be clear, I am not advocating for a full software suite that runs on windows. I am advocating for a docker image for the DFC that can run on any OS.
I understand that the full software suite can’t be completely os-independent when running inference using Hailo hardware.
But model training, parsing and compiling to HEF will generally not happen on edge devices (since edge devices are underpowered).
I envision having a Docker image for preparing my model, and a different execution environment for running inference on that model on an edge device.
The tools needed to do the model preparation steps should be available in a Docker image (ideally hosted on Dockerhub), and should be able to run on any os that has docker installed. This would make it easier to integrate with existing pipelines.
This is doable.
In fact I have a working example (see the tutorial link I shared above). I provided a Dockerfile that successfully builds an image that should be os-agnostic (I’m running it using docker desktop on windows, which uses WSL behind thce scenes, but the image should be portable).
The sdk from Hailo is a crazy nightmare. Thank you for your job, I hop Hailo will invest to create real working tools for their products. One again, thank you so much. Form my side, I spent one week to find dependency for nothing…
