PsiZ Datasets

Author: Brett D. Roads

Overview

A small number of PsiZ-compatible datasets are hosted at the OSF repository psiz-datasets. These datasets can be obtained in raw form—where each trial sequence is stored as a separate JSON file inside a zipped directory—or as pre-formatted TensorFlow Datasets.

Pre-formatted Datasets

Datasets can be loaded as pre-formatted TensorFlow Datasets using the psiz-datasets Python package, which can be installed via pip install psiz-datasets. See the psiz-datasets README for additional package information. See the sections below for instructions on loading specific datasets.

Naming Convention

Datasets follow a two part naming convention. The first part refers to the name of the stimulus dataset. The second part refers to the human behavior collected while using the stimulus dataset. For example, birds16_rank2019 refers to the stimulus dataset birds16 and a set of collected behavior refered to as rank2019.

Stimuli Pointers

Rather than passing around filenames; the data refers to specific stimuli using indices. If using the raw data files (zipped directory from OSF), the index mappings can be found in stimuli.txt. If using the the pre-formatted TensorFlow Datasets, the index mappings are accessible by using with_info=True to return an additional info object when loading the dataset. The index mapping dictionary is located at info.metadata['stimuli'].

Data Timesteps

The TensorFlow Datasets can be loaded with or without a timestep axis by appending /with_timestep or /without_timestep to the dataset name when using tfds.load (see examples below). By default, the dataset is loaded with a timestep axis. If loaded without a timestep axis, the timestep axis is simply unrolled into the batch axis.

Structuring for Training

Your application will likely require that the loaded datasets be structured into inputs, targets, and sample weights for training. This is easily achieved using tf.data.Dataset.map function. See birds16_rank2019 below for an example.

birds16_rank2019

Raw Data Files

To load the pre-formatted TensorFlow Dataset:

import tensorflow_datasets as tfds
import psiz_datasets.birds16_rank2019
ds, info = tfds.load(
    'birds16_rank2019/with_timestep',
    split="train",
    with_info=True
)

Example of dataset formatting:

def format_data_for_training(sample):
    """Format sample as (x, y, w) tuple."""
    x = {
        'given2rank1_stimulus_set': sample['given2rank1_stimulus_set'],
        'given8rank2_stimulus_set': sample['given8rank2_stimulus_set'],
    }
    y = {
        'given2rank1_outcome': sample['given2rank1_outcome'],
        'given8rank2_outcome': sample['given8rank2_outcome'],
    }
    w = {
        'given2rank1_outcome': sample['given2rank1_sample_weight'],
        'given8rank2_outcome': sample['given8rank2_sample_weight'],
    }
    return (x, y, w)

tfds_all = tfds_all.map(
    lambda sample: format_data_for_training(sample)
)

ilsvrc2012_val_hsj

Raw Data Files

To load the pre-formatted TensorFlow Dataset:

import tensorflow_datasets as tfds
import psiz_datasets.ilsvrc2012_val_hsj
ds, info = tfds.load(
    'ilsvrc2012_val_hsj/with_timestep',
    split="train",
    with_info=True
)

skin_lesions2018_rank2018

Raw Data Files

To load the pre-formatted TensorFlow Dataset:

import tensorflow_datasets as tfds
import psiz_datasets.ilsvrc2012_val_hsj
ds, info = tfds.load(
    'skin_lesions2018_rank2018/with_timestep',
    split="train",
    with_info=True
)