\u4f60\u662f\u4e0d\u662f\u4e5f\u9047\u5230\u8fc7\u8fd9\u6837\u7684\u70e6\u607c？\u56e2\u961f\u91cc\u90e8\u7f72\u4e86\u597d\u51e0\u4e2aAI\u6a21\u578b，\u6bcf\u4e2a\u6a21\u578b\u90fd\u6709\u81ea\u5df1\u7684\u73af\u5883\u4f9d\u8d56\u3001\u542f\u52a8\u811a\u672c\u548cAPI\u63a5\u53e3，\u7ba1\u7406\u8d77\u6765\u50cf\u4e00\u76d8\u6563\u6c99\u3002\u4eca\u5929\u8981\u90e8\u7f72\u4e00\u4e2a\u8bed\u97f3\u964d\u566a\u6a21\u578b，\u660e\u5929\u8981\u4e0a\u7ebf\u4e00\u4e2a\u56fe\u50cf\u8bc6\u522b\u670d\u52a1，\u540e\u5929\u53ef\u80fd\u53c8\u8981\u52a0\u4e00\u4e2a\u6587\u672c\u751f\u6210\u5de5\u5177\u3002\u6bcf\u6b21\u90fd\u8981\u5355\u72ec\u914d\u7f6e，\u4e0d\u4ec5\u6548\u7387\u4f4e\u4e0b，\u8fd8\u5bb9\u6613\u51fa\u9519\u3002<\/p>\n
\u5982\u679c\u4f60\u6b63\u5728\u4e3a\u591a\u6a21\u578b\u7ba1\u7406\u5934\u75bc，\u90a3\u4e48NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u53ef\u80fd\u5c31\u662f\u4f60\u7684\u6551\u661f\u3002\u5b83\u5c31\u50cf\u4e00\u4e2a\u201cAI\u6a21\u578b\u7ba1\u5bb6\u201d，\u80fd\u628a\u4e0d\u540c\u6846\u67b6\u3001\u4e0d\u540c\u7248\u672c\u7684\u6a21\u578b\u7edf\u4e00\u7ba1\u7406\u8d77\u6765，\u63d0\u4f9b\u6807\u51c6\u5316\u7684\u670d\u52a1\u63a5\u53e3\u3002<\/p>\n
\u4eca\u5929，\u6211\u5c31\u624b\u628a\u624b\u5e26\u4f60\u7528Triton\u6765\u90e8\u7f72\u4e00\u4e2a\u975e\u5e38\u5b9e\u7528\u7684\u8bed\u97f3\u964d\u566a\u6a21\u578b\u2014\u2014FRCRN\u3002\u8fd9\u4e2a\u6a21\u578b\u6765\u81ea\u963f\u91cc\u5df4\u5df4\u8fbe\u6469\u9662，\u4e13\u95e8\u5904\u7406\u5355\u901a\u9053\u97f3\u9891\u7684\u566a\u58f0\u6d88\u9664，\u6548\u679c\u76f8\u5f53\u4e0d\u9519\u3002\u66f4\u91cd\u8981\u7684\u662f，\u901a\u8fc7\u8fd9\u6b21\u5b9e\u8df5，\u4f60\u80fd\u638c\u63e1\u7528Triton\u7ba1\u7406\u4efb\u610f\u6a21\u578b\u7684\u65b9\u6cd5\u3002<\/p>\n
\u5728\u5f00\u59cb\u52a8\u624b\u4e4b\u524d，\u6211\u4eec\u5148\u641e\u6e05\u695a\u4e24\u4e2a\u95ee\u9898：\u4e3a\u4ec0\u4e48\u8981\u7528Triton？\u4e3a\u4ec0\u4e48\u8981\u90e8\u7f72FRCRN？<\/p>\n
\u60f3\u8c61\u4e00\u4e0b，\u5982\u679c\u6ca1\u6709Triton，\u4f60\u7684AI\u670d\u52a1\u90e8\u7f72\u6d41\u7a0b\u53ef\u80fd\u662f\u8fd9\u6837\u7684：<\/p>\n
\u6709\u4e86Triton\u4e4b\u540e，\u6d41\u7a0b\u5c31\u7b80\u5316\u591a\u4e86：<\/p>\n
FRCRN（Frequency-Recurrent Convolutional Recurrent Network）\u662f\u963f\u91cc\u5df4\u5df4\u8fbe\u6469\u9662\u5f00\u6e90\u7684\u8bed\u97f3\u964d\u566a\u6a21\u578b，\u5728ModelScope\u793e\u533a\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u3002\u5b83\u7684\u7279\u70b9\u662f：<\/p>\n
\u8fd9\u4e2a\u6a21\u578b\u7279\u522b\u9002\u5408\u7528\u5728：<\/p>\n
\u597d\u4e86，\u7406\u8bba\u8bf4\u5b8c\u4e86，\u54b1\u4eec\u5f00\u59cb\u52a8\u624b\u3002\u9996\u5148\u5f97\u628aTriton\u88c5\u8d77\u6765\u3002<\/p>\n
\u5728\u5f00\u59cb\u4e4b\u524d，\u786e\u8ba4\u4f60\u7684\u73af\u5883\u6ee1\u8db3\u4ee5\u4e0b\u8981\u6c42：<\/p>\n
\u5982\u679c\u4f60\u7528\u7684\u662f\u4e91\u670d\u52a1\u5668，\u8fd9\u4e9b\u901a\u5e38\u90fd\u5df2\u7ecf\u9884\u88c5\u597d\u4e86\u3002\u672c\u5730\u73af\u5883\u7684\u8bdd，\u53ef\u4ee5\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\u68c0\u67e5：<\/p>\n
# \u68c0\u67e5Docker\u7248\u672c
\ndocker –version<\/p>\n
# \u68c0\u67e5NVIDIA\u9a71\u52a8
\nnvidia-smi<\/p>\n
# \u68c0\u67e5CUDA\u7248\u672c
\nnvcc –version<\/p>\n
Triton\u9700\u8981Docker\u80fd\u591f\u8bbf\u95eeGPU，\u6240\u4ee5\u8981\u5148\u5b89\u88c5NVIDIA Container Toolkit：<\/p>\n
# \u6dfb\u52a0NVIDIA\u7684\u5305\u4ed3\u5e93
\ndistribution=$(. \/etc\/os-release;echo $ID$VERSION_ID)
\ncurl -s -L https:\/\/nvidia.github.io\/nvidia-docker\/gpgkey | sudo apt-key add –
\ncurl -s -L https:\/\/nvidia.github.io\/nvidia-docker\/$distribution\/nvidia-docker.list | sudo tee \/etc\/apt\/sources.list.d\/nvidia-docker.list<\/p>\n
# \u5b89\u88c5\u5de5\u5177\u5305
\nsudo apt-get update
\nsudo apt-get install -y nvidia-docker2<\/p>\n
# \u91cd\u542fDocker\u670d\u52a1
\nsudo systemctl restart docker<\/p>\n
\u5b89\u88c5\u5b8c\u6210\u540e，\u6d4b\u8bd5\u4e00\u4e0bGPU\u5728Docker\u4e2d\u662f\u5426\u53ef\u7528：<\/p>\n
# \u8fd0\u884c\u4e00\u4e2a\u6d4b\u8bd5\u5bb9\u5668
\nsudo docker run –rm –gpus all nvidia\/cuda:11.0-base nvidia-smi<\/p>\n
\u5982\u679c\u80fd\u770b\u5230GPU\u4fe1\u606f，\u8bf4\u660e\u914d\u7f6e\u6210\u529f\u4e86\u3002<\/p>\n
Triton\u63d0\u4f9b\u4e86\u591a\u4e2a\u7248\u672c\u7684\u955c\u50cf，\u6211\u4eec\u9009\u62e9\u5305\u542bPyTorch\u540e\u7aef\u7684\u7248\u672c，\u56e0\u4e3aFRCRN\u662fPyTorch\u6a21\u578b：<\/p>\n
# \u62c9\u53d6Triton\u670d\u52a1\u5668\u955c\u50cf
\ndocker pull nvcr.io\/nvidia\/tritonserver:22.12-py3<\/p>\n
# \u8fd9\u4e2a\u955c\u50cf\u6bd4\u8f83\u5927，\u7ea610GB，\u9700\u8981\u8010\u5fc3\u7b49\u5f85
\n# \u5982\u679c\u4e0b\u8f7d\u6162，\u53ef\u4ee5\u914d\u7f6eDocker\u955c\u50cf\u52a0\u901f<\/p>\n
\u8fd9\u4e2a\u955c\u50cf\u5305\u542b\u4e86：<\/p>\n
Triton\u9700\u8981\u6a21\u578b\u6309\u7167\u7279\u5b9a\u7684\u76ee\u5f55\u7ed3\u6784\u6765\u7ec4\u7ec7，\u6240\u4ee5\u6211\u4eec\u8981\u5148\u628aFRCRN\u6a21\u578b\u201c\u5305\u88c5\u201d\u6210Triton\u80fd\u8bc6\u522b\u7684\u683c\u5f0f\u3002<\/p>\n
\u9996\u5148，\u521b\u5efa\u4e00\u4e2a\u5de5\u4f5c\u76ee\u5f55，\u7136\u540e\u6309\u7167Triton\u7684\u8981\u6c42\u7ec4\u7ec7\u6587\u4ef6：<\/p>\n
# \u521b\u5efa\u5de5\u4f5c\u76ee\u5f55
\nmkdir -p ~\/triton_models
\ncd ~\/triton_models<\/p>\n
# \u521b\u5efaFRCRN\u6a21\u578b\u7684\u76ee\u5f55\u7ed3\u6784
\nmkdir -p frcrn\/1
\nmkdir -p frcrn\/config<\/p>\n
Triton\u7684\u6a21\u578b\u76ee\u5f55\u7ed3\u6784\u662f\u8fd9\u6837\u7684：<\/p>\n
frcrn\/ # \u6a21\u578b\u540d\u79f0
\n\u251c\u2500\u2500 1\/ # \u7248\u672c\u53f7（\u5fc5\u987b\u662f\u6570\u5b57）
\n\u2502 \u2514\u2500\u2500 model.py # \u6a21\u578b\u63a8\u7406\u811a\u672c
\n\u2514\u2500\u2500 config.pbtxt # \u6a21\u578b\u914d\u7f6e\u6587\u4ef6<\/p>\n
\u5728frcrn\/config.pbtxt\u4e2d，\u6211\u4eec\u9700\u8981\u544a\u8bc9Triton\u8fd9\u4e2a\u6a21\u578b\u7684\u57fa\u672c\u4fe1\u606f：<\/p>\n
name: "frcrn"
\nplatform: "pytorch_libtorch"
\nmax_batch_size: 8<\/p>\n
input [
\n {
\n name: "audio_input"
\n data_type: TYPE_FP32
\n dims: [ -1, 1 ] # \u52a8\u6001\u7ef4\u5ea6，\u652f\u6301\u4e0d\u540c\u957f\u5ea6\u7684\u97f3\u9891
\n }
\n]<\/p>\n
output [
\n {
\n name: "audio_output"
\n data_type: TYPE_FP32
\n dims: [ -1, 1 ] # \u8f93\u51fa\u4e0e\u8f93\u5165\u76f8\u540c\u957f\u5ea6
\n }
\n]<\/p>\n
instance_group [
\n {
\n count: 1
\n kind: KIND_GPU
\n }
\n]<\/p>\n
parameters [
\n {
\n key: "sampling_rate"
\n value: { string_value: "16000" }
\n }
\n]<\/p>\n
\u8fd9\u4e2a\u914d\u7f6e\u6587\u4ef6\u5b9a\u4e49\u4e86：<\/p>\n
\u8fd9\u662f\u6700\u5173\u952e\u7684\u4e00\u6b65，\u6211\u4eec\u9700\u8981\u5728frcrn\/1\/model.py\u4e2d\u5b9e\u73b0\u6a21\u578b\u7684\u52a0\u8f7d\u548c\u63a8\u7406\u903b\u8f91：<\/p>\n
import torch
\nimport torch.nn as nn
\nimport numpy as np
\nfrom typing import Dict, List
\nimport triton_python_backend_utils as pb_utils<\/p>\n
class FRCRNModel(nn.Module):
\n """FRCRN\u6a21\u578b\u5c01\u88c5\u7c7b"""<\/p>\n
def __init__(self):
\n super(FRCRNModel, self).__init__()
\n # \u8fd9\u91cc\u7b80\u5316\u4e86\u6a21\u578b\u7ed3\u6784，\u5b9e\u9645\u4f7f\u7528\u65f6\u9700\u8981\u5bfc\u5165\u5b8c\u6574\u7684FRCRN\u6a21\u578b
\n # \u4e3a\u4e86\u6559\u7a0b\u6e05\u6670，\u6211\u4eec\u5148\u4f7f\u7528\u4e00\u4e2a\u7b80\u5355\u7684\u964d\u566a\u7f51\u7edc
\n self.conv1 = nn.Conv1d(1, 16, kernel_size=3, padding=1)
\n self.conv2 = nn.Conv1d(16, 1, kernel_size=3, padding=1)
\n self.relu = nn.ReLU()<\/p>\n
def forward(self, x):
\n # \u7b80\u5355\u7684\u964d\u566a\u5904\u7406：\u4e24\u5c42\u5377\u79ef
\n x = self.conv1(x)
\n x = self.relu(x)
\n x = self.conv2(x)
\n return x<\/p>\n
class TritonPythonModel:
\n """Triton Python\u540e\u7aef\u6a21\u578b\u7c7b"""<\/p>\n
def initialize(self, args):
\n """\u6a21\u578b\u521d\u59cb\u5316"""
\n self.logger = pb_utils.Logger
\n self.model = FRCRNModel()<\/p>\n
# \u52a0\u8f7d\u9884\u8bad\u7ec3\u6743\u91cd（\u8fd9\u91cc\u9700\u8981\u66ff\u6362\u4e3a\u5b9e\u9645\u7684FRCRN\u6743\u91cd\u8def\u5f84）
\n # checkpoint = torch.load('frcrn_weights.pth')
\n # self.model.load_state_dict(checkpoint)<\/p>\n
self.model.eval()
\n self.logger.log_info("FRCRN\u6a21\u578b\u521d\u59cb\u5316\u5b8c\u6210")<\/p>\n
def execute(self, requests):
\n """\u5904\u7406\u63a8\u7406\u8bf7\u6c42"""
\n responses = []<\/p>\n
for request in requests:
\n # \u83b7\u53d6\u8f93\u5165\u6570\u636e
\n audio_input = pb_utils.get_input_tensor_by_name(request, "audio_input")
\n audio_data = audio_input.as_numpy()<\/p>\n
# \u8f6c\u6362\u4e3aPyTorch\u5f20\u91cf
\n audio_tensor = torch.from_numpy(audio_data).float()<\/p>\n
# \u6267\u884c\u63a8\u7406
\n with torch.no_grad():
\n # \u6dfb\u52a0\u6279\u6b21\u7ef4\u5ea6（\u5982\u679c\u9700\u8981）
\n if len(audio_tensor.shape) == 1:
\n audio_tensor = audio_tensor.unsqueeze(0).unsqueeze(0)
\n elif len(audio_tensor.shape) == 2:
\n audio_tensor = audio_tensor.unsqueeze(1)<\/p>\n
# \u6a21\u578b\u63a8\u7406
\n output_tensor = self.model(audio_tensor)<\/p>\n
# \u79fb\u9664\u6279\u6b21\u7ef4\u5ea6
\n output_tensor = output_tensor.squeeze()
\n if len(output_tensor.shape) == 1:
\n output_tensor = output_tensor.unsqueeze(0)<\/p>\n
# \u521b\u5efa\u8f93\u51fa\u5f20\u91cf
\n output_numpy = output_tensor.numpy()
\n output_tensor = pb_utils.Tensor("audio_output", output_numpy)<\/p>\n
# \u521b\u5efa\u54cd\u5e94
\n response = pb_utils.InferenceResponse(output_tensors=[output_tensor])
\n responses.append(response)<\/p>\n
return responses<\/p>\n
def finalize(self):
\n """\u6e05\u7406\u8d44\u6e90"""
\n self.logger.log_info("FRCRN\u6a21\u578b\u6e05\u7406\u5b8c\u6210")<\/p>\n
\u91cd\u8981\u8bf4\u660e：\u4e0a\u9762\u7684\u4ee3\u7801\u662f\u4e00\u4e2a\u7b80\u5316\u7248\u672c，\u7528\u4e8e\u6f14\u793aTriton\u6a21\u578b\u7684\u57fa\u672c\u7ed3\u6784\u3002\u5b9e\u9645\u90e8\u7f72FRCRN\u65f6，\u4f60\u9700\u8981：<\/p>\n
\u5982\u679c\u4f60\u8981\u90e8\u7f72\u771f\u5b9e\u7684FRCRN\u6a21\u578b，\u53ef\u4ee5\u8fd9\u6837\u51c6\u5907：<\/p>\n
# \u5b89\u88c5ModelScope
\npip install modelscope torchaudio<\/p>\n
# \u4e0b\u8f7dFRCRN\u6a21\u578b
\nfrom modelscope.pipelines import pipeline
\nfrom modelscope.utils.constant import Tasks<\/p>\n
# \u521b\u5efa\u8bed\u97f3\u964d\u566apipeline
\nans_pipeline = pipeline(
\n task=Tasks.acoustic_noise_suppression,
\n model='damo\/speech_frcrn_ans_cirm_16k'
\n)<\/p>\n
# \u4fdd\u5b58\u4e3aTorchScript\u683c\u5f0f（Triton\u9700\u8981）
\ndummy_input = torch.randn(1, 1, 16000) # 1\u79d2\u7684\u97f3\u9891
\ntraced_model = torch.jit.trace(ans_pipeline.model, dummy_input)
\ntraced_model.save("frcrn_model.pt")<\/p>\n
\u7136\u540e\u628a\u4fdd\u5b58\u7684frcrn_model.pt\u653e\u5230\u6a21\u578b\u76ee\u5f55\u4e2d，\u5e76\u4fee\u6539model.py\u6765\u52a0\u8f7d\u8fd9\u4e2a\u6a21\u578b\u3002<\/p>\n
\u6a21\u578b\u51c6\u5907\u597d\u4e86，\u73b0\u5728\u53ef\u4ee5\u542f\u52a8Triton\u670d\u52a1\u5668\u4e86\u3002<\/p>\n
# \u542f\u52a8Triton\u670d\u52a1\u5668
\ndocker run –gpus=all –rm \\\\
\n -p 8000:8000 -p 8001:8001 -p 8002:8002 \\\\
\n -v ~\/triton_models:\/models \\\\
\n nvcr.io\/nvidia\/tritonserver:22.12-py3 \\\\
\n tritonserver –model-repository=\/models<\/p>\n
\u8fd9\u4e2a\u547d\u4ee4\u505a\u4e86\u51e0\u4ef6\u4e8b：<\/p>\n
\u542f\u52a8\u540e，\u4f60\u4f1a\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u8f93\u51fa：<\/p>\n
I1230 10:00:00.000000 1 server.cc:592]
\n+——————+——+——–+
\n| Model | Version | Status |
\n+——————+——+——–+
\n| frcrn | 1 | READY |
\n+——————+——+——–+<\/p>\n
\u8fd9\u8868\u793aFRCRN\u6a21\u578b\u5df2\u7ecf\u52a0\u8f7d\u6210\u529f，\u72b6\u6001\u662fREADY\u3002<\/p>\n
\u4f60\u8fd8\u53ef\u4ee5\u901a\u8fc7HTTP\u63a5\u53e3\u68c0\u67e5\u670d\u52a1\u5668\u72b6\u6001：<\/p>\n
# \u68c0\u67e5\u670d\u52a1\u5668\u5065\u5eb7\u72b6\u6001
\ncurl -v localhost:8000\/v2\/health\/ready<\/p>\n
# \u67e5\u770b\u5df2\u52a0\u8f7d\u7684\u6a21\u578b
\ncurl localhost:8000\/v2\/models<\/p>\n
\u5982\u679c\u542f\u52a8\u5931\u8d25，\u53ef\u4ee5\u68c0\u67e5\u4ee5\u4e0b\u51e0\u70b9：<\/p>\n
\u670d\u52a1\u5668\u8dd1\u8d77\u6765\u4e86，\u73b0\u5728\u6211\u4eec\u6765\u5199\u4e2a\u5ba2\u6237\u7aef\u6d4b\u8bd5\u4e00\u4e0b\u3002<\/p>\n
\u9996\u5148\u5b89\u88c5Triton\u5ba2\u6237\u7aef\u5e93：<\/p>\n
pip install tritonclient[all]<\/p>\n
\u7136\u540e\u5199\u4e00\u4e2a\u7b80\u5355\u7684\u5ba2\u6237\u7aef\u811a\u672c：<\/p>\n
import numpy as np
\nimport tritonclient.http as httpclient
\nimport soundfile as sf
\nimport librosa<\/p>\n
class FRCRNClient:
\n """FRCRN Triton\u5ba2\u6237\u7aef"""<\/p>\n
def __init__(self, url="localhost:8000"):
\n self.client = httpclient.InferenceServerClient(url=url)
\n self.model_name = "frcrn"<\/p>\n
def preprocess_audio(self, audio_path):
\n """\u9884\u5904\u7406\u97f3\u9891：\u52a0\u8f7d\u3001\u91cd\u91c7\u6837\u3001\u5f52\u4e00\u5316"""
\n # \u52a0\u8f7d\u97f3\u9891
\n audio, sr = librosa.load(audio_path, sr=None)<\/p>\n
# \u91cd\u91c7\u6837\u523016kHz
\n if sr != 16000:
\n audio = librosa.resample(audio, orig_sr=sr, target_sr=16000)<\/p>\n
# \u8f6c\u6362\u4e3a\u5355\u58f0\u9053
\n if len(audio.shape) > 1:
\n audio = librosa.to_mono(audio)<\/p>\n
# \u5f52\u4e00\u5316\u5230[-1, 1]
\n audio = audio \/ np.max(np.abs(audio))<\/p>\n
# \u6dfb\u52a0\u6279\u6b21\u7ef4\u5ea6
\n audio = audio.astype(np.float32)
\n audio = audio.reshape(1, -1)<\/p>\n
return audio<\/p>\n
def denoise(self, audio_path, output_path="denoised.wav"):
\n """\u6267\u884c\u964d\u566a"""
\n # \u9884\u5904\u7406\u97f3\u9891
\n audio_data = self.preprocess_audio(audio_path)<\/p>\n
# \u51c6\u5907\u8f93\u5165
\n inputs = [
\n httpclient.InferInput(
\n "audio_input",
\n audio_data.shape,
\n "FP32"
\n )
\n ]
\n inputs[0].set_data_from_numpy(audio_data)<\/p>\n
# \u51c6\u5907\u8f93\u51fa
\n outputs = [
\n httpclient.InferRequestedOutput("audio_output")
\n ]<\/p>\n
# \u53d1\u9001\u8bf7\u6c42
\n response = self.client.infer(
\n model_name=self.model_name,
\n inputs=inputs,
\n outputs=outputs
\n )<\/p>\n
# \u83b7\u53d6\u7ed3\u679c
\n result = response.as_numpy("audio_output")<\/p>\n
# \u4fdd\u5b58\u7ed3\u679c
\n sf.write(output_path, result[0], 16000)<\/p>\n
print(f"\u964d\u566a\u5b8c\u6210，\u7ed3\u679c\u4fdd\u5b58\u5230: {output_path}")
\n return result<\/p>\n
def batch_denoise(self, audio_paths):
\n """\u6279\u91cf\u5904\u7406\u591a\u4e2a\u97f3\u9891"""
\n results = []
\n for path in audio_paths:
\n print(f"\u5904\u7406: {path}")
\n result = self.denoise(path)
\n results.append(result)
\n return results<\/p>\n
# \u4f7f\u7528\u793a\u4f8b
\nif __name__ == "__main__":
\n # \u521b\u5efa\u5ba2\u6237\u7aef
\n client = FRCRNClient()<\/p>\n
# \u5904\u7406\u5355\u4e2a\u97f3\u9891
\n client.denoise("noisy_audio.wav", "clean_audio.wav")<\/p>\n
# \u6279\u91cf\u5904\u7406
\n # audio_list = ["audio1.wav", "audio2.wav", "audio3.wav"]
\n # client.batch_denoise(audio_list)<\/p>\n
\u5982\u679c\u4f60\u89c9\u5f97\u4e0a\u9762\u7684\u4ee3\u7801\u592a\u590d\u6742，Triton\u8fd8\u63d0\u4f9b\u4e86\u66f4\u7b80\u5355\u7684\u8c03\u7528\u65b9\u5f0f：<\/p>\n
# \u4f7f\u7528Triton\u7684\u7b80\u6613\u5ba2\u6237\u7aef
\nfrom tritonclient.utils import *
\nimport tritonclient.http as httpclient
\nimport numpy as np<\/p>\n
# \u8fde\u63a5\u670d\u52a1\u5668
\ntriton_client = httpclient.InferenceServerClient(url="localhost:8000")<\/p>\n
# \u51c6\u5907\u97f3\u9891\u6570\u636e（\u5047\u8bbe\u5df2\u7ecf\u9884\u5904\u7406\u597d\u4e86）
\naudio_data = np.random.randn(1, 16000).astype(np.float32) # 1\u79d2\u7684\u6d4b\u8bd5\u97f3\u9891<\/p>\n
# \u6267\u884c\u63a8\u7406
\ninputs = [httpclient.InferInput("audio_input", audio_data.shape, "FP32")]
\ninputs[0].set_data_from_numpy(audio_data)<\/p>\n
outputs = [httpclient.InferRequestedOutput("audio_output")]<\/p>\n
results = triton_client.infer(
\n model_name="frcrn",
\n inputs=inputs,
\n outputs=outputs
\n)<\/p>\n
# \u83b7\u53d6\u7ed3\u679c
\noutput_data = results.as_numpy("audio_output")
\nprint(f"\u8f93\u51fa\u97f3\u9891\u5f62\u72b6: {output_data.shape}")<\/p>\n
\u6211\u4eec\u8fd8\u53ef\u4ee5\u6d4b\u8bd5\u4e00\u4e0b\u670d\u52a1\u7684\u6027\u80fd：<\/p>\n
import time<\/p>\n
def benchmark_client(client, audio_data, num_requests=100):
\n """\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5"""
\n latencies = []<\/p>\n
for i in range(num_requests):
\n start_time = time.time()<\/p>\n
# \u6267\u884c\u63a8\u7406
\n inputs = [httpclient.InferInput("audio_input", audio_data.shape, "FP32")]
\n inputs[0].set_data_from_numpy(audio_data)
\n outputs = [httpclient.InferRequestedOutput("audio_output")]<\/p>\n
_ = client.infer(
\n model_name="frcrn",
\n inputs=inputs,
\n outputs=outputs
\n )<\/p>\n
latency = (time.time() – start_time) * 1000 # \u8f6c\u6362\u4e3a\u6beb\u79d2
\n latencies.append(latency)<\/p>\n
if (i + 1) % 10 == 0:
\n print(f"\u5df2\u5b8c\u6210 {i + 1}\/{num_requests} \u6b21\u8bf7\u6c42")<\/p>\n
# \u7edf\u8ba1\u7ed3\u679c
\n avg_latency = np.mean(latencies)
\n p95_latency = np.percentile(latencies, 95)<\/p>\n
print(f"\\\\n\u6027\u80fd\u6d4b\u8bd5\u7ed3\u679c:")
\n print(f"\u5e73\u5747\u5ef6\u8fdf: {avg_latency:.2f} ms")
\n print(f"P95\u5ef6\u8fdf: {p95_latency:.2f} ms")
\n print(f"QPS: {1000 \/ avg_latency:.2f}")<\/p>\n
return latencies<\/p>\n
# \u8fd0\u884c\u6d4b\u8bd5
\n# test_audio = np.random.randn(1, 16000).astype(np.float32) # 1\u79d2\u97f3\u9891
\n# benchmark_client(triton_client, test_audio, num_requests=100)<\/p>\n
Triton\u6700\u5f3a\u5927\u7684\u5730\u65b9\u5728\u4e8e\u80fd\u7edf\u4e00\u7ba1\u7406\u591a\u4e2a\u6a21\u578b\u3002\u5047\u8bbe\u6211\u4eec\u9664\u4e86FRCRN，\u8fd8\u60f3\u90e8\u7f72\u4e00\u4e2a\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\u548c\u4e00\u4e2a\u8bed\u97f3\u5408\u6210\u6a21\u578b\u3002<\/p>\n
\u53ea\u9700\u8981\u5728\u6a21\u578b\u4ed3\u5e93\u76ee\u5f55\u4e0b\u521b\u5efa\u65b0\u7684\u6a21\u578b\u6587\u4ef6\u5939：<\/p>\n
# \u6dfb\u52a0\u8bed\u97f3\u8bc6\u522b\u6a21\u578b
\nmkdir -p ~\/triton_models\/speech_recognition\/1
\nmkdir -p ~\/triton_models\/speech_recognition\/config<\/p>\n
# \u6dfb\u52a0\u8bed\u97f3\u5408\u6210\u6a21\u578b
\nmkdir -p ~\/triton_models\/speech_synthesis\/1
\nmkdir -p ~\/triton_models\/speech_synthesis\/config<\/p>\n
\u7136\u540e\u4e3a\u6bcf\u4e2a\u6a21\u578b\u51c6\u5907\u5bf9\u5e94\u7684config.pbtxt\u548cmodel.py\u6587\u4ef6\u3002<\/p>\n
\u6709\u4e86\u591a\u4e2a\u6a21\u578b\u540e，\u6211\u4eec\u53ef\u4ee5\u521b\u5efa\u4e00\u4e2a\u7edf\u4e00\u7684API\u7f51\u5173：<\/p>\n
class AIServiceGateway:
\n """AI\u670d\u52a1\u7f51\u5173，\u7edf\u4e00\u7ba1\u7406\u591a\u4e2a\u6a21\u578b"""<\/p>\n
def __init__(self, triton_url="localhost:8000"):
\n self.client = httpclient.InferenceServerClient(url=triton_url)
\n self.models = {
\n "denoise": "frcrn",
\n "asr": "speech_recognition",
\n "tts": "speech_synthesis"
\n }<\/p>\n
def process_pipeline(self, audio_path):
\n """\u5904\u7406\u6d41\u6c34\u7ebf：\u964d\u566a -> \u8bc6\u522b -> \u5408\u6210"""
\n # 1. \u964d\u566a
\n print("\u6b65\u9aa41: \u8bed\u97f3\u964d\u566a")
\n denoised_audio = self.denoise(audio_path)<\/p>\n
# 2. \u8bed\u97f3\u8bc6\u522b
\n print("\u6b65\u9aa42: \u8bed\u97f3\u8bc6\u522b")
\n text = self.speech_to_text(denoised_audio)<\/p>\n
# 3. \u8bed\u97f3\u5408\u6210（\u53ef\u9009）
\n print("\u6b65\u9aa43: \u8bed\u97f3\u5408\u6210")
\n synthesized_audio = self.text_to_speech(text)<\/p>\n
return {
\n "denoised_audio": denoised_audio,
\n "text": text,
\n "synthesized_audio": synthesized_audio
\n }<\/p>\n
def denoise(self, audio_data):
\n """\u8c03\u7528FRCRN\u964d\u566a"""
\n # … \u8c03\u7528FRCRN\u7684\u4ee3\u7801 …
\n pass<\/p>\n
def speech_to_text(self, audio_data):
\n """\u8c03\u7528\u8bed\u97f3\u8bc6\u522b\u6a21\u578b"""
\n # … \u8c03\u7528ASR\u6a21\u578b\u7684\u4ee3\u7801 …
\n pass<\/p>\n
def text_to_speech(self, text):
\n """\u8c03\u7528\u8bed\u97f3\u5408\u6210\u6a21\u578b"""
\n # … \u8c03\u7528TTS\u6a21\u578b\u7684\u4ee3\u7801 …
\n pass<\/p>\n
# \u4f7f\u7528\u793a\u4f8b
\n# gateway = AIServiceGateway()
\n# result = gateway.process_pipeline("noisy_audio.wav")<\/p>\n
Triton\u652f\u6301\u6a21\u578b\u7248\u672c\u7ba1\u7406，\u4f60\u53ef\u4ee5\u540c\u65f6\u90e8\u7f72\u591a\u4e2a\u7248\u672c\u7684\u6a21\u578b：<\/p>\n
# \u6a21\u578b\u76ee\u5f55\u7ed3\u6784
\nfrcrn\/
\n\u251c\u2500\u2500 1\/ # \u7248\u672c1
\n\u2502 \u2514\u2500\u2500 model.py
\n\u251c\u2500\u2500 2\/ # \u7248\u672c2（\u65b0\u7248\u672c）
\n\u2502 \u2514\u2500\u2500 model.py
\n\u2514\u2500\u2500 config.pbtxt<\/p>\n
\u5728\u5ba2\u6237\u7aef\u8c03\u7528\u65f6，\u53ef\u4ee5\u6307\u5b9a\u7248\u672c\u53f7：<\/p>\n
# \u8c03\u7528\u7279\u5b9a\u7248\u672c\u7684\u6a21\u578b
\nresponse = client.infer(
\n model_name="frcrn",
\n model_version="2", # \u6307\u5b9a\u7248\u672c\u53f7
\n inputs=inputs,
\n outputs=outputs
\n)<\/p>\n
\u5982\u679c\u4e0d\u6307\u5b9a\u7248\u672c\u53f7，Triton\u4f1a\u81ea\u52a8\u4f7f\u7528\u6700\u65b0\u7684\u7248\u672c\u3002<\/p>\n
\u5982\u679c\u4f60\u6253\u7b97\u5728\u751f\u4ea7\u73af\u5883\u4f7f\u7528Triton，\u8fd9\u91cc\u6709\u4e00\u4e9b\u5efa\u8bae：<\/p>\n
# \u5728config.pbtxt\u4e2d\u8c03\u6574
\ndynamic_batching {
\n preferred_batch_size: [4, 8, 16]
\n max_queue_delay_microseconds: 100
\n}<\/p>\n
# \u521b\u5efa\u6a21\u578b\u6d41\u6c34\u7ebf
\nname: "audio_pipeline"
\nplatform: "ensemble"<\/p>\n
input [
\n {
\n name: "audio_input"
\n data_type: TYPE_FP32
\n dims: [ -1, 1 ]
\n }
\n]<\/p>\n
output [
\n {
\n name: "final_output"
\n data_type: TYPE_FP32
\n dims: [ -1, 1 ]
\n }
\n]<\/p>\n
ensemble_scheduling {
\n step [
\n {
\n model_name: "frcrn"
\n model_version: -1
\n input_map {
\n key: "audio_input"
\n value: "audio_input"
\n }
\n output_map {
\n key: "audio_output"
\n value: "denoised_audio"
\n }
\n },
\n {
\n model_name: "vad" # \u8bed\u97f3\u6d3b\u52a8\u68c0\u6d4b
\n model_version: -1
\n input_map {
\n key: "audio_input"
\n value: "denoised_audio"
\n }
\n output_map {
\n key: "vad_output"
\n value: "final_output"
\n }
\n }
\n ]
\n}<\/p>\n
docker run –gpus=all –rm \\\\
\n -p 8000:8000 -p 8001:8001 -p 8002:8002 \\\\
\n -v ~\/triton_models:\/models \\\\
\n nvcr.io\/nvidia\/tritonserver:22.12-py3 \\\\
\n tritonserver –model-repository=\/models \\\\
\n –metrics-port=8002 \\\\
\n –allow-metrics=true \\\\
\n –allow-gpu-metrics=true<\/p>\n
# \u83b7\u53d6\u6027\u80fd\u6307\u6807
\ncurl localhost:8002\/metrics<\/p>\n
# \u4f7f\u7528Prometheus + Grafana\u53ef\u89c6\u5316<\/p>\n
\u5bf9\u4e8e\u751f\u4ea7\u73af\u5883，\u5efa\u8bae\u4f7f\u7528Kubernetes\u90e8\u7f72：<\/p>\n
# triton-deployment.yaml
\napiVersion: apps\/v1
\nkind: Deployment
\nmetadata:
\n name: triton-server
\nspec:
\n replicas: 3 # 3\u4e2a\u526f\u672c
\n selector:
\n matchLabels:
\n app: triton
\n template:
\n metadata:
\n labels:
\n app: triton
\n spec:
\n containers:
\n – name: triton
\n image: nvcr.io\/nvidia\/tritonserver:22.12-py3
\n args: ["tritonserver", "–model-repository=\/models"]
\n ports:
\n – containerPort: 8000
\n – containerPort: 8001
\n – containerPort: 8002
\n volumeMounts:
\n – name: models
\n mountPath: \/models
\n resources:
\n limits:
\n nvidia.com\/gpu: 1
\n volumes:
\n – name: models
\n persistentVolumeClaim:
\n claimName: models-pvc<\/p>\n
\u901a\u8fc7\u8fd9\u4e2a\u6559\u7a0b，\u6211\u4eec\u5b8c\u6210\u4e86FRCRN\u8bed\u97f3\u964d\u566a\u6a21\u578b\u5728NVIDIA Triton\u4e0a\u7684\u90e8\u7f72\u3002\u8ba9\u6211\u4eec\u56de\u987e\u4e00\u4e0b\u5173\u952e\u6b65\u9aa4：<\/p>\n
Triton\u7684\u771f\u6b63\u5a01\u529b\u5728\u4e8e\u5b83\u7684\u6269\u5c55\u6027\u3002\u4eca\u5929\u4f60\u90e8\u7f72\u4e86FRCRN，\u660e\u5929\u5c31\u53ef\u4ee5\u7528\u540c\u6837\u7684\u65b9\u5f0f\u90e8\u7f72\u5176\u4ed6\u4efb\u4f55\u6a21\u578b\u3002\u6240\u6709\u7684\u6a21\u578b\u90fd\u901a\u8fc7\u7edf\u4e00\u7684\u63a5\u53e3\u63d0\u4f9b\u670d\u52a1，\u5927\u5927\u964d\u4f4e\u4e86\u8fd0\u7ef4\u590d\u6742\u5ea6\u3002<\/p>\n
\u5982\u679c\u4f60\u5728\u5b9e\u8df5\u8fc7\u7a0b\u4e2d\u9047\u5230\u95ee\u9898，\u6216\u8005\u60f3\u4e86\u89e3\u66f4\u9ad8\u7ea7\u7684\u529f\u80fd，\u6211\u5efa\u8bae：<\/p>\n
\u5e0c\u671b\u8fd9\u4e2a\u6559\u7a0b\u80fd\u5e2e\u4f60\u8fc8\u51fa\u591a\u6a21\u578b\u7edf\u4e00\u7ba1\u7406\u7684\u7b2c\u4e00\u6b65\u3002\u5728\u5b9e\u9645\u9879\u76ee\u4e2d，\u4f60\u53ef\u80fd\u4f1a\u9047\u5230\u5404\u79cd\u6311\u6218，\u4f46\u6709\u4e86Triton\u8fd9\u4e2a\u5f3a\u5927\u7684\u5de5\u5177，\u81f3\u5c11\u6a21\u578b\u90e8\u7f72\u8fd9\u90e8\u5206\u4f1a\u8f7b\u677e\u5f88\u591a\u3002<\/p>\n
\u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n
\u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
FRCRN\u90e8\u7f72\u6559\u7a0b：\u4f7f\u7528NVIDIA Triton\u63a8\u7406\u670d\u52a1\u5668\u7edf\u4e00\u7ba1\u7406\u591a\u6a21\u578b
\n\u4f60\u662f\u4e0d\u662f\u4e5f\u9047\u5230\u8fc7\u8fd9\u6837\u7684\u70e6\u607c？\u56e2\u961f\u91cc\u90e8\u7f72\u4e86\u597d\u51e0\u4e2aAI\u6a21\u578b，\u6bcf\u4e2a\u6a21\u578b\u90fd\u6709\u81ea\u5df1\u7684\u73af\u5883\u4f9d\u8d56\u3001\u542f\u52a8\u811a\u672c\u548cAPI\u63a5\u53e3，\u7ba1\u7406\u8d77\u6765\u50cf\u4e00\u76d8\u6563\u6c99\u3002\u4eca\u5929\u8981\u90e8\u7f72\u4e00\u4e2a\u8bed\u97f3\u964d\u566a\u6a21\u578b，\u660e\u5929\u8981\u4e0a\u7ebf\u4e00\u4e2a\u56fe\u50cf\u8bc6\u522b\u670d\u52a1，\u540e\u5929\u53ef\u80fd\u53c8\u8981\u52a0\u4e00\u4e2a\u6587\u672c\u751f\u6210\u5de5\u5177\u3002\u6bcf\u6b21\u90fd\u8981\u5355\u72ec\u914d\u7f6e，\u4e0d\u4ec5\u6548\u7387\u4f4e\u4e0b，\u8fd8\u5bb9\u6613\u51fa\u9519\u3002
\n\u5982\u679c\u4f60\u6b63\u5728\u4e3a\u591a\u6a21\u578b\u7ba1\u7406\u5934\u75bc&<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[9281,9282,9280],"topic":[],"class_list":["post-83056","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-nvidia-triton","tag-9280"],"yoast_head":"\n