sequence-models
Analyzing properties of sequence models empirically and theoretically.
Experimentation Framework
- The orchestrator script:
usage: python3 orchestrator.py [-h] [--num_gpus NUM_GPUS]
[--exps_per_gpu EXPS_PER_GPU]
specification_dir out_dir
The experiment orchestrator. This takes as argument a directory containing
experiment specifications (a bunch of jsons) and the desired parallelism (# of gpus, and workers per gpu) with which to run a set of experiments
positional arguments:
specification_dir A directory containing experiment specifications
generated by specifier.py
out_dir The output directory
optional arguments:
-h, --help show this help message and exit
--num_gpus NUM_GPUS Number of GPUs to use
--exps_per_gpu EXPS_PER_GPU
Number of examples per GPU
- The experiment runner: exp_runner,py
-> This is the main entrypoint for experiments, there is a main function called
run_experiment(spec, experiment_directory):
this takes a dictionary called spec (which is the JSON that specifies the experiment, containing name, # of layers, learning rate, batch size, etc)
and then "runs an experiment" with it.
-> This method should basically output all the information we need to a directory given by "experiment_directory"