lm-checkpoints

🤖🚩lm-checkpoints

Simple library for dealing with language model checkpoints to study training dynamics.

lm-checkpoints should make it easier to work with intermediate training checkpoints that are provided for some language models (LMs), like MultiBERTs and Pythia. This library allows you to iterate over the training steps, to define different subsets, to automatically clear the cache for previously seen checkpoints, etc. Nothing fancy, simply a wrapper for 🤗 models that should make it easier to study their training dynamics.

Install using pip install lm-checkpoints.

Checkpoints

Currently implemented for the following models on HuggingFace:

• The Pythia models
• MultiBERTs

Example

Say you want to compute some metrics for all model checkpoints of Pythia 160m, but only seed 0.

from lm_checkpoints import PythiaCheckpoints

for ckpt in PythiaCheckpoints(size=160,seed=[0]):
    # Do something with ckpt.model, ckpt.config or ckpt.tokenizer
    print(ckpt.config)

Or if you only want to load steps 0, 1, 2, 4, 8, 16 for all available seeds:

from lm_checkpoints import PythiaCheckpoints

for ckpt in PythiaCheckpoints(size=160,step=[0, 1, 2, 4, 8, 16]):
    # Do something with ckpt.model, ckpt.config or ckpt.tokenizer
    print(ckpt.config)

Alternatively, you may want to load all final checkpoints of MultiBERTs:

from lm_checkpoints import MultiBERTCheckpoints

for ckpt in MultiBERTCheckpoints.final_checkpoints():
    # Do something with ckpt.model, ckpt.config or ckpt.tokenizer
    print(ckpt.config)

Loading "chunks" of checkpoints for parallel computations

It is possible to split the checkpoints in N "chunks", e.g., useful if you want to run computations in parallel:

chunks = []
checkpoints = PythiaCheckpoints(size=160,seed=[0])
for chunk in checkpoints.split(N):
    chunks.append(chunk)

Dealing with limited disk space

In case you don't want the checkpoints to fill up your disk space, use clean_cache=True to delete earlier checkpoints when iterating over these models (NB: You have to redownload these if you run it again!):

from lm_checkpoints import PythiaCheckpoints

for ckpt in PythiaCheckpoints(size=14,clean_cache=True):
    # Do something with ckpt.model or ckpt.tokenizer

Evaluating checkpoints using lm-evaluation-harness

If you install lm-checkpoints with the eval option (pip install "lm-checkpoints[eval]"), you can use the evaluate function to run lm-evaluation-harness for all checkpoints:

from lm_checkpoints import evaluate, PythiaCheckpoints

ckpts = PythiaCheckpoints(size=14, step=[0, 1, 2, 4], seed=[0], device="cuda")

evaluate(
    ckpts,
    tasks=["triviaqa", "crows_pairs_english"],
    output_dir="test_results",
    log_samples=True,
    skip_if_exists=True,
#    limit=5, # For testing purposes!
)

Or you can use the evaluate_checkpoints script:

evaluate_checkpoints pythia --output test_results --size 14 --seed 1 --step 0 1 2 --tasks blimp crows_pairs_english --device cuda --skip_if_exists

Both examples will create a subdirectory structure in test_results/ for each model and step. This will contain a results json file (e.g., results_crows_pairs_english,triviaqa.json), and if using the --log_samples option, a jsonl file containing the LM responses to the individual test items for each task (e.g., samples_triviaqa.jsonl).