def __init__(self, path_train, path_dev, path_test, aug_num,
distance_train, distance_dev, distance_test):
self.path_train = path_train
self.path_dev = path_dev
self.path_test = path_test
self.aug_num = aug_num
self.old_train = load_dataset_input(self.path_train)
self.old_dev = load_dataset_input(self.path_dev)
self.old_test = load_dataset_input(self.path_test)
self.train_train = load_pickle(distance_train)
self.dev_train = load_pickle(distance_dev)
self.test_train = load_pickle(distance_test)
tprint('load is done, begin transfer data to new format')
self._load_new()
tprint('dataset load done')
def pruning_sem_f1_score(target, predict, predicate_correct, predicate_sum,
out_of_pruning, argument2idx):
predict_args = 0
golden_args = 0
correct_args = 0
num_correct = 0
total = 0
for i in range(len(target)):
pred_i = predict[i]
golden_i = target[i]
if golden_i == argument2idx[_PAD_]:
continue
total += 1
if pred_i == argument2idx[_UNK_]:
pred_i = argument2idx['_']
if golden_i == argument2idx[_UNK_]:
golden_i = argument2idx['_']
if pred_i != argument2idx['_']:
predict_args += 1
if golden_i != argument2idx['_']:
golden_args += 1
if golden_i != argument2idx['_'] and pred_i == golden_i:
correct_args += 1
if pred_i == golden_i:
num_correct += 1
P = (correct_args + predicate_correct) / (predict_args + predicate_sum +
1e-13)
R = (correct_args + predicate_correct) / (golden_args + out_of_pruning +
predicate_sum + 1e-13)
NP = correct_args / (predict_args + 1e-13)
NR = correct_args / (golden_args + 1e-13)
F1 = 2 * P * R / (P + R + 1e-13)
NF1 = 2 * NP * NR / (NP + NR + 1e-13)
tprint(
'\teval accurate:{:.2f} predict:{} golden:{} correct:{} P:{:.2f} R:{:.2f} F1:{:.2f} NP:{:.2f} NR:{:.2f} NF1:{:.2f}'
.format(num_correct / total * 100, predict_args,
golden_args + out_of_pruning, correct_args, P * 100, R * 100,
F1 * 100, NP * 100, NR * 100, NF1 * 100))
return (P, R, F1, NP, NR, NF1)
def sem_f1_score(target,
predict,
predicate_correct,
predicate_sum,
argument2idx,
output_to_file=None):
predict_args = 0
golden_args = 0
correct_args = 0
num_correct = 0
total = 0
for i in range(len(target)):
pred_i = predict[i]
golden_i = target[i]
if golden_i == argument2idx[_PAD_]:
continue
total += 1
if pred_i == argument2idx[_UNK_]:
pred_i = argument2idx['_']
if golden_i == argument2idx[_UNK_]:
golden_i = argument2idx['_']
if pred_i != argument2idx['_']:
predict_args += 1
if golden_i != argument2idx['_']:
golden_args += 1
if golden_i != argument2idx['_'] and pred_i == golden_i:
correct_args += 1
if pred_i == golden_i:
num_correct += 1
P = (correct_args + predicate_correct) / (predict_args + predicate_sum +
1e-13)
R = (correct_args + predicate_correct) / (golden_args + predicate_sum +
1e-13)
NP = correct_args / (predict_args + 1e-13)
NR = correct_args / (golden_args + 1e-13)
F1 = 2 * P * R / (P + R + 1e-13)
NF1 = 2 * NP * NR / (NP + NR + 1e-13)
outs = tprint(
'eval accurate:{:.2f} predict:{} golden:{} correct:{} P:{:.2f} R:{:.2f} F1:{:.2f} NP:{:.2f} NR:{:.2f} NF1:{:.2f}'
.format(num_correct / total * 100, predict_args, golden_args,
correct_args, P * 100, R * 100, F1 * 100, NP * 100, NR * 100,
NF1 * 100))
if output_to_file is not None:
with open(output_to_file, 'a') as files:
files.write(outs)
return (P, R, F1, NP, NR, NF1)
def eval_train_batch(epoch,
batch_i,
loss,
golden_batch,
predict_batch,
argument2idx,
output_to_file=None):
predict_args = 0
golden_args = 0
correct_args = 0
num_correct = 0
batch_total = 0
for i in range(len(golden_batch)):
pred_i = predict_batch[i]
golden_i = golden_batch[i]
if golden_i == argument2idx[_PAD_]:
continue
batch_total += 1
if pred_i == argument2idx[_UNK_]:
pred_i = argument2idx['_']
if golden_i == argument2idx[_UNK_]:
golden_i = argument2idx['_']
if pred_i != argument2idx['_']:
predict_args += 1
if golden_i != argument2idx['_']:
golden_args += 1
if golden_i != argument2idx['_'] and pred_i == golden_i:
correct_args += 1
if pred_i == golden_i:
num_correct += 1
recall = correct_args / (golden_args + 1e-13)
precision = correct_args / (predict_args + 1e-13)
F = 2 * recall * precision / (recall + precision + 1e-13)
outs = tprint(
'epoch {} batch {} loss:{:4f} accurate:{:.2f} precision:{} recall:{} F:{}'
.format(epoch, batch_i, loss, num_correct / batch_total * 100,
precision, recall, F))
if output_to_file is not None:
with open(output_to_file, 'a') as out_file:
out_file.write(outs)
return correct_args, golden_args, predict_args
res_tmp = []
for i,e in enumerate(ele):
res_tmp.append([i,e])
res_tmp = sorted(res_tmp,key = lambda x:x[1])[:300]
train_tmp.append([ele[0] for ele in res_tmp])
pickle_save(train_tmp,result_data_path )
if __name__ == '__main__':
target_name = sys.argv[1]
if target_name == 'train':
train = get_sentences('./data/train.txt')
# train file is so large, we need to split them to 18 file for process
target_2 = get_sentences('./data/train.txt')
cur_idx = 0
while cur_idx < 179014:
tprint('begin processing %d/18' % (cur_idx // 10000 + 1))
target = target_2[cur_idx:cur_idx+10000]
result = get_max_similar(target, train)
pickle_save(result, './temp/train-train-%d-result.bin' % (cur_idx))
analyse_train_file('./temp/train-train-%d-result.bin' % (cur_idx), './temp/train-train-%d-part.bin' % (cur_idx))
cur_idx += 10000
total = []
for i in range(18):
total.extend(pickle_load('./temp/train-train-%d-part.bin' % (i)))
pickle_save(total, './temp/train_train.bin')
elif target_name == 'dev':
target = get_sentences('./data/dev.txt')
train = get_sentences('./data/train.txt')
result = get_max_similar(target, train)
pickle_save(result, './temp/dev-train-result.bin')