Merge branch 'develop' of https://github.com/kaiidams/NeMoOnnxSharp i…

…nto develop

Python/export_models.py

@@ -1,5 +1,5 @@
-import nemo
 import importlib
+from omegaconf import OmegaConf
 
 
 def get_class(cls_path):
@@ -12,46 +12,15 @@ def get_class(cls_path):
 def export(cls_path: str, model_name: str):
     cls = get_class(cls_path)
     model = cls.from_pretrained(model_name)
-    print(model)
     model.export(f'{model_name}.onnx')
+    print(OmegaConf.to_yaml(model._cfg))
 
 
-
-#export("nemo_asr.models.EncDecClassificationModel")
-
-#import nemo
-#import nemo.collections.asr as nemo_asr
-#import nemo.collections.asr
-
-
-cls_path = "nemo.collections.asr.models.EncDecClassificationModel"
-#export(cls_path, 'vad_marblenet')
-cls = get_class(cls_path)
-model = cls.from_pretrained('vad_marblenet')
-print(model.preprocessor)
-print(model._cfg.preprocessor)
-import librosa
-import torch
-import numpy as np
-import struct
-from glob import glob
-for wave_file in glob("../test_data/*.wav"):
-    audio, sample_rate = librosa.load(wave_file, sr=16000)
-    audio_signal = torch.from_numpy(audio / 32768.0).to(torch.float32)
-    audio_signal = torch.unsqueeze(audio_signal, 0)
-    audio_signal_len = torch.tensor([audio.shape[0]], dtype=torch.int64)
-    processed_signal, processed_signal_len = model.preprocessor(
-        input_signal=audio_signal, length=audio_signal_len,
-    )
-    print(processed_signal.shape, processed_signal_len)
-    processed_signal = processed_signal[0]
-    bin_file = wave_file.replace('.wav', '.bin')
-    with open(bin_file, "wb") as fp:
-        fp.write(struct.pack('2i', *list(processed_signal.shape)))
-        fp.write(processed_signal.numpy().tobytes())
-
-# audio_signal, audio_signal_len = batch
-#     audio_signal, audio_signal_len = audio_signal.to(vad_model.device), audio_signal_len.to(vad_model.device)
-#     processed_signal, processed_signal_len = vad_model.preprocessor(
-#         input_signal=audio_signal, length=audio_signal_len,
-#     )
+cls_path = 'nemo.collections.asr.models.EncDecClassificationModel'
+cls_path = 'nemo.collections.asr.models.EncDecCTCModel'
+cls_path = 'nemo.collections.asr.models.EncDecClassificationModel'
+model_name = 'vad_marblenet'
+model_name = 'stt_en_quartznet15x5'
+model_name = 'stt_en_jasper10x5dr'
+model_name = 'commandrecognition_en_matchboxnet3x1x64_v2'
+export(cls_path, model_name)

Python/make_test.py

@@ -1,3 +1,6 @@
+# Copyright (c) Katsuya Iida.  All Rights Reserved.
+# See LICENSE in the project root for license information.
+
 import librosa
 import torch
 from nemo.collections.asr.modules import (
@@ -7,7 +10,7 @@
 
 
 def main():
-    wavpath = "61-70968-0000.wav"
+    wavpath = "../NemoOnnxSharp.Tests/Data/61-70968-0000.wav"
     sr = 16000
     audio_signal, sr = librosa.load(wavpath, sr=sr, mono=True)
     assert audio_signal.ndim == 1
@@ -35,7 +38,7 @@ def convert_mel_spectrogram(audio_signal, length):
         processed_signal, processed_signal_length = preprocessor(input_signal=audio_signal, length=length)
     print(processed_signal, processed_signal_length)
     print(processed_signal.shape, processed_signal_length)
-    with open('mel_spectrogram.bin', 'wb') as fp:
+    with open("../NemoOnnxSharp.Tests/Data/mel_spectrogram.bin", 'wb') as fp:
         fp.write(processed_signal[0].T.numpy().tobytes("C"))
 
 
@@ -52,7 +55,7 @@ def convert_mfcc(audio_signal, length):
         processed_signal, processed_signal_length = preprocessor(input_signal=audio_signal, length=length)
     print(processed_signal, processed_signal_length)
     print(processed_signal.shape, processed_signal_length)
-    with open('mfcc.bin', 'wb') as fp:
+    with open("../NemoOnnxSharp.Tests/Data/mfcc.bin", 'wb') as fp:
         fp.write(processed_signal[0].T.numpy().tobytes("C"))