pad_length in streaming decode

[1215thebqtic]

Hi,

I'm confused about the pad_length in set_features. Since the frames that are less than chunk_size*2+7+2*3 have already been padded in streaming_decode.py, why are we still adding 7+2*3 when initially receiving the feature here? Is the 7+2*3 necessary in set_features?

icefall/egs/librispeech/ASR/zipformer/decode_stream.py

Lines 108 to 121 in cea0dbe

	def set_features(
	self,
	features: torch.Tensor,
	tail_pad_len: int = 0,
	) -> None:
	"""Set features tensor of current utterance."""
	assert features.dim() == 2, features.dim()
	self.features = torch.nn.functional.pad(
	features,
	(0, 0, 0, self.pad_length + tail_pad_len),
	mode="constant",
	value=self.LOG_EPS,
	)
	self.num_frames = self.features.size(0)

icefall/egs/librispeech/ASR/zipformer/streaming_decode.py

Lines 472 to 484 in cea0dbe

	# Make sure the length after encoder_embed is at least 1.
	# The encoder_embed subsample features (T - 7) // 2
	# The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling
	tail_length = chunk_size * 2 + 7 + 2 * 3
	if features.size(1) < tail_length:
	pad_length = tail_length - features.size(1)
	feature_lens += pad_length
	features = torch.nn.functional.pad(
	features,
	(0, 0, 0, pad_length),
	mode="constant",
	value=LOG_EPS,
	)

Thanks!