def _decode_sample(self, x, topk,
average_encoders = False,
selected_encoder = -1,
visualization = False,
rows_visualization = None,
cols_visualization = None):
attns, rx = self._get_sample_attns(x, average_encoders = average_encoders,
selected_encoder = selected_encoder)
if visualization: Visualization.visualize_attentions(attns, 16, 16, rows_visualization, cols_visualization)
attn = self._get_sample_attn(attns)
x_lines = x.split(" . ")
lx_lines = len(x_lines)
sent_pad_required = max(0, self.document_max_sents - lx_lines)
if visualization: Visualization.visualize_attentions(attn, 16, 16, 1, 1)
attn = attn[sent_pad_required:, sent_pad_required:]
if visualization: Visualization.visualize_attentions(attn, 16, 16, 1, 1)
sentence_attn = attn.sum(axis = 0) / attn.shape[0]
topk_sentences = sorted(np.argsort(sentence_attn)[::-1][:topk])
if visualization: Visualization.visualize_attentions(sentence_attn, 16, 16, 1, 1)
summary = [x_lines[i] for i in topk_sentences]
return " . ".join(summary)
l_words = s.split()
print("Pred: %s" % (pred))
print("Truth: %s" % (y_dv[inp]))
pad = 0
while len(l_words) < max_words:
l_words.insert(0, "<pad>")
pad += 1
l_words = l_words[pad:]
attn_i = attn_i[:, pad:, pad:]
attn_i = attn_i[2]
#attn_i = attn_i.sum(axis=0)
output_file = "sentiment_att.pdf"
Visualization.visualize_attentions(attn_i,
30,
30,
rows=1,
columns=1,
ticks=l_words,
output_file=output_file,
save=True)
inp = int(input())
"""
attn_i = attns[inp]
l_words = x_dv[inp].split()
print("Pred: %s, True: %s" % (preds[inp], truths[inp]))
pad = 0
while len(l_words) < max_words:
l_words.insert(0, "<pad>")
pad += 1
l_words = l_words[pad:]
attn_i = attn_i[:, pad:, pad:]
#attn_i = attn_i.sum(axis=0)