July Instructions for Chris - NGN
From SnOwy - Ed's Wiki Notebook
- This is a really really dirty hackjob, so ... here.
- Note: if you screw up anytime during the build, just do "make clean" to nuke everything and start over
Contents |
Making the Source Files for the NGN
Required: Python 2.6, flex, bison
- Untargz the archive in a place in Ubuntu somewhere
- cd into the untarred directory.
- do "make build_source"
- at this point, you'll have working source files-- you now have three options:
Option 1: Run a Local Test
- do "make compile_source" (will produce binaries for your local machine)
- do "make run_local"
- this will launch A FEW processes, to terminate them, kill them in this order--
- "killall make"
- "killall python"
- "killall smiles"
- "killall inchi"
Option 2 and 3: Run this on SharcNet
- go up to the previous directory, and targz the directory
- sftp the archive into SharcNet (Use WHALE if possible)
- untar the archive in SharcNet (in /work/<your_user_directory>/ or a subdirectory here-- BUT NOT /home/<your_user_directory>)
- cd into the untarred directory.
- do "make compile_source" (will produce binaries for Linux on SharcNet)
- you can now finish option 2 or 3 below...
Option 2: Run a TEST on SharcNet
- do "make run_sharcnet"
Option 3: Run actual experiment on SharcNet
- do "make run_sharctest"
General notes about SharcNet
- to kill your processes, use sqkillall.py [1] (instructions here).
How do I check on experiments?
- use "sqjobs" to look at time run
- cd into <untarred_directory>/expdir/* and look at the number of weight files
- in the above directories, you can also cat *trainlog*txt and cat *testlog*txt to eyeball convergence ratios
What do I do after experiments are done?
- archive the entire experiment and get it back with sftp; we can look at the data together afterward.
