def visualize_attention(wts,x_test_pad,word_to_id,filename):
wts_add = torch.sum(wts,1)
wts_add_np = wts_add.data.numpy()
wts_add_list = wts_add_np.tolist()
id_to_word = {v:k for k,v in word_to_id.items()}
text= []
for test in x_test_pad:
text.append(" ".join([id_to_word.get(i) for i in test]))
createHTML(text, wts_add_list, filename)
print("Attention visualization created for {} samples".format(len(x_test_pad)))
return
def visualize_attention(attention_model, wts, x_test_pad, word_to_id,
word_to_word, y_test, filename):
print(filename, "{} samples".format(len(x_test_pad)))
labels = load_label_data(data_params['labels_csv'])
wts_add = torch.sum(wts, 1)
wts_add_np = wts_add.data.numpy()
wts_add_list = wts_add_np.tolist()
id_to_word = {v: word_to_word[k] for k, v in word_to_id.items()}
result = []
text = []
correct = 0
correct2 = 0
n = 0
for test in x_test_pad:
attention_model.batch_size = 1
attention_model.hidden_state = attention_model.init_hidden()
x_test_var = Variable(torch.from_numpy(test).type(torch.LongTensor))
y_test_pred, _ = attention_model(x_test_var)
# 結果のリストを降順に並べる
m = 0
dic = {}
for x in y_test_pred[0]:
dic[x] = m
m += 1
yy = sorted(dic.items(), reverse=True)
m = 0
pred = []
for y in yy:
m += 1
l = str(y[1])
if len(labels) > 0:
l = labels[y[1]]
pred.append(l)
if m > 5:
break
l = str(y_test[n])
if len(labels) > 0:
l = labels[y_test[n]]
if l == pred[0]:
correct += 1
for r in pred:
if l == r:
correct2 += 1
break
text.append(" ".join([id_to_word.get(i) for i in test]))
result.append([l, pred[0], pred])
n += 1
# print(text[0])
# 20個ずつhtmlに出力
m = 20
n = len(result)
for i in range(0, n, m):
j = i + m
createHTML(result[i:j], text[i:j], wts_add_list[i:j],
filename + '_' + str(i + 1) + '_' + str(j) + '.html')
return (correct, correct2, result)
for i in x_test:
np_sentence = np.asarray(x_test[i])
np_labels = np.asarray(y_test[i])
dtest = xgb.DMatrix(np_sentence)
labels = bst.predict(dtest)
#print(labels)
y_pred_roc.extend(labels)
#print(labels)
y_pred.append([item for item in labels])
y_test_prob.append(y_test[i])
#print(y_test[i])
#y_pred_tensor = torch.FloatTensor()
#visualize_attention(y_pred_tensor,,)
text_flat = []
scores_flat = []
for test in corpus.test.words:
text_flat.append(" ".join(test))
embed()
attention_visualization.createHTML(text_flat, y_pred, "res/xgboost.html")
y_pred_roc = np.array([1 if i >= 0.5 else 0 for i in y_pred_roc])
y_test_roc = []
for i in y_test_prob:
y_test_roc.extend(i)
y_test_roc = np.array([1 if i >= 0.5 else 0 for i in y_test_roc])
match_M(y_pred, y_test_prob)
topK(y_pred, y_test_prob)
print(roc_auc_score(y_test_roc, y_pred_roc))
embed()