[ssl] support pack dataset

wenet/ssl/init_dataset.py

@@ -1,10 +1,61 @@
+from collections.abc import Callable
 from functools import partial
 import sys
+from typing import List
 
 import torch
 from torch.nn.utils.rnn import pad_sequence
+from torch.utils.data import IterDataPipe, functional_datapipe
 from wenet.dataset import processor
-from wenet.dataset.datapipes import WenetRawDatasetSource, WenetTarShardDatasetSource
+from wenet.dataset.datapipes import (WenetRawDatasetSource,
+                                     WenetTarShardDatasetSource)
+
+
+@functional_datapipe("pack_speech")
+class PackSpeechDatapipe(IterDataPipe):
+
+    def __init__(
+        self,
+        dataset: IterDataPipe,
+        length_fn: Callable,
+        merge_speech_fn: Callable,
+        max_length: int = 30000,
+    ) -> None:
+        super().__init__()
+        self.dp = dataset
+        self._iter = None
+        self.length_fn = length_fn
+        self.max_length = max_length
+        self.merge_fn = merge_speech_fn
+        self.buf = []
+
+        self.length = 0
+
+    def __iter__(self):
+        for elem in self.dp:
+            elem_length = self.length_fn(elem)
+            current_length = self.length + elem_length
+            if current_length >= self.max_length:
+                long_elem = self.merge_fn(self.buf)
+                yield long_elem
+                del self.buf
+                self.buf = []
+                self.length = 0
+            self.buf.append(elem)
+            self.length += elem_length
+        if len(self.buf) > 0:
+            yield self.merge_fn(self.buf)
+        del self.buf
+        self.buf = []
+        self.length = 0
+
+
+def cat_speech(buffer: List):
+    assert len(buffer) > 0
+    waves = [sample['wav'] for sample in buffer]
+    sample_rate = buffer[0]['sample_rate']
+    wav = torch.cat(waves, dim=1)
+    return {"wav": wav, "sample_rate": sample_rate}
 
 
 def padding(data):
@@ -39,6 +90,11 @@ def padding(data):
     return batch
 
 
+def wav_length_fn(sample):
+    wav = sample['wav']
+    return wav.size(1)
+
+
 def Dataset(data_type, data_list_file, conf=None, partition=True):
     """ Construct dataset from arguments for ssl model
 
@@ -81,6 +137,10 @@ def Dataset(data_type, data_list_file, conf=None, partition=True):
     dataset = dataset.map(
         partial(processor.singal_channel, **singal_channel_conf))
 
+    pack_conf = conf.get('pack_conf', {})
+    if pack_conf:
+        dataset = dataset.pack_speech(wav_length_fn, cat_speech,
+                                      pack_conf['max_speech_length'])
     filter_conf = conf.get('filter_conf', {})
     dataset = dataset.filter(partial(processor.filter, **filter_conf))