def runtest():
test_predicts = model_slots.predict([X_test])
testPred = predict_classes(test_predicts)
testPred = [
map(lambda x: idx2labels[x], remove_padding(c, d))
for c, d in zip(testPred, X_test)
]
print conlleval(testPred, testAct, wordsTest,
output_dir + 'current.test.txt')
intentPred = predict_classes(model_intent.predict([X_test])[0])
print "Test Accuracy : " + str(accuracy(intentPred, testL))
def evaluate(self, label_list, seq_len_list, data, epoch=None):
label2tag = {}
for tag, label in self.tag2label.items():
label2tag[label] = tag if label != 0 else label
model_predict = []
for label_, (sent, tag) in zip(label_list, data):
tag_ = [label2tag[label__] for label__ in label_]
sent_res = []
if len(label_) != len(sent):
print(sent)
print(len(label_))
print(tag)
for i in range(len(sent)):
sent_res.append(
[sent[i], tag[i], tag_[i]]
) #sent_res format like : [['中','B-LOC','0'],['国','I-LOC','I-LOC']...]
model_predict.append(sent_res)
epoch_num = str(epoch + 1) if epoch != None else 'test'
# 将模型的预测写到这个文件中，每一轮迭代都要写一个
label_path = os.path.join(self.result_path, 'label_' + epoch_num)
# 将评测结果写到这个文件中，每一轮迭代都要写一个
metric_path = os.path.join(self.result_path,
'result_metric_' + epoch_num)
for _ in conlleval(model_predict, label_path, metric_path):
print(_)
def visualize(args):
score = []
x = [i+1 for i in range(args.epoch_num)]
for eidx in x:
score.append(conlleval(eidx))
plt.plot(x, score, 'b',)
plt.xlabel("epoch_num")
plt.ylabel("accuracy(%)")
plt.title("Accuracy")
plt.show()
def evaluate(data, sess, model, epoch=None):
# seqs_list = []
labels_pred = []
label_references = []
# (sents, tags) = data
label2tag = []
label2tag = {label: tag for tag, label in hp.tag2label.items()}
# for tag, label in hp.tag2label.items():
# label2tag[label] = tag
for seqs, labels, seqs_len in get_batch(data,
hp.batch_size,
hp.vocab_path,
hp.tag2label,
shuffle=False):
_logits, _transition_params = sess.run(
[logits, transition_params],
feed_dict={
model.sent_input: seqs,
model.label: labels,
model.sequence_length: seqs_len
})
# seqs_list.extend(seqs)
label_references.extend(labels)
for logit, seq_len in zip(_logits, seqs_len):
viterbi_seq, _ = tf.contrib.crf.viterbi_decode(
logit[:seq_len], _transition_params)
labels_pred.append(viterbi_seq)
# print(seqs_list)
# print(label_references)
model_pred = []
epoch_num = str(epoch) if epoch != None else 'test'
if not os.path.exists(hp.result_path): os.mkdir(hp.result_path)
with open(hp.result_path + 'results_epoch_' + (epoch_num),
'w',
encoding='utf-8') as fw:
for label_pred, (sent, tag) in zip(labels_pred, data):
fw.write(''.join(sent) + '\n')
fw.write(''.join(tag) + '\n')
tag_pred = [label2tag[i] for i in label_pred]
fw.write(''.join(tag_pred) + '\n')
sent_res = []
if len(label_pred) != len(sent):
print(sent)
print(len(label_pred))
print(len(sent))
for i in range(len(sent)):
sent_res.append([sent[i], tag[i], tag_pred[i]])
model_pred.append(sent_res)
# label_path = os.path.join(hp.result_path, 'label_' + epoch_num)
# metric_path = os.path.join(hp.result_path, 'result_metric_' + epoch_num)
result = conlleval(model_pred)
print(result)
def train(args):
logger = logging.getLogger("hit-cosem-2018")
logger.info('Load data_set and vocab...')
with open(os.path.join(args.vocab_dir, 'vocab.data'), 'rb') as fin:
vocab = pickle.load(fin)
atis_data = AtisDataSet(args.max_len, args.slot_name_file, train_files=args.train_files, test_files=args.test_files)
logger.info('Converting text into ids...')
atis_data.convert_to_ids(vocab)
atis_data.dynamic_padding(vocab.token2id[vocab.pad_token])
train_data, test_data = atis_data.get_numpy_data()
train_data = AtisLoader(train_data)
train_loader = DataLoader(dataset=train_data, batch_size=args.batch_size, shuffle=True)
test_data = AtisLoader(test_data)
test_loader = DataLoader(dataset=test_data, batch_size=args.batch_size, shuffle=False)
model = biLSTM(args, vocab.size(), len(atis_data.slot_names))
optimizer = model.get_optimizer(args.learning_rate, args.embed_learning_rate, args.weight_decay)
loss_fn = torch.nn.CrossEntropyLoss()
score = []
losses = []
for eidx, _ in enumerate(range(args.epoch_num), 1):
for bidx, data in enumerate(train_loader, 1):
optimizer.zero_grad()
sentences, labels = data
if args.has_cuda:
sentences, labels = Variable(sentences).cuda(), Variable(labels).cuda()
else:
sentences, labels = Variable(sentences), Variable(labels)
output = model(sentences)
labels = labels.view(-1)
loss = loss_fn(output, labels)
loss.backward()
optimizer.step()
_, predicted = torch.max(output.data, 1)
predicted = predicted.cpu().numpy()
labels = labels.data.cpu().numpy()
losses.append(loss.data[0])
logger.info('epoch: {} batch: {} loss: {:.4f} acc: {:.4f}'.format(eidx, bidx, loss.data[0],
accuracy(predicted, labels)))
if eidx >= args.begin_epoch:
if args.embed_learning_rate == 0:
args.embed_learning_rate = 2e-4
elif args.embed_learning_rate > 0:
args.embed_learning_rate *= args.lr_decay
if args.embed_learning_rate <= 1e-5:
args.embed_learning_rate = 1e-5
args.learning_rate = args.learning_rate * args.lr_decay
optimizer = model.get_optimizer(args.learning_rate,
args.embed_learning_rate,
args.weight_decay)
logger.info('do eval on test set...')
f = open("./result/result_epoch"+str(eidx)+".txt", 'w', encoding='utf-8')
for data in test_loader:
sentences, labels = data
if args.has_cuda:
sentences, labels = Variable(sentences).cuda(), Variable(labels).cuda()
else:
sentences, labels = Variable(sentences), Variable(labels) # batch_size * max_len
output = model(sentences)
sentences = sentences.data.cpu().numpy().tolist()
sentences = [vocab.recover_from_ids(remove_pad(s, vocab.token2id[vocab.pad_token]))
for s in sentences]
labels = labels.data.cpu().numpy().tolist()
_, predicted = torch.max(output.data, 1)
predicted = predicted.view(-1, args.max_len).cpu().numpy().tolist()
iter = [zip(s,
map(lambda x:atis_data.slot_names[x], labels[i][:len(s)]),
map(lambda x:atis_data.slot_names[x], predicted[i][:len(s)])
) for i, s in enumerate(sentences)]
for it in iter:
for z in it:
z = list(map(str, z))
f.write(' '.join(z)+'\n')
f.close()
score.append(conlleval(eidx))
torch.save(model.state_dict(), args.save_path+"biLSTM_epoch"+str(eidx)+".model")
max_score_eidx = score.index(max(score))+1
logger.info('epoch {} gets max score.'.format(max_score_eidx))
os.system('perl ./eval/conlleval.pl < ./result/result_epoch' + str(max_score_eidx) + '.txt')
x = [i + 1 for i in range(len(losses))]
plt.plot(x, losses, 'r')
plt.xlabel("time_step")
plt.ylabel("loss")
plt.title("CrossEntropyLoss")
plt.show()
def test_lstm(**kwargs):
"""
Wrapper function for training and testing LSTM
:type fold: int
:param fold: fold index of the ATIS dataset, from 0 to 4.
:type lr: float
:param lr: learning rate used (factor for the stochastic gradient).
:type nepochs: int
:param nepochs: maximal number of epochs to run the optimizer.
:type win: int
:param win: number of words in the context window.
:type nhidden: int
:param n_hidden: number of hidden units.
:type emb_dimension: int
:param emb_dimension: dimension of word embedding.
:type verbose: boolean
:param verbose: to print out epoch summary or not to.
:type decay: boolean
:param decay: decay on the learning rate if improvement stop.
:type savemodel: boolean
:param savemodel: save the trained model or not.
:type normal: boolean
:param normal: normalize word embeddings after each update or not.
:type folder: string
:param folder: path to the folder where results will be stored.
"""
# process input arguments
param = {
'experiment': 'standard',
'lr': 0.1,
'verbose': True,
'decay': True,
'win': 3,
'nhidden': 300,
'nhidden2': 300,
'seed': 345,
'emb_dimension': 90,
'nepochs': 40,
'savemodel': False,
'normal': True,
'layer_norm': False,
'minibatch_size': 4978,
'folder': '../result'
}
param_diff = set(kwargs.keys()) - set(param.keys())
if param_diff:
raise KeyError("invalid arguments:" + str(tuple(param_diff)))
param.update(kwargs)
if param['verbose']:
for k, v in param.items():
print("%s: %s" % (k, v))
# create result folder if not exists
check_dir(param['folder'])
# load the dataset
print('... loading the dataset')
train_set, valid_set, test_set, dic = load_data(3)
train_set = list(train_set)
valid_set = list(valid_set)
# Add validation set to train set
for i in range(3):
train_set[i] += valid_set[i]
# create mapping from index to label, and index to word
idx2label = dict((k, v) for v, k in dic['labels2idx'].items())
idx2word = dict((k, v) for v, k in dic['words2idx'].items())
# unpack dataset
train_lex, train_ne, train_y = train_set
test_lex, test_ne, test_y = test_set
n_trainbatches = len(train_lex) // param['minibatch_size']
print("Sentences in train: %d, Words in train: %d" %
(count_of_words_and_sentences(train_lex)))
print("Sentences in test: %d, Words in test: %d" %
(count_of_words_and_sentences(test_lex)))
vocsize = len(dic['words2idx'])
nclasses = len(dic['labels2idx'])
nsentences = len(train_lex)
groundtruth_test = [[idx2label[x] for x in y] for y in test_y]
words_test = [[idx2word[x] for x in w] for w in test_lex]
# instanciate the model
numpy.random.seed(param['seed'])
random.seed(param['seed'])
print('... building the model')
lstm = LSTM(n_hidden=param['nhidden'],
n_hidden2=param['nhidden2'],
n_out=nclasses,
n_emb=vocsize,
dim_emb=param['emb_dimension'],
cwind_size=param['win'],
normal=param['normal'],
layer_norm=param['layer_norm'],
experiment=param['experiment'])
# train with early stopping on validation set
print('... training')
best_f1 = -numpy.inf
param['clr'] = param['lr']
for e in range(param['nepochs']):
# shuffle
shuffle([train_lex, train_ne, train_y], param['seed'])
param['ce'] = e
tic = timeit.default_timer()
for minibatch_index in range(n_trainbatches):
for i in range(minibatch_index * param['minibatch_size'],
(1 + minibatch_index) * param['minibatch_size']):
x = train_lex[i]
y = train_y[i]
res = lstm.train(x, y, param['win'], param['clr'])
predictions_test = [[
idx2label[x] for x in lstm.classify(
numpy.asarray(contextwin(x, param['win'])).astype('int32'))
] for x in test_lex]
# evaluation // compute the accuracy using conlleval.pl
res_test = conlleval(predictions_test, groundtruth_test,
words_test,
param['folder'] + '/current.test.txt',
param['folder'])
if res_test['f1'] > best_f1:
if param['savemodel']:
lstm.save(param['folder'])
best_lstm = copy.deepcopy(lstm)
best_f1 = res_test['f1']
if param['verbose']:
print(
'NEW BEST: epoch %d, minibatch %d/%d, best test F1: %.3f'
% (e, minibatch_index + 1, n_trainbatches,
res_test['f1']))
param['tf1'] = res_test['f1']
param['tp'] = res_test['p']
param['tr'] = res_test['r']
param['be'] = e
os.rename(param['folder'] + '/current.test.txt',
param['folder'] + '/best.test.txt')
else:
if param['verbose']:
print('')
# learning rate decay if no improvement in 10 epochs
if param['decay'] and abs(param['be'] - param['ce']) >= 10:
param['clr'] *= 0.5
print("Decay happened. New Learning Rate:", param['clr'])
lstm = best_lstm
if param['clr'] < 0.00001:
break
print('BEST RESULT: epoch', param['be'], 'best test F1', param['tf1'],
'with the model', param['folder'])
return lstm, dic
i = 0
while (True):
model_slots.fit([X_train], [y_train, l_train],
batch_size=batch_size,
epochs=1,
validation_data=([X_valid], [y_valid, l_valid]))
model_slots.save(output_dir + 'model_slots_intents_' + str(i) +
'.ckpt')
valid_predicts = model_slots.predict([X_valid])
validPred = predict_classes(valid_predicts[0])
validPred = [
map(lambda x: idx2labels[x], remove_padding(c, d))
for c, d in zip(validPred, X_valid)
]
print conlleval(validPred, validAct, wordsValid,
output_dir + 'current.valid.txt')
print "Valid Accuracy : " + str(
accuracy(predict_classes(valid_predicts[1]), validL))
test_predicts = model_slots.predict([X_test])
testPred = predict_classes(test_predicts[0])
testPred = [
map(lambda x: idx2labels[x], remove_padding(c, d))
for c, d in zip(testPred, X_test)
]
print conlleval(testPred, testAct, wordsTest,
output_dir + 'current.test.txt')
print "Test Accuracy : " + str(
accuracy(predict_classes(test_predicts[1]), testL))
def evaluate(self, session, data, id2char, id2tag):
# 先预测再评估
prediction = self.predict(session, data, id2char, id2tag)
return utils.conlleval(prediction)
