Open EuphoriaCelestial opened 4 years ago
+1 on this one. Currently it's a big trial & error to get it running when installing from scratch
+1 the requirements is outdated, and I don't know which version I should use on tensorflow/numpy/numba/llvmlite ...
And this error occured:
ImportError: numpy.core.multiarray failed to import
For reference for people having issues installing deps later on, I'm using conda and this sequence of commands seemed to work well to create a new env that can run it, though it's likely not necessary to stretch it out like this:
conda config --add channels conda-forge
conda create --name tacotron2
conda install python=3.7
conda install tensorflow=1
pip install -r requirements.txt
conda install pytorch
I also deleted all the versions (except mpl) in requirements.txt before running the above, and removed tensorflow from it:
matplotlib==2.2
# tensorflow==1
numpy
inflect
librosa
scipy
Unidecode
pillow
Matplotlib will throw an error in the inference jupyter notebook on import, but from what I can tell it doesn't impact any of the default functionality
I've got everything to run with this but I'm worried the version numbers are affecting my output. I can't remember the reasoning for using python 3.6 but it was to get it to run on windows.
conda create -n tacotron python=3.6 conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.2 -c pytorch (if on windows) conda install -c conda-forge nvidia-apex
Then requirements.txt: matplotlib==2.1.0 tensorflow==1.15.2 numpy==1.16.0 numba==0.48.0 inflect==0.2.5 librosa==0.6.0 scipy==1.0.0 Unidecode==1.0.22 pillow future
I've got everything to run with this but I'm worried the version numbers are affecting my output. I can't remember the reasoning for using python 3.6 but it was to get it to run on windows.
conda create -n tacotron python=3.6 conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.2 -c pytorch (if on windows) conda install -c conda-forge nvidia-apex
Then requirements.txt: matplotlib==2.1.0 tensorflow==1.15.2 numpy==1.16.0 numba==0.48.0 inflect==0.2.5 librosa==0.6.0 scipy==1.0.0 Unidecode==1.0.22 pillow future
the librosa==0.6.0 lead numpy==1.16.0 change to numpy==1.19.5 why and how change
I'm not sure what you mean exactly.
Make sure you are using python 3.6. Its been a long time but I think you might be using that numpy because of a different cuda toolkit? I'm guessing windows? What GPU?
The numpy version in Tacotron 2 requirement file is 1.13.3, when numba, imgaug require >= 1.15, tensorflow require >= 1.16; torch==1.0 in Waveglow's requirement file is also outdated since the code is using torch.utils.tensorboard, which only appear in newer version of Pytorch