mila-udem / fuel

    return synsets, cost_matrix, raw_valid_groundtruth

def read_metadata_mat_file(meta_mat):

    """Read ILSVRC2010 metadata from the distributed MAT file.

    Parameters

    ----------

    meta_mat : str or file-like object

        The filename or file-handle for `meta.mat` from the

        ILSVRC2010 development kit.

    Returns

    -------

    synsets : ndarray, 1-dimensional, compound dtype

        A table containing ILSVRC2010 metadata for the "synonym sets"

        or "synsets" that comprise the classes and superclasses,

        including the following fields:

         * `ILSVRC2010_ID`: the integer ID used in the original

           competition data.

         * `WNID`: A string identifier that uniquely identifies

           a synset in ImageNet and WordNet.

         * `wordnet_height`: The length of the longest path to

           a leaf node in the FULL ImageNet/WordNet hierarchy

           (leaf nodes in the FULL ImageNet/WordNet hierarchy

           have `wordnet_height` 0).

         * `gloss`: A string representation of an English

           textual description of the concept represented by

           this synset.

         * `num_children`: The number of children in the hierarchy

           for this synset.

         * `words`: A string representation, comma separated,

           of different synoym words or phrases for the concept

           represented by this synset.

         * `children`: A vector of `ILSVRC2010_ID`s of children

           of this synset, padded with -1. Note that these refer

           to `ILSVRC2010_ID`s from the original data and *not*

           the zero-based index in the table.

         * `num_train_images`: The number of training images for

           this synset.

    cost_matrix : ndarray, 2-dimensional, uint8

        A 1000x1000 matrix containing the precomputed pairwise

        cost (based on distance in the hierarchy) for all

        low-level synsets (i.e. the thousand possible output

        classes with training data associated).

    """

    mat = loadmat(meta_mat, squeeze_me=True)

    synsets = mat['synsets']

    cost_matrix = mat['cost_matrix']

    new_dtype = numpy.dtype([

        ('ILSVRC2010_ID', numpy.int16),

        ('WNID', ('S', max(map(len, synsets['WNID'])))),

        ('wordnet_height', numpy.int8),

        ('gloss', ('S', max(map(len, synsets['gloss'])))),

        ('num_children', numpy.int8),

        ('words', ('S', max(map(len, synsets['words'])))),

        ('children', (numpy.int8, max(synsets['num_children']))),

        ('num_train_images', numpy.uint16)

    ])

    new_synsets = numpy.empty(synsets.shape, dtype=new_dtype)

    for attr in ['ILSVRC2010_ID', 'WNID', 'wordnet_height', 'gloss',

                 'num_children', 'words', 'num_train_images']:

        new_synsets[attr] = synsets[attr]

    children = [numpy.atleast_1d(ch) for ch in synsets['children']]

    padded_children = [

        numpy.concatenate((c,

                           -numpy.ones(new_dtype['children'].shape[0] - len(c),

                                       dtype=numpy.int16)))

        for c in children

    ]

    new_synsets['children'] = padded_children

    return new_synsets, cost_matrix

def extract_patch_images(f, which_set):

    return synsets, raw_valid_groundtruth

def read_metadata_mat_file(meta_mat):

    """Read ILSVRC2012 metadata from the distributed MAT file.

    Parameters

    ----------

    meta_mat : str or file-like object

        The filename or file-handle for `meta.mat` from the

        ILSVRC2012 development kit.

    Returns

    -------

    synsets : ndarray, 1-dimensional, compound dtype

        A table containing ILSVRC2012 metadata for the "synonym sets"

        or "synsets" that comprise the classes and superclasses,

        including the following fields:

         * `ILSVRC2012_ID`: the integer ID used in the original

           competition data.

         * `WNID`: A string identifier that uniquely identifies

           a synset in ImageNet and WordNet.

         * `wordnet_height`: The length of the longest path to

           a leaf node in the FULL ImageNet/WordNet hierarchy

           (leaf nodes in the FULL ImageNet/WordNet hierarchy

           have `wordnet_height` 0).

         * `gloss`: A string representation of an English

           textual description of the concept represented by

           this synset.

         * `num_children`: The number of children in the hierarchy

           for this synset.

         * `words`: A string representation, comma separated,

           of different synoym words or phrases for the concept

           represented by this synset.

         * `children`: A vector of `ILSVRC2012_ID`s of children

           of this synset, padded with -1. Note that these refer

           to `ILSVRC2012_ID`s from the original data and *not*

           the zero-based index in the table.

         * `num_train_images`: The number of training images for

           this synset.

    """

    mat = loadmat(meta_mat, squeeze_me=True)

    synsets = mat['synsets']

    new_dtype = numpy.dtype([

        ('ILSVRC2012_ID', numpy.int16),

        ('WNID', ('S', max(map(len, synsets['WNID'])))),

        ('wordnet_height', numpy.int8),

        ('gloss', ('S', max(map(len, synsets['gloss'])))),

        ('num_children', numpy.int8),

        ('words', ('S', max(map(len, synsets['words'])))),

        ('children', (numpy.int8, max(synsets['num_children']))),

        ('num_train_images', numpy.uint16)

    ])

    new_synsets = numpy.empty(synsets.shape, dtype=new_dtype)

    for attr in ['ILSVRC2012_ID', 'WNID', 'wordnet_height', 'gloss',

                 'num_children', 'words', 'num_train_images']:

        new_synsets[attr] = synsets[attr]

    children = [numpy.atleast_1d(ch) for ch in synsets['children']]

    padded_children = [

        numpy.concatenate((c,

                           -numpy.ones(new_dtype['children'].shape[0] - len(c),

                                       dtype=numpy.int16)))

        for c in children

    ]

    new_synsets['children'] = padded_children

    return new_synsets