benchmark

游雁

2023-03-17 745ac1b175baf157aeb843c373b2221d5ecb9a21

benchmark

 funasr/runtime/python/benchmark_onnx.md

@@ -1,6 +1,8 @@
Benchmark [Paraformer-large](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) based on Aishell1 test set , the total audio duration is 36108.919 seconds.
# Benchmark 

(Note: The service has been fully warm up.)
Aishell1 test set , the total audio duration is 36108.919 seconds. (Note: The service has been fully warm up.)

## [Paraformer-large](https://www.modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/summary) 

 ### Intel(R) Xeon(R) Platinum 8369B CPU @ 2.90GHz   16core-32processor    with avx512_vnni

@@ -46,4 +48,6 @@
|  64 (onnx fp32)  |        153         | 0.0042 |    235.9     |
|  64 (onnx int8)  |        103         | 0.0029 |    349.9     |
|  96 (onnx fp32)  |        146         | 0.0041 |    247.0     |
|  96 (onnx int8)  |        108         | 0.0030 |    334.1     |
|  96 (onnx int8)  |        108         | 0.0030 |    334.1     |

## [Paraformer](https://modelscope.cn/models/damo/speech_paraformer_asr_nat-zh-cn-16k-common-vocab8358-tensorflow1/summary)

 funasr/runtime/python/utils/test_rtf.py

@@ -2,17 +2,24 @@
import time
import sys
import librosa
backend=sys.argv[1]
model_dir=sys.argv[2]
wav_file=sys.argv[3]


import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--model_dir', type=str, required=True)
parser.add_argument('--backend', type=str, default='onnx', help='["onnx", "torch"]')
parser.add_argument('--wav_file', type=int, default=0, help='amp fallback number')
parser.add_argument('--quantize', type=bool, default=False, help='quantized model')
args = parser.parse_args()


from torch_paraformer import Paraformer
if backend == "onnxruntime":
if args.backend == "onnxruntime":
    from rapid_paraformer import Paraformer
    
model = Paraformer(model_dir, batch_size=1, device_id="-1")
model = Paraformer(args.model_dir, batch_size=1, quantize=args.quantize)

wav_file_f = open(wav_file, 'r')
wav_file_f = open(args.wav_file, 'r')
wav_files = wav_file_f.readlines()

# warm-up

 funasr/runtime/python/utils/test_rtf.sh

@@ -5,28 +5,31 @@
backend=libtorch
model_dir="/nfs/zhifu.gzf/export/damo/amp_int8/libtorch"
tag=${backend}_fp32
quantize='False'
!

:<<!
backend=libtorch
model_dir="/nfs/zhifu.gzf/export/damo/amp_int8/libtorch_fb20"
tag=${backend}_amp_fb20
quantize='True'
!

:<<!
backend=onnxruntime
model_dir="/nfs/zhifu.gzf/export/damo/amp_int8/onnx"
tag=${backend}_fp32
quantize='False'
!

:<<!
backend=onnxruntime
model_dir="/nfs/zhifu.gzf/export/damo/amp_int8/onnx_dynamic"
tag=${backend}_fp32
quantize='True'
!

#scp=/nfs/haoneng.lhn/funasr_data/aishell-1/data/test/wav.scp
scp="/nfs/zhifu.gzf/data_debug/test/wav_1500.scp"
scp=/nfs/haoneng.lhn/funasr_data/aishell-1/data/test/wav.scp
local_scp_dir=/nfs/zhifu.gzf/data_debug/test/${tag}/split$nj

rtf_tool=test_rtf.py
@@ -45,7 +48,7 @@
for JOB in $(seq ${nj}); do
  {
    core_id=`expr $JOB - 1`
    taskset -c ${core_id} python ${rtf_tool} ${backend} ${model_dir} ${local_scp_dir}/wav.$JOB.scp &> ${local_scp_dir}/log.$JOB.txt
    taskset -c ${core_id} python ${rtf_tool} --backend ${backend} --model_dir ${model_dir} --wav_file ${local_scp_dir}/wav.$JOB.scp --quantize ${quantize} &> ${local_scp_dir}/log.$JOB.txt
  }&

done