From 1d4ab65c8bfebaecbcb0eec0064bae9a321cad75 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期二, 14 二月 2023 16:27:37 +0800
Subject: [PATCH] export model

---
 funasr/datasets/iterable_dataset.py |  159 +++++++++++++++++++++++++++++++++++-----------------
 1 files changed, 107 insertions(+), 52 deletions(-)

diff --git a/funasr/datasets/iterable_dataset.py b/funasr/datasets/iterable_dataset.py
index bed295b..2001df9 100644
--- a/funasr/datasets/iterable_dataset.py
+++ b/funasr/datasets/iterable_dataset.py
@@ -11,14 +11,16 @@
 
 import kaldiio
 import numpy as np
-import soundfile
 import torch
+import torchaudio
 from torch.utils.data.dataset import IterableDataset
 from typeguard import check_argument_types
 import os.path
 
 from funasr.datasets.dataset import ESPnetDataset
 
+
+SUPPORT_AUDIO_TYPE_SETS = ['flac', 'mp3', 'ogg', 'opus', 'wav', 'pcm']
 
 def load_kaldi(input):
     retval = kaldiio.load_mat(input)
@@ -58,9 +60,14 @@
     array = np.frombuffer((middle_data.astype(dtype) - offset) / abs_max, dtype=np.float32)
     return array
 
+def load_pcm(input):
+    with open(input,"rb") as f:
+        bytes = f.read()
+    return load_bytes(bytes)
 
 DATA_TYPES = {
-    "sound": lambda x: soundfile.read(x)[0],
+    "sound": lambda x: torchaudio.load(x)[0][0].numpy(),
+    "pcm": load_pcm,
     "kaldi_ark": load_kaldi,
     "bytes": load_bytes,
     "waveform": lambda x: x,
@@ -98,6 +105,7 @@
                 [str, Dict[str, np.ndarray]], Dict[str, np.ndarray]
             ] = None,
             float_dtype: str = "float32",
+            fs: dict = None,
             int_dtype: str = "long",
             key_file: str = None,
     ):
@@ -113,6 +121,7 @@
         self.float_dtype = float_dtype
         self.int_dtype = int_dtype
         self.key_file = key_file
+        self.fs = fs
 
         self.debug_info = {}
         non_iterable_list = []
@@ -165,64 +174,94 @@
     def __iter__(self) -> Iterator[Tuple[Union[str, int], Dict[str, np.ndarray]]]:
         count = 0
         if len(self.path_name_type_list) != 0 and (self.path_name_type_list[0][2] == "bytes" or self.path_name_type_list[0][2] == "waveform"):
+            linenum = len(self.path_name_type_list)
             data = {}
-            value = self.path_name_type_list[0][0]
-            uid = 'utt_id'
-            name = self.path_name_type_list[0][1]
-            _type = self.path_name_type_list[0][2]
-            func = DATA_TYPES[_type]
-            array = func(value)
-            data[name] = array
+            for i in range(linenum):
+                value = self.path_name_type_list[i][0]
+                uid = 'utt_id'
+                name = self.path_name_type_list[i][1]
+                _type = self.path_name_type_list[i][2]
+                func = DATA_TYPES[_type]
+                array = func(value)
+                if self.fs is not None and (name == "speech" or name == "ref_speech"):
+                    audio_fs = self.fs["audio_fs"]
+                    model_fs = self.fs["model_fs"]
+                    if audio_fs is not None and model_fs is not None:
+                        array = torch.from_numpy(array)
+                        array = array.unsqueeze(0)
+                        array = torchaudio.transforms.Resample(orig_freq=audio_fs,
+                                                       new_freq=model_fs)(array)
+                        array = array.squeeze(0).numpy()
+                data[name] = array
 
-            if self.preprocess is not None:
-                data = self.preprocess(uid, data)
-            for name in data:
-                count += 1
-                value = data[name]
-                if not isinstance(value, np.ndarray):
-                    raise RuntimeError(
-                        f'All values must be converted to np.ndarray object '
-                        f'by preprocessing, but "{name}" is still {type(value)}.')
-                # Cast to desired type
-                if value.dtype.kind == 'f':
-                    value = value.astype(self.float_dtype)
-                elif value.dtype.kind == 'i':
-                    value = value.astype(self.int_dtype)
-                else:
-                    raise NotImplementedError(
-                        f'Not supported dtype: {value.dtype}')
-                data[name] = value
+                if self.preprocess is not None:
+                    data = self.preprocess(uid, data)
+                for name in data:
+                    count += 1
+                    value = data[name]
+                    if not isinstance(value, np.ndarray):
+                        raise RuntimeError(
+                            f'All values must be converted to np.ndarray object '
+                            f'by preprocessing, but "{name}" is still {type(value)}.')
+                    # Cast to desired type
+                    if value.dtype.kind == 'f':
+                        value = value.astype(self.float_dtype)
+                    elif value.dtype.kind == 'i':
+                        value = value.astype(self.int_dtype)
+                    else:
+                        raise NotImplementedError(
+                            f'Not supported dtype: {value.dtype}')
+                    data[name] = value
 
             yield uid, data
 
         elif len(self.path_name_type_list) != 0 and self.path_name_type_list[0][2] == "sound" and not self.path_name_type_list[0][0].lower().endswith(".scp"):
+            linenum = len(self.path_name_type_list)
             data = {}
-            value = self.path_name_type_list[0][0]
-            uid = os.path.basename(self.path_name_type_list[0][0]).split(".")[0]
-            name = self.path_name_type_list[0][1]
-            _type = self.path_name_type_list[0][2]
-            func = DATA_TYPES[_type]
-            array = func(value)
-            data[name] = array
+            for i in range(linenum):
+                value = self.path_name_type_list[i][0]
+                uid = os.path.basename(self.path_name_type_list[i][0]).split(".")[0]
+                name = self.path_name_type_list[i][1]
+                _type = self.path_name_type_list[i][2]
+                if _type == "sound":
+                    audio_type = os.path.basename(value).split(".")[1].lower()
+                    if audio_type not in SUPPORT_AUDIO_TYPE_SETS:
+                        raise NotImplementedError(
+                            f'Not supported audio type: {audio_type}')
+                    if audio_type == "pcm":
+                        _type = "pcm"
 
-            if self.preprocess is not None:
-                data = self.preprocess(uid, data)
-            for name in data:
-                count += 1
-                value = data[name]
-                if not isinstance(value, np.ndarray):
-                    raise RuntimeError(
-                        f'All values must be converted to np.ndarray object '
-                        f'by preprocessing, but "{name}" is still {type(value)}.')
-                # Cast to desired type
-                if value.dtype.kind == 'f':
-                    value = value.astype(self.float_dtype)
-                elif value.dtype.kind == 'i':
-                    value = value.astype(self.int_dtype)
-                else:
-                    raise NotImplementedError(
-                        f'Not supported dtype: {value.dtype}')
-                data[name] = value
+                func = DATA_TYPES[_type]
+                array = func(value)
+                if self.fs is not None and (name == "speech" or name == "ref_speech"):
+                    audio_fs = self.fs["audio_fs"]
+                    model_fs = self.fs["model_fs"]
+                    if audio_fs is not None and model_fs is not None:
+                        array = torch.from_numpy(array)
+                        array = array.unsqueeze(0)
+                        array = torchaudio.transforms.Resample(orig_freq=audio_fs,
+                                                               new_freq=model_fs)(array)
+                        array = array.squeeze(0).numpy()
+                data[name] = array
+
+                if self.preprocess is not None:
+                    data = self.preprocess(uid, data)
+                for name in data:
+                    count += 1
+                    value = data[name]
+                    if not isinstance(value, np.ndarray):
+                        raise RuntimeError(
+                            f'All values must be converted to np.ndarray object '
+                            f'by preprocessing, but "{name}" is still {type(value)}.')
+                    # Cast to desired type
+                    if value.dtype.kind == 'f':
+                        value = value.astype(self.float_dtype)
+                    elif value.dtype.kind == 'i':
+                        value = value.astype(self.int_dtype)
+                    else:
+                        raise NotImplementedError(
+                            f'Not supported dtype: {value.dtype}')
+                    data[name] = value
 
             yield uid, data
 
@@ -286,9 +325,25 @@
                 data = {}
                 # 2.a. Load data streamingly
                 for value, (path, name, _type) in zip(values, self.path_name_type_list):
+                    if _type == "sound":
+                        audio_type = os.path.basename(value).split(".")[1].lower()
+                        if audio_type not in SUPPORT_AUDIO_TYPE_SETS:
+                            raise NotImplementedError(
+                                f'Not supported audio type: {audio_type}')
+                        if audio_type == "pcm":
+                            _type = "pcm"
                     func = DATA_TYPES[_type]
                     # Load entry
                     array = func(value)
+                    if self.fs is not None and name == "speech":
+                        audio_fs = self.fs["audio_fs"]
+                        model_fs = self.fs["model_fs"]
+                        if audio_fs is not None and model_fs is not None:
+                            array = torch.from_numpy(array)
+                            array = array.unsqueeze(0)
+                            array = torchaudio.transforms.Resample(orig_freq=audio_fs,
+                                                                   new_freq=model_fs)(array)
+                            array = array.squeeze(0).numpy()
                     data[name] = array
                 if self.non_iterable_dataset is not None:
                     # 2.b. Load data from non-iterable dataset

--
Gitblit v1.9.1