def predict(data_set=None, data_per_batch=32, epoch=1, model_path='model.h5'):
Dataset = None
if data_set == None:
dataloaders = os.listdir('Dataset')
for dataloader in dataloaders:
loader_path = os.path.join('Dataset', dataloader)
if dataloader.endswith('.py') and os.path.isfile(
loader_path) and dataloader != '__init__.py':
try:
Dataset = importlib.import_module("Dataset." +
dataloader[:-3])
except Exception as ex:
print('failed to load Dataset from "%s".' % dataloader, ex)
else:
print('successfuly loaded Dataset from "%s"!' % dataloader)
break
if Dataset == None:
raise Exception('No vaild dataset found!')
else:
try:
Dataset = importlib.import_module("Dataset." + data_set)
except Exception as ex:
raise Exception('"%s" is not a vaild dataset!' % data_set)
data_loader = Dataset.DataLoader(1024, 1, 13)
# 加载网络模型
model = NN.model.create_pridict_model()
# 输出网络结构
model.summary()
# 加载之前训练的数据
model.load_weights(model_path)
# 验证集
validation_data = data_loader.get_validation_generator()
data = next(validation_data)[0]
r = model.predict(data['speech_data_input'])
r = K.ctc_decode(r, data['input_length'][0])
r1 = K.get_value(r[0][0])
r1 = r1[0]
tokens = NN.model.get_tokens()
print('predict: [', end='')
for i in r1:
print(tokens[i], end=', ')
print(']')
print('truth : [', end='')
for i in range(data['label_length'][0][0]):
print(tokens[int(data['speech_labels'][0][i])], end=', ')
print(']')
pass
def evaluate(self, batch_num):
batch_acc = 0
o_acc = 0
generator = self.val_seq.generator()
#关闭学习率
for i in range(batch_num):
inputs = next(generator)
x_test, y_test, source_str = (inputs["the_input"],
inputs["the_labels"],
inputs["source_str"])
out = self.validation_func([x_test, 0])[0]
current_acc = np.zeros([out.shape[0]])
c_acc = np.zeros([out.shape[0]])
#example one
# ctc_decode = K.ctc_decode(y_pred[:, 2:, :], input_length=np.ones(shape[0]) * shape[1])[0][0]
ctc_decode = K.get_value(
K.ctc_decode(out,
input_length=np.ones(out.shape[0]) * out.shape[1],
greedy=True)[0][0])
# print(ctc_decode)
for j in range(ctc_decode.shape[0]):
print("ctc_decode", ctc_decode[j], y_test[j][:4])
# out_best = list(np.argmax(decode_out[j, 2:], 1))
out_best = list(ctc_decode[j])
out_best = [k for k, g in itertools.groupby(out_best)]
if self.val_seq.equals_after_trim(y_test[j],
np.asarray(out_best)):
c_acc[j] = 1
print(source_str[j], y_test[j], out_best)
o_acc += c_acc.mean()
# print(" ctc_acc: %f%%" % (o_acc))
for j in range(out.shape[0]):
# 该层的输出结果是使用 max ,此处推断出最有可能的结果, 对每一列
out_best = list(np.argmax(out[j, 2:], 1))
out_best = [k for k, g in itertools.groupby(out_best)]
if self.val_seq.equals_after_trim(y_test[j],
np.asarray(out_best)):
current_acc[j] = 1
print(source_str[j], y_test[j], out_best)
batch_acc += current_acc.mean()
return batch_acc / batch_num, o_acc / batch_num