funasr/runtime/triton_gpu/README.md
@@ -8,8 +8,8 @@ pretrained_model_dir=$(pwd)/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch cp $pretrained_model_dir/tokens.txt ./model_repo_paraformer_large_offline/scoring/ cp $pretrained_model_dir/am.mvn ./model_repo_paraformer_large_offline/feature_extractor/ cp $pretrained_model_dir/config.yaml ./model_repo_paraformer_large_offline/feature_extractor/ # Refer here to get model.onnx (https://github.com/alibaba-damo-academy/FunASR/blob/main/funasr/export/README.md) cp <exported_onnx_dir>/model.onnx ./model_repo_paraformer_large_offline/encoder/1/ @@ -33,10 +33,9 @@ `-- scoring |-- 1 | `-- model.py |-- config.pbtxt `-- tokens.txt `-- config.pbtxt 8 directories, 10 files 8 directories, 9 files ``` 2. Follow below instructions to launch triton server funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/1/model.py
@@ -229,22 +229,24 @@ if key == "config_path": with open(str(value), 'rb') as f: config = yaml.load(f, Loader=yaml.Loader) if key == "cmvn_path": cmvn_path = str(value) opts = kaldifeat.FbankOptions() opts.frame_opts.dither = 1.0 # TODO: 0.0 or 1.0 opts.frame_opts.window_type = config['WavFrontend']['frontend_conf']['window'] opts.mel_opts.num_bins = int(config['WavFrontend']['frontend_conf']['n_mels']) opts.frame_opts.frame_shift_ms = float(config['WavFrontend']['frontend_conf']['frame_shift']) opts.frame_opts.frame_length_ms = float(config['WavFrontend']['frontend_conf']['frame_length']) opts.frame_opts.samp_freq = int(config['WavFrontend']['frontend_conf']['fs']) opts.frame_opts.window_type = config['frontend_conf']['window'] opts.mel_opts.num_bins = int(config['frontend_conf']['n_mels']) opts.frame_opts.frame_shift_ms = float(config['frontend_conf']['frame_shift']) opts.frame_opts.frame_length_ms = float(config['frontend_conf']['frame_length']) opts.frame_opts.samp_freq = int(config['frontend_conf']['fs']) opts.device = torch.device(self.device) self.opts = opts self.feature_extractor = Fbank(self.opts) self.feature_size = opts.mel_opts.num_bins self.frontend = WavFrontend( cmvn_file=config['WavFrontend']['cmvn_file'], **config['WavFrontend']['frontend_conf']) cmvn_file=cmvn_path, **config['frontend_conf']) def extract_feat(self, waveform_list: List[np.ndarray] funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.pbtxt
@@ -34,6 +34,10 @@ value: { string_value: "16000"} }, { key: "cmvn_path" value: { string_value: "./model_repo_paraformer_large_offline/feature_extractor/am.mvn"} }, { key: "config_path" value: { string_value: "./model_repo_paraformer_large_offline/feature_extractor/config.yaml"} } funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/feature_extractor/config.yaml
File was deleted funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/1/model.py
@@ -21,6 +21,7 @@ import json import os import yaml class TritonPythonModel: """Your Python model must use the same class name. Every Python model @@ -73,9 +74,9 @@ """ load lang_char.txt """ with open(str(vocab_file), 'r') as f: token_list = [line.strip() for line in f] return token_list with open(str(vocab_file), 'rb') as f: config = yaml.load(f, Loader=yaml.Loader) return config['token_list'] def execute(self, requests): """`execute` must be implemented in every Python model. `execute` funasr/runtime/triton_gpu/model_repo_paraformer_large_offline/scoring/config.pbtxt
@@ -23,7 +23,7 @@ }, { key: "vocabulary", value: { string_value: "./model_repo_paraformer_large_offline/scoring/tokens.txt"} value: { string_value: "./model_repo_paraformer_large_offline/feature_extractor/config.yaml"} }, { key: "lm_path"
