1_simple

1. simple

hydra_jupyter_sample.ipynb

You could tinker with the cfg variable (omegaconf.DictConfig instance) and see how it works.
Compare the contents of cfg with config/train_cfg.yaml.

hydra_sample.py

Recommend running the following on the shell script as a tutorial.
You should notice that outputs/ directory is created when the scripts are single-run, and multirun/ directory is created when the scripts are multi-run.
Also, each directory that is created contains .hydra/ directory which has:

• the current user configurations (.hydra/config.yaml) for the process
• the current hydra configurations (.hydra/hydra.yaml) for the process
• the current overridden configurations (.hydra/overrides.yaml) for the process

The followings are example runs:

• python hydra_sample.py # basic functionality
• python hydra_sample.py -m run.random_seed=0,1,2,3,4 # multirun parameter sweep with -m flag
• python hydra_sample.py -m "run.random_seed=range(0,5) # multirun sweep with predefined command line hydra syntax
• python hydra_sample.py +new_variable=0 # Inserting new config values (you could also add predefined config_groups that was not included in the main config)
• python hydra_sample.py -m +new_variable=0,1,2,3,4 # Insert new config & sweep

configs that start with hydra are internal hydra configs that changes the behavior of the runs. They are internally defined so you have to refer to the documents.
Single run examples:

• python hydra_sample.py hydra.job.chdir=True # Change working directory (directory of process) when run. You can change this behavior to be the default by setting hydra.job.chdir: True in the config/train_cfg.yaml file
• python hydra_sample.py hydra.job.chdir=True hydra.run.dir=outputs/custom_directory # Specifying custom directory by overriding (hidden) hydra configuration
• python hydra_sample.py hydra.run.dir=outputs/custom_directory # Note that this does not change the working directory, but the logging file (hydra_sample.log) & .hydra/ files are generated in the hydra.run.dir.

Multirun examples:

• python hydra_sample.py -m hydra.job.chdir=True +random_seed=0,1,2,3,4 # Multirun with changing working directory
• python hydra_sample.py -m hydra.job.chdir=True hydra.sweep.dir=multirun/custom_directory +random_seed=0,1,2,3,4 # Specifying custom multirun directory by overriding (hidden) hydra configuration
• python hydra_sample.py -m hydra.sweep.dir=multirun/custom_directory +random_seed=0,1,2,3,4 # Specifying custom multirun directory to store hydra_sample.log & .hydra/, but does not change working direcotry
• python hydra_sample.py -m hydra.job.chdir=True hydra.sweep.dir=multirun/custom_directory hydra.sweep.subdir=run_${random_seed} +random_seed=0,1,2,3,4 # Specifying custom multirun directory & subdir
• python hydra_sample.py -m hydra.job.chdir=True hydra/launcher=joblib hydra.launcher.n_jobs=2 "+random_seed=range(0,5)" # Multirun sweep with parallel processes & n_job limit. Need to install hydra-joblib-launcher first! (pip install hydra-joblib-launcher)

Debug mode:

• python hydra_sample.py hydra.verbose=True # Run using different logging levels