search

AnacondaRecipes / tensorflow_recipes

Tensorflow conda recipes
27 stars 26 forks source link

Tensorflow 2.3 conda dependencies conflict with pip dependencies #27

Open ZOUG opened 3 years ago

ZOUG commented 3 years ago

The conda dependencies differ from (and conflict with certain) pip dependencies. Details of tensorflow 2.3.0 are shown in the following. This is for both Python 3.7 and 3.8 environments.

conda dependencies pip dependencies
gast >=0.4.0,<0.4.1.0a0 gast==0.3.3
scipy >=1.5.2 scipy==1.4.1
keras-preprocessing >=1.1.0 keras-preprocessing<1.2,>=1.1.1
numpy >=1.16.6,<2.0a0 numpy<1.19.0,>=1.16.0

If only one of conda or pip is used, it would be fine. However, in some cases pip is needed to install packages in a conda environment. If a package built on top of tensorflow needs to be installed using pip, a conflict would cause problems.

I notice that the comments in the configuration file says "Some are higher than TF's lower bound. These are higher because they are the lowest version that we have.". But scipy 1.4.1 and gast 0.3.3 are available for Python 3.8 (as well as Python 3.7) on Anaconda (see below). So I wonder what that really means.

> conda search scipy=1.4.1 --info

...

scipy 1.4.1 py38h9439919_0
--------------------------
file name   : scipy-1.4.1-py38h9439919_0.conda
name        : scipy
version     : 1.4.1
build       : py38h9439919_0
...
md5         : cb1746d20633a0267a609b7882346f0d
timestamp   : 2020-02-07 17:17:46 UTC
dependencies:
  - blas 1.0 mkl
  - icc_rt >=2019.0.0
  - mkl >=2019.4,<2021.0a0
  - mkl-service >=2,<3.0a0
  - numpy >=1.14.6,<2.0a0
  - python >=3.8,<3.9.0a0
> conda search gast=0.3.3 --info

Loading channels: done
gast 0.3.3 py_0
---------------
file name   : gast-0.3.3-py_0.conda
name        : gast
version     : 0.3.3
build       : py_0
...
md5         : 59bdcc125a9b94505a63d10371bafec3
timestamp   : 2020-02-13 16:14:22 UTC
dependencies:
  - python