kantts
TTS appalication based on modelscope KAN-TTS
demo
python demo.py
Features - 特性
-
增加语速调节,scale参数调节
-
推理加速
1.HifiGAN tensorrt加速
2.sambert tensorrt加速
HifiGAN tensorrt加速
cd tensorrt_onnx
# 导出hifigan onnx
python hifigan_onnx_export.py
# 利用trt.OnnxParser生成engine
python build_from_onnx.py
本项目原始模型直接利用modelscope下载
from modelscope.utils.constant import Tasks
model_id = 'damo/speech_sambert-hifigan_tts_zh-cn_16k'
sambert_hifigan_tts = pipeline(task=Tasks.text_to_speech, model=model_id)
text = "你好"
output = sambert_hifigan_tts(input=text, voice="zhitian_emo") # zhibei_emo zhitian_emo zhiyan_emo zhizhe_emo
sambert tensorrt加速
python sambert_convert_weights.py
cd tensorrt_onnx
# 生成engine
python build_from_network.py
ps. 主要思路来自kantts群的大佬
1、统计sambert各个模块耗时,发现主要耗时都集中在MelPNCADecoder;
2、将MelPNCADecoder部分由一个循环的mel_dec函数组成;
它在memory的第二个维度上循环调用,下一次调用会依赖上一次的结果
3、将mel_dec函数利用tensorrt python api重写,因为mel_del的调用会依赖上一次mel_del的中间变量和输出结果,故修改该函数输入输出,将需要的中间结果都输入到下一次调用;
step=0时和step>0时的输入不一致,所有只对step>0时的推理部分进行tensorrt搭建。
输入部分:
input = network.add_input(
name="input", dtype=trt.float32, shape=(1, 1, 80))
memory = network.add_input(
name="memory", dtype=trt.float32, shape=(1, -1, 160))
memory_step = network.add_input(
name="memory_step", dtype=trt.float32, shape=(1, 1, 160))
pnca_x_attn_mask_step_part1 = network.add_input(
name="pnca_x_attn_mask_step_part1", dtype=trt.float32, shape=(8, 1, -1))
pnca_x_attn_mask_step_part2 = network.add_input(
name="pnca_x_attn_mask_step_part2", dtype=trt.float32, shape=(8, 1, 1))
pnca_h_attn_mask_step = network.add_input(
name="pnca_h_attn_mask_step", dtype=trt.float32, shape=(8, 1, -1))
pre_x_k = network.add_input(
name="pre_x_k", dtype=trt.float32, shape=(8*12, -1, 16))
pre_x_v = network.add_input(
name="pre_x_v", dtype=trt.float32, shape=(8*12, -1, 16))
推理输入输出转换
pnca_x_attn_mask_step_ = pnca_x_attn_mask_step.repeat(8, 1, 1)
pnca_h_attn_mask_step_ = pnca_h_attn_mask_step.repeat(8, 1, 1)
pnca_x_attn_mask_step_part1 = pnca_x_attn_mask_step_[:, :, :-1]
pnca_x_attn_mask_step_part2 = pnca_x_attn_mask_step_[:, :, -1:]
pre_x_k_merge = None
pre_x_v_merge = None
for i in range(len(pre_x_k_list)):
if i == 0:
pre_x_k_merge = pre_x_k_list[i]
pre_x_v_merge = pre_x_v_list[i]
else:
pre_x_k_merge = torch.cat([pre_x_k_merge, pre_x_k_list[i]], dim=0)
pre_x_v_merge = torch.cat([pre_x_v_merge, pre_x_v_list[i]], dim=0)
# print(pre_x_k_merge.shape)
# print(pre_x_v_merge.shape)
output_trt = self.trt_model({"input": input, "memory": memory, "memory_step": memory_step,
"pnca_x_attn_mask_step_part1": pnca_x_attn_mask_step_part1.float(),
"pnca_x_attn_mask_step_part2": pnca_x_attn_mask_step_part2.float(),
"pnca_h_attn_mask_step": pnca_h_attn_mask_step_.float(),
"pre_x_k": pre_x_k_merge,
"pre_x_v": pre_x_v_merge})
dec_output_step = output_trt['output']
dec_pnca_attn_x_step = []
dec_pnca_attn_h_step = []
for i in range(12):
dec_pnca_attn_x_step += [output_trt[f'dec_pnca_attn_x_{i}']]
dec_pnca_attn_h_step += [output_trt[f'dec_pnca_attn_h_{i}']]
pre_x_k_list = []
pre_x_v_list = []
for i in range(12):
pre_x_k_list += [output_trt[f'x_k_{i}']]
pre_x_v_list += [output_trt[f'x_v_{i}']]
模型搭建部分 masked_fill方法有两个参数,maske和value,mask是一个pytorch张量(Tensor),元素是布尔值,value是要填充的值,填充规则是mask中取值为True位置对应于待填充的相应位置用value填充。 例如本项目中,需要实现 attn = attn.masked_fill(mask, -np.inf) 填充值为负无穷 trt实现步骤 ● 将mask由bool转换成float32,变为由0和1组成的tensor ● 然后将mask利用逐元素相乘的方式,乘以一个很大的负数,这里设置为-100000000000000000.0
inf_const = network.add_constant((1,1,1), (-100000000000000000.0)*np.ones((1,1,1)).astype(np.float32))
mask = network.add_elementwise(mask, inf_const.get_output(0), trt.ElementWiseOperation.PROD)
● 最后attn和mask逐元素相加
layernorm算子可以使用8.6的tensorrt,也可以自己按照公式实现