def compute_confuse_matrix(fname, classes):
"""
Give a file, compute confuse matrix of y_true and y_pred.
"""
print('im in')
y_true = []
with codecs.open(fname, 'r', 'utf8') as f:
for line in f:
line = line.strip().split('\t')[-1]
y_true.append(line)
checkpoint_dir = "output/self_attention/multi_attention_0802/"
pred_path = "tmp/eval_y_self_attention.txt"
if os.path.exists(checkpoint_dir + 'config.pkl'):
config = pickle.load(open(checkpoint_dir + 'config.pkl', 'rb'))
else:
config = Config()
config.mode = 'inference'
word2id, id2word = read_vocab(config.word_vocab_file)
tag2id, id2tag = read_vocab(config.tag_vocab_file)
with tf.Session(config=get_config_proto(
log_device_placement=False)) as sess:
model = get_model(config.model, config, sess)
model.build()
model.restore_model(checkpoint_dir)
y_pred = infer_file(model, word2id, id2tag, fname, pred_path)
cmatrix = confusion_matrix(y_true, y_pred, classes)
print(cmatrix)
correct = [x == y for x, y in list(zip(y_true, y_pred))]
print(correct.count(True) / len(correct))
return cmatrix
def main(_argv):
"""Program entry point.
"""
if not FLAGS.model_dir:
raise Exception("you must provide model directory")
with tf.Session(config=get_config_proto(False)) as sess:
# read config from pickle
model_dir = Path(FLAGS.model_dir)
config = pickle.load((model_dir / 'config.pkl').open('rb'))
config.mode = "inference"
model = NERModel(sess, config)
model.build()
# using best dev model
best_dev_dir = model_dir / "best_dev"
model.restore_model(best_dev_dir)
export_path = FLAGS.export
builder = saved_model_builder.SavedModelBuilder(export_path)
prediction_input_tokens = utils.build_tensor_info(model.source_tokens)
prediction_input_length = utils.build_tensor_info(model.source_length)
prediction_input_segments = utils.build_tensor_info(model.segment_tokens)
prediction_output_sequence = utils.build_tensor_info(model.viterbi_sequence)
prediction_output_score = utils.build_tensor_info(model.viterbi_score)
prediction_signature = signature_def_utils.build_signature_def(
inputs={'query': prediction_input_tokens, 'query_length': prediction_input_length, "segement": prediction_input_segments},
outputs={"sequence": prediction_output_sequence,"score":prediction_output_score},
method_name=signature_constants.PREDICT_METHOD_NAME)
legacy_init_op = tf.group(
tf.tables_initializer(), name='legacy_init_op')
builder.add_meta_graph_and_variables(
sess, [tag_constants.SERVING],
signature_def_map={
'predict':
prediction_signature,
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
prediction_signature,
},
legacy_init_op=legacy_init_op,
clear_devices=True)
builder.save()
print('Done exporting!')
def main():
checkpoint_dir = "output/self_attention/multi_attention_0802/"
# inference 用的还是文件中的config，并不是当初的config了。应该读取check_point中的config.pkl
# TODO: read config from checkpoint/config.pkl
if os.path.exists(checkpoint_dir + 'config.pkl'):
config = pickle.load(open(checkpoint_dir+'config.pkl', 'rb'))
else:
config = Config()
config.mode = 'inference'
# 每次训练word_vocab和tag_vocab都会变化，而inference的时候用的是当初训练该模型时的词表；所以最好把词表定下来。
# TODO：data_utils::fix vocab
word2id, id2word = read_vocab(config.word_vocab_file)
tag2id, id2tag = read_vocab(config.tag_vocab_file)
with tf.Session(config=get_config_proto(log_device_placement=False)) as sess:
model = get_model(config.model, config, sess)
model.build()
model.restore_model(checkpoint_dir)
infer_cmd(model, word2id, id2tag)
config = pickle.load(
open(os.path.join(checkpoint_dir + "config.pkl"), 'rb'))
else:
config.checkpoint_dir = checkpoint_dir
config.vocab_file = os.path.join(checkpoint_dir, "word.vocab")
config.encode_cell_type = 'COUPLED_LSTM'
#config.external_word_file = external_words_fname
#config.pretrained_embedding_file = os.path.join(DATA_DIR,"wiki_100.utf8")
config.num_tags = len(tag_vocab)
config.segment_vocab_file = os.path.join(checkpoint_dir, "seg.vocab")
config.tag_vocab_file = os.path.join(checkpoint_dir, "tag.vocab")
pickle.dump(
config,
open(os.path.join(config.checkpoint_dir, "config.pkl"), 'wb'))
with tf.Session(config=get_config_proto(
log_device_placement=False)) as sess:
model = NERModel(sess, config)
model.build()
model.init()
print("Config:")
for k, v in config.__dict__.items():
print(k, '-', v, sep='\t')
try:
model.restore_model()
print("restored model from checkpoint dir")
except:
print("create with fresh parameters")
pass
def main():
parser = argparse.ArgumentParser()
add_argument(parser)
args = parser.parse_args()
config = Config()
train_data = read_data(config.train_data_files, config.model)
# 试试对负样本进行降采样
# train_data = sample(train_data)
eval_data = read_data(config.eval_data_files, config.model)
# train_data_sen = read_data_sen("data/data_tech.train")
# eval_data_sen = read_data_sen("data/data_tech.eval")
# 这里使用了预训练的词向量的词表作为了模型的词表
create_vocab_from_pretrained_w2v(config.w2v_path, config.word_vocab_file)
create_tag_vocab_from_data(train_data, config.tag_vocab_file)
word2id, id2word = read_vocab(config.word_vocab_file)
tag2id, id2tag = read_vocab(config.tag_vocab_file)
# convert word into ids
train_data = convert_dataset(train_data, word2id, tag2id,
config.sentence_length, config.num_classes,
config.model)
# train_data_sen = convert_dataset_sen(train_data_sen, word2id, tag2id, config.num_classes, one_hot_label=True)
print(train_data[0])
eval_data = convert_dataset(eval_data, word2id, tag2id,
config.sentence_length, config.num_classes,
config.model)
# eval_data_sen = convert_dataset_sen(eval_data_sen, word2id, tag2id, config.num_classes, one_hot_label=True)
print("train_data size: {0}".format(len(train_data)))
if os.path.exists(os.path.join(config.checkpoint_dir, "config.pkl")):
config = pickle.load(
open(os.path.join(config.checkpoint_dir, "config.pkl"), 'rb'))
else:
pickle.dump(
config,
open(os.path.join(config.checkpoint_dir, "config.pkl"), 'wb'))
with tf.Session(config=get_config_proto(
log_device_placement=False)) as sess:
model = get_model(config.model, config, sess)
model.build()
model.init()
batch_manager = Batch_self_attention(train_data, config.batch_size)
batch_manager_eval = Batch_self_attention(eval_data, config.batch_size)
# batch_manager = Batch(train_data, config.batch_size)
# batch_manager_eval = Batch(eval_data, config.batch_size)
epoches = config.epoch
max_acc = 0
for i in range(epoches):
for batch in batch_manager.next_batch():
# print(batch)
loss, accuracy, global_step = model.train_one_step(*zip(
*batch))
# key_shape, query_shape = model.test(*zip(*batch))
# print(key_shape, query_shape)
# break
train_accuracy = evaluate(model, batch_manager)
eval_accuracy = evaluate(model, batch_manager_eval)
# train_accuracy = evaluate_attention(model, train_data_sen, id2tag)
# eval_accuracy = evaluate_attention(model, eval_data_sen, id2tag)
print("epoch - {0} step - {1} loss - {2} train_accuracy - {3} eval_accuracy - {4}"\
.format(i, global_step, loss, train_accuracy, eval_accuracy))
# train_accuracy = evaluate_attention(model, train_data_sen, id2tag)
# eval_accuracy = evaluate_attention(model, eval_data_sen, id2tag)
# print("epoch - {0} step - {1} loss - {2} train_accuracy - {3} eval_accuracy - {4}"\
# .format(i, global_step, loss, train_accuracy, eval_accuracy))
if max_acc < eval_accuracy:
max_acc = eval_accuracy
model.save_model()
