Installation is not reproducible: dependency/version mismatches (Windows)

[srividya-0001]

Hi all,

While testing DeepLenseSim end-to-end on Microsoft Windows 10 Pro, I noticed the current installation workflow is not fully reproducible due to missing dependencies and version mismatches.

Environment tested

• OS: Microsoft Windows 10 Pro
• Python: 3.11 (fresh venv)

What happens

Even after installing the repo, running the simulation scripts fails unless additional dependencies are installed manually.

1) Missing dependencies during runtime

Example errors observed during execution:

• ModuleNotFoundError: No module named 'matplotlib'
• ModuleNotFoundError: No module named 'mcfit'

These packages are required by the scripts/pyHalo stack, but they are not installed automatically.

2) Version mismatch between pyHalo and lenstronomy

The README suggests:

• lenstronomy==1.9.2

However, with current pyHalo (1.4.3) I encountered:

ModuleNotFoundError: No module named 'lenstronomy.LensModel.Profiles.pseudo_double_powerlaw'

which indicates pyHalo expects a newer compatible lenstronomy version.

Suggested improvement

Instead of adding a new requirements file, it may be better to update setup.py to include the required dependencies via install_requires (and optionally specify compatible version ranges).
This would make pip install -e . or pip install . reliably install everything needed to run the scripts.

It may also help to clarify supported Python versions (Python 3.10/3.11 worked reliably; Python 3.14 caused dependency build/wheel issues on Windows).

I’m happy to submit a PR updating setup.py dependencies if that’s helpful.

Note: I tested using the latest commits from mwt5345/DeepLenseSim#1 (PR branch), since the fixes are not merged into main yet.

Thanks...!

[srividya-0001]

Update: I submitted a PR with the proposed fix (setup.py dependency updates + ignoring generated output files) here:
mwt5345/DeepLenseSim (mwt5345/DeepLenseSim#2)

This PR is currently based on PR (mwt5345/DeepLenseSim#1 the fixes are not merged yet.

Thanks..!

[ShauryaGupta2006]

Hello @srividya-0001
i have gone through the whole issue and found it intersting i would love to work on this and take these problems to some solution
Key problems that i figured out :
1. Missing dependencies during runtime
2. Version mismatch between dependent libraries
3. working fails in clean virtual enviroment
4. Unsupported Python versions
5. platform based incompatibility

**Approach **

To resolve the issue related missing dependencies i will mention all the require liberaries in the setup.py file under install_requires so they get installed automatically at the runtime time.

Then in version mismatch issue which mainly arisies due to conflict of two liberaries due to required version so for that we can figure out the suitable version for each other and let programe install those specific versions only during runtime.

In the third issue we can set up fresh virtual enviroment and verify the working of pip whether it requires manual working or not.

For unsupported python version we can handle this similarly as we are handling with other liberaries here.

5th platform incompatibilty for this we have make check by running on diffrent platforms and figure out the actual reasons of crashing and then specificly resolve them... These are some of my solutions on the issues. i am ready to work on these from my side.

Thank you

[manshirahangdale]

Hi, I’ve opened a PR addressing this issue by adding a minimal requirements.txt
to improve installation reproducibility.
While testing on Windows with a fresh virtual environment, I encountered
missing dependencies (e.g., einops when running
DeepLense_Regression_Zhongchao_Guan/CNNT.py). The PR adds commonly used
dependencies so users can install them with:
pip install -r requirements.txt

PR: #185

[Akarshak51]

The setup.py vs requirements.txt debate is worth clarifying for reproducibility. Here's a structured approach:

Root Cause Analysis:

The installation failures stem from three distinct dependency management gaps:

1. Runtime vs Install-time Dependencies: Core packages (matplotlib, mcfit, einops) missing from install_requires
2. Transitive Version Conflicts: pyHalo 1.4.3 expects lenstronomy features from versions >=1.11.x, but README pins 1.9.2
3. Platform-specific Build Issues: Python 3.14 wheel compilation failures on Windows (likely due to immature ecosystem support)

Recommended Solution - Hybrid Approach:

# setup.py
setup(
    name="DeepLenseSim",
    install_requires=[
        "numpy>=1.21.0,<2.0.0",
        "lenstronomy>=1.11.0,<1.13.0",  # Compatible with pyHalo 1.4.3
        "pyHalo>=1.4.0,<1.5.0",
        "matplotlib>=3.5.0",
        "mcfit>=0.0.18",
        "einops>=0.6.0",
        "torch>=1.12.0",
    ],
    extras_require={
        "dev": ["pytest>=7.0", "black>=22.0"],
        "gpu": ["cupy-cuda11x>=11.0.0"],
    },
    python_requires=">=3.10,<3.14",  # Explicit Python version bounds
)

Why This Works:

• install_requires ensures pip install . captures all runtime dependencies
• Version ranges allow minor updates while preventing breaking changes
• python_requires fails fast on unsupported interpreters
• Platform-specific dependencies isolated in extras

Requirements.txt for Reproducibility:

# requirements.txt - Generated from working environment
lenstronomy==1.11.4
pyHalo==1.4.3
matplotlib==3.7.1
numpy==1.24.3
# ... pinned versions for exact reproducibility

Testing Protocol:

# Clean test on Windows 10
python -m venv test_env
test_env\Scripts\activate
pip install -e .
python -m pytest tests/  # Validate installation

Alternative - Poetry for Modern Dependency Resolution:

If migrating to Poetry is an option, it handles transitive conflicts automatically:

[tool.poetry.dependencies]
python = "^3.10,<3.14"
lenstronomy = "^1.11.0"
pyHalo = "^1.4.0"

Poetry's lockfile ensures reproducible installs across platforms without manual version detective work.

Would be happy to test a PR with updated setup.py on Windows 10/11 environments to validate cross-platform compatibility.