Skip to content

Latest commit

 

History

History
101 lines (69 loc) · 2.98 KB

README.md

File metadata and controls

101 lines (69 loc) · 2.98 KB
Raw

config files are defined using facebook hydra.

Hydra defines a modular config, which can be altered using the command line.

Everything at once

SPLADE_CONFIG_NAME=config_splade python -m src.all will train, index, retrieve and evaluate.

More examples (train)

Train toy splade model (default)

python -m src.train config.checkpoint_dir=<MY_RES_DIR>

Train another model

Option 1: create/use another config file such as: conf/config_splade.yaml:

# FILES
defaults: # (these specify which config FILES to use)
  ############## TRAIN ###################################
  - train/config: splade
  - train/data: msmarco
  - train/model: splade
  ############## INDEX ###################################
  - index: msmarco
  ############## RETRIEVE ################################
  - retrieve_evaluate: all

# Direct PARAMETER setting
config: # to be provided for each run
  checkpoint_dir: ???
  index_dir: ???
  out_dir: ???
  fp16: false

To tell the code to use this config file, either modify CONFIG_NAME to config_splade in conf/CONFIG_CHOICE.py to use it instead of config_default.yaml

or

equivalently, define and use an environment variable: SPLADE_CONFIG_NAME=config_splade python -m src.train

It is also possible to re-use a config saved from a previous experiment; in this case provide its full path: SPLADE_CONFIG_FULLPATH=/path/to/my/exp/config_splade.yaml python -m src.train

Option 2: provide command line arguments for everything that changes from default or desired config, specifying files or parameters to change:

python -m src.train config.checkpoint_dir=<MY_RES_DIR> train/data=msmarco train/model=splade init_dict.model_type_or_dir='bert-base-uncased' ...

Note on naming in hydra

TLDR: for overriding a parameter, use the parameter package naming, for overriding a package, use the package path

Overriding a parameter

If you are overriding a parameter from the command line you should name is using its package position. For example, overriding model_type_or_dir: this parameter is defined in train/config/splade.yaml, and the content of this file is:

# @package _global_

init_dict:
  model_type_or_dir: distilbert-base-uncased
  model_type_or_dir_q: null
  freeze_d_model: 0
  agg: max
  fp16: true

config:
  tokenizer_type: distilbert-base-uncased

Therefore, the parameter is referred to with init_dict.model_type_or_dir=...

python -m src.train init_dict.model_type_or_dir=<CHOICE_FOR_MODEL>

if the header in train/config/splade.yaml was # @package awesome_package the command would thus be:

python -m src.train awesome_package.init_dict.model_type_or_dir=<CHOICE_FOR_MODEL>

Overriding a package

Override a package by referring to its path:

python -m src.train train/data=msmarco

Index, retrieve

same thing with src.index, src.retrieve.