Single predictor models¶

Single predictor models investigating the effect of a range of lower- and higher-level visual and auditory predictors. Analyses are run using Neuroscout’s pyNS interaface for the Neuroscout API. Identical analyses are specified for each dataset / task combination.

from collections import defaultdict
from pyns import Neuroscout
from pathlib import Path
from create import create_models

import sys
sys.path.append("..") 
from utils import dump_collection, load_collection

%matplotlib inline

/Users/rr48396/opt/anaconda3/lib/python3.7/site-packages/nilearn/datasets/__init__.py:96: FutureWarning: Fetchers from the nilearn.datasets module will be updated in version 0.9 to return python strings instead of bytes and Pandas dataframes instead of Numpy arrays.
  "Numpy arrays.", FutureWarning)

Note: To push models to Neuroscout, authenticate pyNS by setting `NEUROSCOUT_USER` and `NEUROSCOUT_PASSWORD` environment variables, or directly passing your auth details to `Neuroscout()`. Google SSO logins are not currently supported.

api = Neuroscout()

Define predictors and confounds¶

predictors = ['speech', 'rms', 'text', 'brightness', 'shot_change', 'landscape', 'building', 'tool']
confounds = [
    'a_comp_cor_00', 'a_comp_cor_01', 'a_comp_cor_02', 'a_comp_cor_03', 'a_comp_cor_04', 'a_comp_cor_05', 
    'trans_x', 'trans_y', 'trans_z', 'rot_x', 'rot_y', 'rot_z'
]
datasets = api.datasets.get() # Get all Neuroscout datasets

Create models¶

The following will create a neuroscout analysis for every dataset/task that and predictor combination.

Note: If a predictor is unavailable for a task, it will be skipped. Examples include visual predictor for auditory narratives, and "shot_change" for NNDb, due to issues with large stimuli in their API

Uncomment the follwing line to re-execute and re-create models for the logged in acccount.

filename = Path('models') / 'single_predictor.json'

# single_models = defaultdict(list)
# for pred in predictors:
#     single_models[pred] = create_models(name=pred, predictors=[pred], confounds=confounds, datasets=datasets)

# Save models out to disk:
# dump_collection(single_models, filename)

# Load models from cache
single_models = load_collection(filename)

# Three models for "building" Clarifai visual predictor
single_models['building'][0:3]

[{'dataset': 'Budapest',
  'task': 'movie',
  'hash_id': 'A13DD',
  'analysis': <Analysis hash_id=A13DD name=building dataset_id=27>},
 {'dataset': 'LearningTemporalStructure',
  'task': 'movie',
  'hash_id': 'AVrDK',
  'analysis': <Analysis hash_id=AVrDK name=building dataset_id=19>},
 {'dataset': 'Life',
  'task': 'life',
  'hash_id': 'Ar6D0',
  'analysis': <Analysis hash_id=Ar6D0 name=building dataset_id=9>}]

We can inspect the BIDS StatsModel generate for any analysis.

For example: building for the Budapest dataset:

# BIDS StatsModel generated for a single analysis. 
# In this case, "studyforrest" for the "building" feature
single_models['building'][-1]['analysis'].model

{'Input': {'Run': [1, 2, 3, 4, 5, 6, 7, 8],
  'Subject': ['10',
   '19',
   '04',
   '03',
   '01',
   '18',
   '15',
   '09',
   '16',
   '14',
   '05',
   '20',
   '06'],
  'Task': 'movie'},
 'Name': 'building',
 'Steps': [{'Contrasts': [],
   'DummyContrasts': {'Conditions': ['building'], 'Type': 't'},
   'Level': 'Run',
   'Model': {'X': ['a_comp_cor_00',
     'a_comp_cor_01',
     'a_comp_cor_02',
     'a_comp_cor_03',
     'a_comp_cor_04',
     'a_comp_cor_05',
     'trans_x',
     'trans_y',
     'trans_z',
     'rot_x',
     'rot_y',
     'rot_z',
     'building']},
   'Transformations': [{'Input': ['building'], 'Name': 'Convolve'}]},
  {'DummyContrasts': {'Type': 'FEMA'}, 'Level': 'Subject'},
  {'DummyContrasts': {'Type': 't'}, 'Level': 'Dataset'}]}

Generate reports¶

For every individual analysis created, we can generate reports to inspect the design matrix

analysis = single_models['building'][0]['analysis']
analysis.generate_report(run_id=analysis.runs[0]) # Only generate for a single example run, to save time

{'generated_at': '2022-03-28T07:2',
 'result': None,
 'sampling_rate': None,
 'scale': False,
 'status': 'PENDING',
 'traceback': None,
 'warnings': []}

analysis.plot_report(plot_type='design_matrix_plot')

analysis.plot_report(plot_type='design_matrix_corrplot')

Compile models¶

The following will “compile” every created model, validating the model and producing an executable bundle

for pred, models in single_models.items():
    for model in models:
        # If analysis is still in "DRAFT" model, compile it.
        if model['analysis'].get_status()['status'] == 'DRAFT':
            model['analysis'].compile()

Model Execution¶

Models can be executed using Neuroscout-CLI. For more information see the Neuroscout documentation.

For example, to re-run the following model using Docker:

single_models['building'][0]

{'dataset': 'Budapest',
 'task': 'movie',
 'hash_id': 'A13DD',
 'analysis': <Analysis hash_id=A13DD name=building dataset_id=27>}

Use the following command:

docker run --rm -it -v /home/results/:/out neuroscout/neuroscout-cli run --force-upload A13DD /out

This will download the bundle for M8LYl, the necessary preprocessed imaging data from the Budapest dataset, run the analysis workflow and upload results to NeuroVault (even if an upload already exists).

Note: Replace `/home/results` with a real directory in your local environment