Add pip setup extra_require for tensorflow (1.x, 2.x) and pytorch

[leonardozcm]

Collect extra pakages used in examples:

# pytorch
matplotlib
ray=0.8.6
Pillow=8.1.0
torch 
torchvision
jep==3.9.0
h5py==2.9.0
cloudpickle==1.6.0

# tensorflow 1.15
h5py==2.9.0
tensorlfow==1.15
Pillow=8.1.0
pypandoc==1.5
pandas
scikit-learn
tensorflow_datasets==3.2.0
psutil
tensorflow-gan

# tensorflow 2
tensorlfow==2.3.0
conda install -y cmake==3.16.0 -c conda-forge

Any suggestions?

[yangw1234]

remove the conda install, this is only for horovod

[yangw1234]

I think for tensorflow 1.15 only tensorflow==1.15.0 and tensorflow_datasets==3.2.0 is enough.

[yangw1234]

@hkvision @qiuxin2012 any suggestions on the PyTorch dependencies?

[hkvision]

ray and jep shouldn't be needed at the same time; also we may not need pillow for general pytorch applications?

[hkvision]

why is matplotlib needed? h5py and cloudpickle is only needed for spark backend right?

[leonardozcm]

why is matplotlib needed? h5py and cloudpickle is only needed for spark backend right?

Pillow and matplot are needed in torch examples. And Pillow are also imported in orca.image.file

[hkvision]

why is matplotlib needed? h5py and cloudpickle is only needed for spark backend right?

Pillow and matplot are needed in torch examples. And Pillow are also imported in orca.image.file

But they would only be needed for image applications; other applications don't need them. Shouldn't add them as a must.

[leonardozcm]

why is matplotlib needed? h5py and cloudpickle is only needed for spark backend right?

Pillow and matplot are needed in torch examples. And Pillow are also imported in orca.image.file

But they would only be needed for image applications; other applications don't need them. Shouldn't add them as a must.

ok, I'll not mark them as necessaray

[yangw1234]

ray and jep shouldn't be needed at the same time; also we may not need pillow for general pytorch applications?

any idea to differentiate ray and spark backend?