('%s_vs_deepwalk_distcorr' % method):
compute_distance_correlation(topology_weights[method],
topology_weights['deepwalk'],
tokens, unattacked_indx)
})
exp_result.update({
('%s_vs_glove_distcorr' % method):
compute_distance_correlation(topology_weights[method],
topology_weights['glove'], tokens,
unattacked_indx)
})
results.append(exp_result)
# Compute cooccurrences
(cooccurrence_list, index_vocab_list, vocab_index_lookup,
tokenized_cooccurrences) = count_cooccurrences(walks_prime, 5)
# Compute mrr curves
for method in topology_weights:
normalize_rows = False
normalize_cols = False
exp_result.update({
('%s_mrr_curve_full' % method): [
compute_mrr_curve(topology_weights[method],
tokenized_cooccurrences,
tokens,
normalize_rows=normalize_rows,
normalize_cols=normalize_cols)
]
})
exp_result.update({
def __init__(self,
vector_size,
vocab_filename=None,
covariate_size=0,
random_seed=12345,
init_weight_dir=None,
random_walks=None,
covariate_data=None,
window_size=5):
"""Initializes the data reading and model variables.
Args:
vector_size: size of the word vectors.
vocab_filename: filename for getting word tokens.
covariate_size: size of the covariate embedding dimension
random_seed: seed the initialization generator
init_weight_dir: directory to pull initial weights from. defaults to a
uniform initializer if none.
random_walks: a list of tokenized sentences
covariate_data: a keyed list of float lists, where each key identifies a
token in the corpus, and each float list is a row of covariate data
window_size: window size to use for cooccurrence counting, if needed
Returns: (none)
"""
print('setting up basic stuff...')
# Get word tokens
self._vector_size = vector_size
self._covariate_size = covariate_size
self._tokens = []
self._vocab_index_lookup = None
if vocab_filename:
with open(vocab_filename, 'r') as f:
for line in f:
self._tokens.append(line.split()[0])
self._vocab_index_lookup = dict(
zip(self._tokens, list(range(len(self._tokens)))))
self._cooccurrences = None
self._cooccurrence_dict = None
print('loading or computing co-occurrences...')
if random_walks:
(self._cooccurrences, self._tokens, self._vocab_index_lookup,
self._cooccurrence_dict) = glove_util.count_cooccurrences(
random_walks, window_size, self._vocab_index_lookup)
self._vocab_size = len(self._tokens)
# Get covariate data
print('setting other placeholders...')
if covariate_data is not None:
self._covariate_data = np.array([covariate_data[t] for t in self._tokens])
# Placeholders for parameter tensors and other trackers
io_dict = {'input': None, 'outpt': None}
self._word = copy.deepcopy(io_dict)
self._bias = copy.deepcopy(io_dict)
self._iter = 0
self._sum_cost = 0
self._sum_adv_cost_g = 0
self._sum_adv_cost_d = 0
self._random_seed = random_seed
self._init_weight_dir = init_weight_dir
# Pointers to variables needed for covariate model
self._cvrt = copy.deepcopy(io_dict)
self._cvrt_transformation = copy.deepcopy(io_dict)
# Initialize the cooccurrence read format
self._cooccurrence_fmt = 'iid'
self._cooccurrence_fmt_length = struct.calcsize(self._cooccurrence_fmt)
self._struct_unpack = struct.Struct(self._cooccurrence_fmt).unpack_from