aind-disrnn-utils

aind_disrnn_utils

⚠️ Retired — no longer maintained (2026-09-01)

This package is retired. It has no active maintainer and will receive no further releases, fixes, or dependency bumps.

Nothing you have installed will break. This repository is preserved read-only: existing git+https://...@<sha> pins keep resolving, and every published release stays installable from PyPI. It is kept in place partly because frozen reproducibility artifacts reference it by SHA (for example studies/01-gru-scaling-law/environment.lock in aind-disrnn-dispatcher). Do not delete it.

Where the code went

The two functions that had real consumers — create_disrnn_dataset and add_model_results — were vendored into the wrapper that used them:

• aind-disrnn-wrapper → code/data_loaders/disrnn_dataset.py

They were copied byte-identically from commit 74de874d, then improved in place. See aind-disrnn-wrapper#66 for the audit and rationale.

data_models.py (disRNNInputSettings, disRNNOutputSettings) was not carried over. Nothing imported it; Hydra configs replaced it as the configuration surface. This closes #21 by obsolescence.

Known issues, now permanent

Two defects are present in the final release and will not be fixed here:

1. int64 truncation of continuous features. create_disrnn_dataset allocates xs with np.full(..., -1), an int64 array, so float feature columns are truncated toward zero before the later .astype(float). Harmless for the stock integer features (animal_response, rewarded), but it destroys any continuous input — in testing, 24,149 distinct log-reaction-time values collapsed to 7 integers. The wrapper carries a float-safe builder alongside the vendored one.
2. Slow session loading. The per-session df.query() loop re-parses the expression and re-scans the frame for every session, which dominates dataset construction for large cohorts. Fixed in the vendored copy (see the closed PR #29).

If you still depend on this

Pin a SHA or a released version and you are fine indefinitely. If you want the fixes above, copy disrnn_dataset.py from the wrapper rather than reviving this package.

Usage

Creating a dataset

• Obtain a list of NWB files you wish to fit the model to

import aind_dynamic_foraging_multisession_analysis.multisession_load as ms_load
import aind_disrnn_utils as dl

nwbs, df_trials = ms_load.make_multisession_trials_df(nwb_files)
dataset = dl.create_disrnn_dataset(df_trials)

• You don't need to use make_multisession_trials_df, but the trials data frame does need to have a column "ses_idx" that splits trials into sessions.

Predefined datasets

This Code Ocean Capsule can be used for loading a list of sessions and saving the result as a dataframe: Code Ocean Capsule

The resulting data assets can be used like:

import pandas as pd
import aind_disrnn_utils.data_loader as dl
df = pd.read_csv('/data/disrnn_dataset_774212/disrnn_dataset.csv')
dataset = dl.create_disrrn_dataset(df)

The datasets can be combined to fit easily:

import pandas as pd
import aind_disrnn_utils.data_loader as dl

mice = [77412, 779531, 781173, 781162, 778077]
dfs = []
for mouse in mice:
    dfs.append(pd.read_csv('/data/disrnn_dataset_{}/disrnn_dataset.csv'.format(mouse)))
df = pd.concat(dfs)
dataset = dl.create_disrrn_dataset(df)

Saving results

After fitting the network, you can add the latent states and predictions back into the dataframe of trials:

df_trials = dl.add_model_results(df_trials, network_states.__array__(), yhat, ignore_policy=ignore_policy)

Installation

To install the software from PyPi

pip install aind-disrnn-utils

To use the software, in the root directory, run

pip install -e .

To develop the code, run

pip install -e . --group dev

Note: --group flag is available only in pip versions >=25.1

Alternatively, if using uv, run

uv sync

Level of Support

• Unsupported (retired 2026-09-01). This repository is archived and read-only. Issues and pull requests are closed; existing installs and SHA pins continue to work. See the notice at the top of this README for where the code moved.