I'm currently working on a machine learning project that includes several pre-trained models written in Python. I'm looking to turn this project into executable files for easier use by others, but I'm unsure if PyInstaller is the right tool for this job. I initially considered using Kivy to create an Android app, but I've decided to focus on desktop applications instead. I would love to hear about others' honest experiences with PyInstaller—does it really work well for what I'm trying to achieve?
5 Answers
If you're working on smaller projects, PyInstaller can be fine, but it gets tricky with larger applications. I've had success minimizing the final executable by excluding unnecessary modules using the `--exclude-module` flag. Just be cautious, though—stay away from heavy libraries like PyTorch!
Honestly, I've never liked PyInstaller. It tends to create huge files, and I find it pretty slow and fragile. Python isn't really designed for distributing as a binary. I think it's generally better to distribute your project as a package, and let users install it using tools like pip.
I've had good results with PyInstaller for a few projects, but keep in mind that executables can become quite large, especially with libraries like TensorFlow. For one of my projects, the final executable weighed in at 500MB just because of the dependencies! If you're looking for something with a GUI, auto-py-to-exe is a good option, but you're still going to face the size issue.
Does it actually make a difference in performance if everything's packed into one .exe versus another file in the same folder? I wonder how RAM usage plays into that.
Just a heads up: you might run into false positives from antivirus software when using PyInstaller. This happens a lot with any tool that generates executable files from Python scripts. To avoid this, you’ll need to purchase a yearly certificate to sign your binaries.
Yeah, I'm with you on that! If you’re dealing with machine learning models, I’d recommend using ONNX for deployment. You could just download the model weights and store them locally instead of relying on a bulky package.
That size problem exists across the board! Once compiled, you can't really shrink the library—unless you get into some complex stuff with Nuitka and advanced compression techniques.