Recipes for creating Tensorflow conda packages.
In the defaults
channel Tensorflow is provided via a number of packages.
As of version 1.11.0, the library itself is provided by the tensorflow-base
package. Different variants of this package are created for each platform from
the tensorflow-base-cpu and tensorflow-base-gpu recipes.
Installing the tensorflow
package using conda installs both the tensorflow
library as well as tensorboard. The tensorboard recipe is used to create
the tensorboard
package.
The tensorflow
metapackage package is created by the tensorflow recipe.
The tensorflow
metapackage depends on tensorboard
, an exact
build of tensorflow-base
and the version of the _tflow_select
package
which matches the tensorflow-base
variant.
The _tflow_select
package, created from the _tflow_select recipe,
establishes the priority of the variants using the version number. The variant
with the highest version will be installed by default. The non-default variant
can be installed using the tensorflow-mkl
, tensorflow-eigen
and
tensorflow-gpu
packages which are created from the tensorflow-variants
recipe. Note that not some platforms do not support all variants.
Available Recipe:
- tensorboard : Tensorboard.
- tensorflow : Metapackage which installs tensorflow-base and tensorboard.
- tensorflow-base-cpu : Eigen and MKL variants of the Tensorflow library.
- tensorflow-base-gpu : GPU variant of the Tensorflow library.
- tensorflow-variants : Recipe used to create tensorflow variant packages, e.g. tensorflow-mkl.
- _tflow_select : Metapackage to establish priority in tensorflow-base packages.