# CalculatedContent/WeightWatcher

The WeightWatcher tool for predicting the accuracy of   Deep Neural Networks

Repository: https://github.com/CalculatedContent/WeightWatcher
Canonical: https://ross.abutalabs.com/products/weightwatcher
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-05-11T16:40:41+00:00

## Health v2 (maintenance only)
Score: 59/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 81, release rhythm 8, longevity 100
- inputs: {"age_days": 2835, "days_push": 114, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1772, forks 145 (observed 2026-08-28T04:05:34.248183+00:00)

## What it is
WeightWatcher is an open-source Python diagnostic tool for analyzing pre-trained deep neural networks without access to training or test data. It applies Random Matrix Theory and heavy-tailed self-regularization theory to predict test accuracy, detect over/under-training, and flag issues in model compression or fine-tuning.

## Use cases
- predict test accuracy of a neural network without test data
- check if my model is over-trained or over-parameterized
- analyze pretrained pytorch or keras model layer quality
- detect problems when fine-tuning or compressing a model
- compare quality of different deep learning models
- get layer-level warning labels for under-trained layers

## When to choose
- you need data-free diagnostics of trained Conv2D or Dense layer models in PyTorch or Keras
- you want to rank or compare models by predicted accuracy without running evaluations
- you are researching why deep learning works or model quality metrics

## When to avoid
- you need full model evaluation with real test data
- you work with architectures beyond Conv2D and Dense layers, like transformers or attention layers
- you need production model monitoring rather than offline analysis

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, monitoring, analytics, benchmarking
- domain: deep-learning, machine-learning, developer-tools, data-science
- platform: python
- tags: neural-network-analysis, random-matrix-theory, model-diagnostics, heavy-tailed-self-regularization, model-quality, pytorch, keras

## Member repositories
- CalculatedContent/WeightWatcher (main) score 59

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:34.248183+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:25:56.776068+00:00, confidence not recorded.
  - readme: https://github.com/CalculatedContent/WeightWatcher (fetched 2026-08-28T04:05:34.248183+00:00, sha 997102bb615d)
  - registry_pypi: https://pypi.org/pypi/weightwatcher/json (fetched 2026-08-29T11:04:03.439804+00:00, sha 65ef245524d0)
- Data as of 2026-08-30T08:39:29.467469+00:00.