Merge pull request #140 from alibaba-damo-academy/dev_lzr

zhifu gao

2023-02-21 676957a1e7000a948bb2cc876ecec0f1b7768afa

Merge pull request #140 from alibaba-damo-academy/dev_lzr

support grpc+onnxruntime

 funasr/runtime/python/grpc/grpc_main_server.py

@@ -9,7 +9,8 @@
      server = grpc.server(futures.ThreadPoolExecutor(max_workers=10),
                        # interceptors=(AuthInterceptor('Bearer mysecrettoken'),)
                           )
      paraformer_pb2_grpc.add_ASRServicer_to_server(ASRServicer(args.user_allowed, args.model, args.sample_rate), server)
      paraformer_pb2_grpc.add_ASRServicer_to_server(
          ASRServicer(args.user_allowed, args.model, args.sample_rate, args.backend, args.onnx_dir), server)
      port = "[::]:" + str(args.port)
      server.add_insecure_port(port)
      server.start()
@@ -37,7 +38,18 @@
    parser.add_argument("--sample_rate",
                        type=int,
                        default=16000,
                        help="audio sample_rate from client")    
                        help="audio sample_rate from client")

    parser.add_argument("--backend",
                        type=str,
                        default="pipeline",
                        choices=("pipeline", "onnxruntime"),
                        help="backend, optional modelscope pipeline or onnxruntime")

    parser.add_argument("--onnx_dir",
                        type=str,
                        default="/nfs/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
                        help="onnx model dir")
                        



 funasr/runtime/python/grpc/grpc_server.py

@@ -3,20 +3,32 @@
import json
import time

from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
import paraformer_pb2_grpc
from paraformer_pb2 import Response
from utils.frontend import load_bytes


class ASRServicer(paraformer_pb2_grpc.ASRServicer):
    def __init__(self, user_allowed, model, sample_rate):
    def __init__(self, user_allowed, model, sample_rate, backend, onnx_dir):
        print("ASRServicer init")
        self.backend = backend
        self.init_flag = 0
        self.client_buffers = {}
        self.client_transcription = {}
        self.auth_user = user_allowed.split("|")
        self.inference_16k_pipeline = pipeline(task=Tasks.auto_speech_recognition, model=model)
        if self.backend == "pipeline":
            try:
                from modelscope.pipelines import pipeline
                from modelscope.utils.constant import Tasks
            except ImportError:
                raise ImportError(f"Please install modelscope")
            self.inference_16k_pipeline = pipeline(task=Tasks.auto_speech_recognition, model=model)
        elif self.backend == "onnxruntime":
            try:
                from paraformer_onnx import Paraformer
            except ImportError:
                raise ImportError(f"Please install onnxruntime requirements, reference https://github.com/alibaba-damo-academy/FunASR/tree/main/funasr/runtime/python/onnxruntime/rapid_paraformer")
            self.inference_16k_pipeline = Paraformer(model_dir=onnx_dir)
        self.sample_rate = sample_rate

    def clear_states(self, user):
@@ -90,12 +102,16 @@
                        result["text"] = ""
                        print ("user: %s , delay(ms): %s, info: %s " % (req.user, delay_str, "waiting_for_more_voice"))
                        yield Response(sentence=json.dumps(result), user=req.user, action="waiting", language=req.language)
                    else:                           
                        asr_result = self.inference_16k_pipeline(audio_in=tmp_data, audio_fs = self.sample_rate)
                        if "text" in asr_result:
                            asr_result = asr_result['text']
                        else:
                            asr_result = ""
                    else:
                        if self.backend == "pipeline":
                            asr_result = self.inference_16k_pipeline(audio_in=tmp_data, audio_fs = self.sample_rate)
                            if "text" in asr_result:
                                asr_result = asr_result['text']
                            else:
                                asr_result = ""
                        elif self.backend == "onnxruntime":
                            array = load_bytes(tmp_data)
                            asr_result = self.inference_16k_pipeline(array)[0]
                        end_time = int(round(time.time() * 1000))
                        delay_str = str(end_time - begin_time)
                        print ("user: %s , delay(ms): %s, text: %s " % (req.user, delay_str, asr_result))

 funasr/runtime/python/grpc/paraformer_onnx.py

New file
@@ -0,0 +1 @@
../onnxruntime/paraformer/rapid_paraformer/paraformer_onnx.py

 funasr/runtime/python/grpc/utils

New file
@@ -0,0 +1 @@
../onnxruntime/paraformer/rapid_paraformer/utils

 funasr/runtime/python/onnxruntime/rapid_paraformer/utils/frontend.py

@@ -134,3 +134,18 @@
        vars = np.array(vars_list).astype(np.float64)
        cmvn = np.array([means, vars])
        return cmvn

def load_bytes(input):
    middle_data = np.frombuffer(input, dtype=np.int16)
    middle_data = np.asarray(middle_data)
    if middle_data.dtype.kind not in 'iu':
        raise TypeError("'middle_data' must be an array of integers")
    dtype = np.dtype('float32')
    if dtype.kind != 'f':
        raise TypeError("'dtype' must be a floating point type")

    i = np.iinfo(middle_data.dtype)
    abs_max = 2 ** (i.bits - 1)
    offset = i.min + abs_max
    array = np.frombuffer((middle_data.astype(dtype) - offset) / abs_max, dtype=np.float32)
    return array