Running sv-channels without training

[lsantuari]

• rename run.sh as train.sh
• create a new bash script call.sh:

input:

• 5 pre-trained models, one for each SV type
• the array windows.npz to run the models on for prediction/calling

The script

• generates the (unlabelled) windows from the input files:
  run.sh from start till add_win_channels
• use the pre-trained models on the windows to predict the SVs:
  use predict (related to Generate SV predictions in VCF format #64)
• outputs the SV calls in VCF format:
  first merge the calls, then convert to VCF

The snakemake way

• 2 parallel workflows, one for the training and one for the calling:
• the training workflow outputs the model(.hdf5)
• the test sample output windows(.npz)
• the calling is done by processing the windows of the test sample with the model trained on the training sample

[arnikz]

* rename [run.sh](https://github.com/GooglingTheCancerGenome/sv-channels/blob/dev/run.sh) as _train.sh_

* create a new bash script _call.sh_:

Instead, these could be the two modes (param) of the same "workflow" (to-be) implemented in one script run.sh. Could you indicate it in this figure?

[arnikz]

I've just noticed that this was added half-way

sv-channels/run.sh

Line 265 in aad71ff

exit 0

which means that the final checks were not executed. For now, I've disabled the code related to training (923b9ce)

[arnikz]

Ooops

[lsantuari]

Add a smoke test with the predict.py function here:

sv-channels/run.sh

Line 332 in 3a9be2d

with the windows generated by run.sh

[arnikz]

I don't know what you'd like to pass via CLI so added a place-holder here:

sv-channels/run.sh

Line 336 in 7d0d18e

#p=predict.py

Please continue in dev-merge branch (7d0d18e).