max_sent_size = min(max([len(s) for s in sents_idxs]), max_sent_size_overall)
sents_idxs_padded = [
s + [padding_idx] *
(max_sent_size - len(s)) if len(s) < max_sent_size else s[:max_sent_size]
for s in sents_idxs
]
reshaped_sentences = np.reshape(np.array(sents_idxs_padded),
(1, len(sents), max_sent_size))
reshaped_sentences_tensor = _to_tensor(
reshaped_sentences,
dtype='float32') # a layer, unlike a model, requires tf tensor as input
print('== attention over words ==')
sents_att_coeffs = TimeDistributed(get_sent_att_coeffs)(
reshaped_sentences_tensor)
word_coeffs = sents_att_coeffs.eval(session=sess)
word_coeffs = np.reshape(word_coeffs, (len(sents), max_sent_size))
doc_att_tensor = get_doc_attention_coeffs(reshaped_sentences_tensor)
doc_att = doc_att_tensor.eval(session=sess)[0]
res_tensor = han(reshaped_sentences_tensor)
res = res_tensor.eval(session=sess)
print(doc_att)
my_wcs = []
my_values_array = []
my_keys_array = []
for my_idx, wc in enumerate(word_coeffs):
my_keys = sents_tokenized[my_idx]
my_values = [round(elt, 2) for elt in wc.tolist()[:len(my_keys)]]
my_values_array.append(my_values[:len(my_values)])
my_keys_array.append(my_keys)
