Fix Apple Silicon (ARM64) compatibility

[ilan-theodoro]

Summary

This PR addresses installation failures on macOS with Apple Silicon (M1/M2/M3) processors. The core functionality is working, but deep learning modules need further compatibility testing and optimization.

Status: DRAFT

• ✅ Core installation fixed
• ✅ CLI commands working
• ⚠️ Deep learning modules (TensorFlow/PyTorch) need compatibility verification
• ⚠️ Metal GPU acceleration needs testing

Problem

Users with Apple Silicon Macs cannot install aydin due to:

1. numpy 1.19.3 compilation failures (missing ARM64 support)
2. TensorFlow API import errors
3. Dependency version conflicts

Solution

• Added platform-specific dependencies for Apple Silicon
• Fixed TensorFlow import paths with try-except fallbacks
• Updated dependency constraints for ARM64 compatibility

Installation Instructions

For users who want to test this branch on Apple Silicon:

# Create conda environment
conda create -n aydin python=3.9 -y
conda activate aydin

# Install from this branch
pip install "aydin @ git+https://github.com/royerlab/aydin.git@fix/apple-silicon-compatibility"

Or if you need to install with editable mode:

# Clone and checkout branch
git clone https://github.com/royerlab/aydin.git
cd aydin
git checkout fix/apple-silicon-compatibility

# Create environment and install
conda create -n aydin python=3.9 -y
conda activate aydin
pip install -e .

Changes Made

1. TensorFlow Import Compatibility

• aydin/regression/nn_utils/models.py: Added try-except for keras imports
• aydin/nn/tf/models/utils/conv_block.py: Fixed BatchNormalization import path

2. Platform-Specific Dependencies in setup.cfg

# Apple Silicon specific
tensorflow-macos==2.8.0; platform_system=='Darwin' and platform_machine=='arm64'
tensorflow-metal==0.4.0; platform_system=='Darwin' and platform_machine=='arm64'
numpy>=1.21.2,<2.0; platform_system=='Darwin' and platform_machine=='arm64'

# Other platforms
tensorflow==2.8.1; platform_system!='Darwin' or platform_machine!='arm64'
numpy>=1.21.2,<1.24; platform_system!='Darwin' or platform_machine!='arm64'

3. Additional Constraints

• Added pydantic>=1.9.0,<2.0 for napari compatibility
• Updated scikit-image and scipy to use flexible versioning
• Added torch upper bound <2.0 for compatibility

Testing

• ✅ Tested on macOS 14.6 with Apple M-series processor
• ✅ Python 3.9 environment
• ✅ CLI commands (aydin --help, aydin denoise, etc.)
• ⚠️ Deep learning models need extensive testing

TODO Before Merge

• Test TensorFlow models on Apple Silicon with Metal acceleration
• Verify PyTorch compatibility and performance
• Test on Intel Mac to ensure backward compatibility
• Update CI/CD to include ARM64 testing
• Consider upgrading to newer TensorFlow/PyTorch versions
• Add installation documentation for Apple Silicon users

Notes

• tensorflow-metal provides GPU acceleration on Apple Silicon
• Some deep learning operations may need optimization for Metal Performance Shaders
• Consider migrating to tensorflow 2.13+ which has better ARM64 support

Related Issues

Fixes installation issues reported by Apple Silicon users.

🤖 Generated with Claude Code