ARM 64bit - Failed to initialize NumPy: module compiled against API version 0xe but this version of numpy is 0xd

🐛 Description

I wanted to do some testing on Raspberry Pi OS Bullseye 64bit and encountered a dependency problem with Numpy 1.19.5. After successful cloning and building the 0.5.0 version running ‘tts’ gives the error:

UserWarning: Failed to initialize NumPy: module compiled against API version 0xe but this version of numpy is 0xd (Triggered internally at  /root/pytorch/torch/csrc/utils/tensor_numpy.cpp:68.)

This seems to be related to an incompatibility between Numpy 1.19.5 and Torch 1.10.1, 1.9.0 and 1.8.1 (all I could try).

I tried upgrading Numpy: pip install numpy --upgrade which will install Numpy 1.22.0, but break everything else since the Numpy 1.19.5 dependency seems to be very strict.

To Reproduce

• Install Raspberry Pi OS Bullseye 64bit
• Install Python and some dependencies. I had to run something like: sudo apt-get install python3-venv libatlas-base-dev llvm-9 libmecab-dev and set sudo ln -s /usr/bin/llvm-config-9 /usr/bin/llvm-config (not sure if llvm is actually required but I needed it when I played with YourTTS)
• Create virtual env: python3 -m venv env && source env/bin/activate
• Clone the repo: git clone https://github.com/coqui-ai/TTS
• Build: cd TTS && pip install -e .
• Try something simple like: tts --list_models

Expected behavior

All dependencies resolve properly and TTS works.

Environment

• 🐸TTS Version: 0.5.0
• PyTorch Version: 1.10.1
• Python version: 3.9.2
• OS: Raspberry Pi OS 64bit (Aarch64), Kernel: 5.10.63-v8+
• CUDA/cuDNN version: -
• How you installed PyTorch (conda, pip, source): pip

[jbroeckling]

Did you downgrade from 3.9? I usually find it pretty troublesome on Pi to use a different version than the default.

yea, but for another reason. I just made a Python 3.9 venv and installing TTS with numpy==1.22.1 works and gives me better results. Downgrading Python is not recommended, cause it makes more trouble 😄

[fquirin]

installed numpy==1.21.5 and the issue is gone (and till now, nothing else broke down)

nice.

I was using Python 3.7.11

Did you downgrade from 3.9? I usually find it pretty troublesome on Pi to use a different version than the default.