def test(args):
from dataio.task_data import TaskData
from predict.predictor import Predictor
data = TaskData(args.test_data_num)
labels, sents = data.read_data(
raw_data_path=config["test_path"],
data_dir=config["data_dir"],
preprocessor=Preprocessor(config["preprocessor"])(
stopwords_path=config["stopwords_path"],
userdict_path=config["userdict_path"]),
is_train=False)
lines = list(zip(sents, labels))
processor = Postprocessor(
config["postprocessor"])(do_lower_case=args.do_lower_case)
label_list = processor.get_labels(config['data_dir'] / "labels.txt")
id2label = {i: label for i, label in enumerate(label_list)}
test_data = processor.get_test(lines=lines)
test_examples = processor.create_examples(
lines=test_data,
example_type='test',
cached_examples_file=config['data_dir'] /
"cached_test_examples_{}".format(args.pretrain))
test_features = processor.create_features(
examples=test_examples,
max_seq_len=args.eval_max_seq_len,
cached_features_file=config['data_dir'] /
"cached_test_features_{}_{}".format(args.eval_max_seq_len,
args.pretrain))
test_dataset = processor.create_dataset(test_features)
test_sampler = SequentialSampler(test_dataset)
test_dataloader = DataLoader(test_dataset,
sampler=test_sampler,
batch_size=args.train_batch_size)
if config["pretrain"] == "Nopretrain":
config["vocab_size"] = processor.vocab_size
model = Classifier(config["classifier"], config["pretrain"],
config["checkpoint_dir"])(num_labels=len(label_list))
########### predict ###########
logger.info('model predicting....')
predictor = Predictor(model=model, logger=logger, n_gpu=args.n_gpu)
logits, y_pred = predictor.predict(data=test_dataloader, thresh=0.5)
pred_labels = []
for item in y_pred.tolist():
tmp = []
for i, v in enumerate(item):
if v == 1:
tmp.append(label_list[i])
pred_labels.append(",".join(tmp))
assert len(pred_labels) == y_pred.shape[0]
df_pred_labels = pd.DataFrame(pred_labels, columns=["predict_labels"])
df_test_raw = pd.read_csv(config["test_path"])
if args.test_data_num > 0:
df_test_raw = df_test_raw.head(args.test_data_num)
df_labels = pd.DataFrame(logits, columns=label_list)
df = pd.concat([df_test_raw, df_pred_labels, df_labels], axis=1)
df.to_csv(config["result"] / "output.csv", index=False)
from predict.predictor import Predictor
from time import time
import json
# Read model definition
model_definition = ModelDefinition()
print("Reading exported model")
predictor = Predictor(model_definition)
# Sample input: First file word (sequence with all pad elements)
input = predictor.get_empty_element()
print(input)
json_test = json.dumps(input)
print("Prediction:", predictor.predict(input))
n_repetitions = 1000
print("Testing performance, n. repetitions:", n_repetitions)
start = time()
for i in range(n_repetitions):
predictor.predict_json(json_test)
end = time()
print("Total time:", end - start, "s")
print("Prediction performance:", ((end - start) / n_repetitions) * 1000, "ms")
# RNN: House computer (Linux):
# seq_len = 16, rnn size = 64 -> With GPU: 1.7 ms / With CPU: 0.85 ms
# seq_len= 64, rnn_size = 256 -> With GPU: 2.76 ms / With CPU: 4.24 ms
# RNN: Work computer (Windows 10):