In a 'new computer' with Ubuntu 20.04 (using docker and pulling ubuntu:20.04), if I install miniconda3 and just run:

conda install -c anaconda tensorflow-gpu

Everything is good to go to use GPU for machine learning, because I can run:

import tensorflow as tf  
print('Num GPUs Available: ', len(tf.config.list_physical_devices('GPU')))
Num GPUs Available:  1

This is okay.

But the 'problem' is that anaconda installs a lot of packages when I run conda install -c anaconda tensorflow-gpu

Before run the command conda install -c anaconda tensorflow-gpu, if I run conda list I get:

After run the command conda install -c anaconda tensorflow-gpu, if I run conda list I get:

I know that the following packages are needed:

cudnn
tensorflow-gpu

, so is anything else needed to run the commands?:

import tensorflow as tf  
print('Num GPUs Available: ', len(tf.config.list_physical_devices('GPU')))
Num GPUs Available:  1

,or are all packages installed with conda install -c anaconda tensorflow-gpu necessary?

As the title says, I would like to know which are the strictly required (bare minimum) libraries/packages to run this

Thanks in advance

CodePudding user response：

Yes, it needs a lot more.

Typically, a large and complicated package like tensorflow has a whole tree of dependencies. If I take your list of packages after the install and remove the packages before the install, the following results:

'_tflow_select', 'absl-py', 'aiohttp', 'aiosignal', 'astor',
'astunparse', 'async-timeout', 'attrs', 'blas', 'blinker',
'c-ares', 'cachetools', 'click', 'cudatoolkit', 'cudnn',
'cupti', 'dataclasses', 'frozenlist', 'gast', 'google-auth',
'google-auth-oauthlib', 'google-pasta', 'grpcio', 'h5py',
'hdf5', 'importlib-metadata', 'intel-openmp', 
'keras-preprocessing', 'libgfortran-ng', 'libgfortran4',
'libprotobuf', 'markdown', 'mkl', 'mkl-service', 'mkl_fft',
'mkl_random', 'multidict', 'numpy', 'numpy-base', 'oauthlib',
'opt_einsum', 'protobuf', 'pyasn1', 'pyasn1-modules', 'pyjwt',
'python-flatbuffers', 'requests-oauthlib', 'rsa', 'scipy',
'tensorboard', 'tensorboard-data-server', 
'tensorboard-plugin-wit', 'tensorflow', 'tensorflow-base',
'tensorflow-estimator', 'tensorflow-gpu', 'termcolor', 
'typing-extensions', 'typing_extensions', 'werkzeug',
'wrapt', 'yarl', 'zipp'

Tensorflow depends both on a number of Python and C/C libraries. Each of those may have dependencies of their own. For example, tensorflow requires keras which requires hdf5. And tensorflow requires numpy which requires a BLAS library (in this case mkl) which requires the Fortran runtime.

Now, it may be that some of those dependencies are optional. But at first glance I don't see any of those.

Trying to pare down the dependencies is a significant task; you would basically have to build the whole dependency tree from source, for every dependency checking which of its dependencies are optional and if you want to do without them.

Personally, I would not bother in this case.