multigeom.fit.tvsd

Data loaders for the THINGS Ventral-Stream Spiking Dataset (TVSD) of Papale et al. (extra [paper]: needs h5py).

Two TVSD files (MATLAB v7.3 / HDF5) are read:

  monkeyN_THINGS_normMUA.mat
    - `train_MUA`  stored as [n_train_images, n_channels] in HDF5 order
    - `test_MUA`   stored as [n_test_images,  n_channels], trial-averaged
    - `oracle`     [n_channels] split-half reliability
  things_imgs.mat
    - `train_imgs/things_path`, `test_imgs/things_path`: object references to
      the THINGS image paths, in the same order as the response rows.

Functions

existing_image_mask(paths, images_root)	Boolean mask of the paths that exist under images_root.
load_images(paths, images_root[, transform, ...])	The images at paths (under images_root) as one tensor [N, 1, 64, 64] in [-1, 1], in the order of paths.
load_tvsd(area[, n_neurons, tvsd_dir])	Load TVSD responses + image paths and select the top-n_neurons channels in area by oracle reliability.

Classes

TVSDDataset(*args, **kwargs)

Pairs THINGS images with macaque normMUA responses.