refactor data feed order

.notebook/dataloader.ipynb

@@ -338,7 +338,7 @@
     }
    ],
    "source": [
-    "for idx, (audio, text, audio_len, text_len) in enumerate(batch_reader()):\n",
+    "for idx, (audio, audio_len, text, text_len) in enumerate(batch_reader()):\n",
     "    print('test', text)\n",
     "    print(\"test raw\", ''.join( chr(i) for i in text[0][:int(text_len[0])] ))\n",
     "    print(\"test raw\", ''.join( chr(i) for i in text[-1][:int(text_len[-1])] ))\n",

.notebook/train_test.ipynb

@@ -249,7 +249,7 @@
     }
    ],
    "source": [
-    "    for idx, (audio, text, audio_len, text_len) in enumerate(batch_reader()):\n",
+    "    for idx, (audio, audio_len, text, text_len) in enumerate(batch_reader()):\n",
     "        print('test', text)\n",
     "        print(\"test raw\", ''.join(batch_reader.dataset.vocab_list[i] for i in text[0]))\n",
     "        print(\"test raw\", ''.join(batch_reader.dataset.vocab_list[i] for i in text[-1]))\n",
@@ -835,7 +835,7 @@
     "\n",
     "        return logits, probs, audio_len\n",
     "\n",
-    "    def forward(self, audio, text, audio_len, text_len):\n",
+    "    def forward(self, audio, audio_len, text, text_len):\n",
     "        \"\"\"\n",
     "        audio: shape [B, D, T]\n",
     "        text: shape [B, T]\n",
@@ -877,10 +877,10 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "audio, text, audio_len, text_len = None, None, None, None\n",
+    "audio, audio_len, text, text_len = None, None, None, None\n",
     "\n",
     "for idx, inputs in enumerate(batch_reader):\n",
-    "    audio, text, audio_len, text_len = inputs\n",
+    "    audio, audio_len, text, text_len = inputs\n",
     "#     print(idx)\n",
     "#     print('a', audio.shape, audio.place)\n",
     "#     print('t', text)\n",
@@ -960,7 +960,7 @@
     }
    ],
    "source": [
-    "outputs = dp_model(audio, text, audio_len, text_len)\n",
+    "outputs = dp_model(audio, audio_len, text, text_len)\n",
     "logits, _, logits_len = outputs\n",
     "print('logits len', logits_len)\n",
     "loss = loss_fn.forward(logits, text, logits_len, text_len)\n",

deepspeech/__init__.py

@@ -222,6 +222,31 @@ if not hasattr(paddle.Tensor, 'relu'):
     logger.warn("register user relu to paddle.Tensor, remove this when fixed!")
     setattr(paddle.Tensor, 'relu', paddle.nn.functional.relu)
 
+
+def type_as(x: paddle.Tensor, other: paddle.Tensor) -> paddle.Tensor:
+    return x.astype(other.dtype)
+
+
+if not hasattr(paddle.Tensor, 'type_as'):
+    logger.warn(
+        "register user type_as to paddle.Tensor, remove this when fixed!")
+    setattr(paddle.Tensor, 'type_as', type_as)
+
+
+def to(x: paddle.Tensor, *args, **kwargs) -> paddle.Tensor:
+    assert len(args) == 1
+    if isinstace(args[0], str):  # dtype
+        return x.astype(args[0])
+    elif isinstance(args[0], paddle.Tensor):  #Tensor
+        return x.astype(args[0].dtype)
+    else:  # Device
+        return x
+
+
+if not hasattr(paddle.Tensor, 'to'):
+    logger.warn("register user to to paddle.Tensor, remove this when fixed!")
+    setattr(paddle.Tensor, 'to', to)
+
 ########### hcak paddle.nn.functional #############
 
 

deepspeech/exps/deepspeech2/bin/tune.py

@@ -103,7 +103,7 @@ def tune(config, args):
                 trans.append(''.join([chr(i) for i in ids]))
             return trans
 
-        audio, text, audio_len, text_len = infer_data
+        audio, audio_len, text, text_len = infer_data
         target_transcripts = ordid2token(text, text_len)
         num_ins += audio.shape[0]
 

deepspeech/io/collator.py

@@ -17,6 +17,7 @@ import numpy as np
 from collections import namedtuple
 
 from deepspeech.io.utility import pad_sequence
+from deepspeech.utils.tensor_utils import IGNORE_ID
 
 logger = logging.getLogger(__name__)
 
@@ -29,10 +30,6 @@ class SpeechCollator():
         Padding audio features with zeros to make them have the same shape (or
         a user-defined shape) within one bach.
 
-        If ``padding_to`` is -1, the maximun shape in the batch will be used
-        as the target shape for padding. Otherwise, `padding_to` will be the
-        target shape (only refers to the second axis).
-
         if ``is_training`` is True, text is token ids else is raw string.
         """
         self._is_training = is_training
@@ -48,8 +45,8 @@ class SpeechCollator():
         Returns:
             tuple(audio, text, audio_lens, text_lens): batched data.
                 audio : (B, Tmax, D)
-                text : (B, Umax)
                 audio_lens: (B)
+                text : (B, Umax)
                 text_lens: (B)
         """
         audios = []
@@ -76,7 +73,9 @@ class SpeechCollator():
 
         padded_audios = pad_sequence(
             audios, padding_value=0.0).astype(np.float32)  #[B, T, D]
-        padded_texts = pad_sequence(texts, padding_value=-1).astype(np.int32)
         audio_lens = np.array(audio_lens).astype(np.int64)
+        # (TODO:Hui Zhang) ctc loss does not support int64 labels
+        padded_texts = pad_sequence(
+            texts, padding_value=IGNORE_ID).astype(np.int32)
         text_lens = np.array(text_lens).astype(np.int64)
-        return padded_audios, padded_texts, audio_lens, text_lens
+        return padded_audios, audio_lens, padded_texts, text_lens

deepspeech/models/deepspeech2.py

@@ -168,13 +168,13 @@ class DeepSpeech2Model(nn.Layer):
             dropout_rate=0.0,
             reduction=True)
 
-    def forward(self, audio, text, audio_len, text_len):
+    def forward(self, audio, audio_len, text, text_len):
         """Compute Model loss
 
         Args:
             audio (Tenosr): [B, T, D]
-            text (Tensor): [B, U]
             audio_len (Tensor): [B]
+            text (Tensor): [B, U]
             text_len (Tensor): [B]
 
         Returns:

deepspeech/models/u2.py

@@ -28,7 +28,12 @@ from paddle import jit
 from paddle import nn
 from paddle.nn import functional as F
 from paddle.nn import initializer as I
+
 from paddle.nn.utils.rnn import pad_sequence
+from deepspeech.modules.mask import make_pad_mask
+from deepspeech.modules.mask import mask_finished_preds
+from deepspeech.modules.mask import mask_finished_scores
+from deepspeech.modules.mask import subsequent_mask
 
 from deepspeech.modules.cmvn import GlobalCMVN
 from deepspeech.modules.encoder import ConformerEncoder
@@ -36,10 +41,6 @@ from deepspeech.modules.encoder import TransformerEncoder
 from deepspeech.modules.ctc import CTCDecoder
 from deepspeech.modules.decoder import TransformerDecoder
 from deepspeech.modules.label_smoothing_loss import LabelSmoothingLoss
-from deepspeech.modules.mask import make_pad_mask
-from deepspeech.modules.mask import mask_finished_preds
-from deepspeech.modules.mask import mask_finished_scores
-from deepspeech.modules.mask import subsequent_mask
 
 from deepspeech.utils import checkpoint
 from deepspeech.utils import layer_tools
@@ -101,6 +102,8 @@ class U2Model(nn.Module):
             speech_lengths: (Batch, )
             text: (Batch, Length)
             text_lengths: (Batch,)
+        Returns:
+            total_loss, attention_loss, ctc_loss
         """
         assert text_lengths.dim() == 1, text_lengths.shape
         # Check that batch_size is unified
@@ -109,21 +112,19 @@ class U2Model(nn.Module):
                                          text.shape, text_lengths.shape)
         # 1. Encoder
         encoder_out, encoder_mask = self.encoder(speech, speech_lengths)
-        encoder_out_lens = encoder_mask.squeeze(1).sum(1)
+        encoder_out_lens = encoder_mask.squeeze(1).sum(1)  #[B, 1, T] -> [B]
 
         # 2a. Attention-decoder branch
+        loss_att = None
         if self.ctc_weight != 1.0:
             loss_att, acc_att = self._calc_att_loss(encoder_out, encoder_mask,
                                                     text, text_lengths)
-        else:
-            loss_att = None
 
         # 2b. CTC branch
+        loss_ctc = None
         if self.ctc_weight != 0.0:
             loss_ctc = self.ctc(encoder_out, encoder_out_lens, text,
                                 text_lengths)
-        else:
-            loss_ctc = None
 
         if loss_ctc is None:
             loss = loss_att
@@ -139,6 +140,17 @@ class U2Model(nn.Module):
             encoder_mask: paddle.Tensor,
             ys_pad: paddle.Tensor,
             ys_pad_lens: paddle.Tensor, ) -> Tuple[paddle.Tensor, float]:
+        """Calc attention loss.
+
+        Args:
+            encoder_out (paddle.Tensor): [B, Tmax, D]
+            encoder_mask (paddle.Tensor): [B, 1, Tmax]
+            ys_pad (paddle.Tensor): [B, Umax]
+            ys_pad_lens (paddle.Tensor): [B]
+
+        Returns:
+            Tuple[paddle.Tensor, float]: attention_loss, accuracy rate
+        """
         ys_in_pad, ys_out_pad = add_sos_eos(ys_pad, self.sos, self.eos,
                                             self.ignore_id)
         ys_in_lens = ys_pad_lens + 1
@@ -163,6 +175,20 @@ class U2Model(nn.Module):
             num_decoding_left_chunks: int=-1,
             simulate_streaming: bool=False,
     ) -> Tuple[paddle.Tensor, paddle.Tensor]:
+        """Encoder pass.
+
+        Args:
+            speech (paddle.Tensor): [B, Tmax, D]
+            speech_lengths (paddle.Tensor): [B]
+            decoding_chunk_size (int, optional): chuck size. Defaults to -1.
+            num_decoding_left_chunks (int, optional): nums chunks. Defaults to -1.
+            simulate_streaming (bool, optional): streaming or not. Defaults to False.
+
+        Returns:
+            Tuple[paddle.Tensor, paddle.Tensor]: 
+                encoder hiddens (B, Tmax, D), 
+                encoder hiddens mask (B, 1, Tmax).
+        """
         # Let's assume B = batch_size
         # 1. Encoder
         if simulate_streaming and decoding_chunk_size > 0:
@@ -205,7 +231,7 @@ class U2Model(nn.Module):
         """
         assert speech.shape[0] == speech_lengths.shape[0]
         assert decoding_chunk_size != 0
-        device = speech.device
+        device = speech.place
         batch_size = speech.shape[0]
 
         # Let's assume B = batch_size and N = beam_size
@@ -223,14 +249,14 @@ class U2Model(nn.Module):
             1, beam_size, 1, 1).view(running_size, 1,
                                      maxlen)  # (B*N, 1, max_len)
 
-        hyps = torch.ones(
+        hyps = paddle.ones(
-            [running_size, 1], dtype=torch.long,
+            [running_size, 1], dtype=paddle.long).fill_(self.sos)  # (B*N, 1)
-            device=device).fill_(self.sos)  # (B*N, 1)
+        # log scale score
-        scores = paddle.tensor(
+        scores = paddle.to_tensor(
-            [0.0] + [-float('inf')] * (beam_size - 1), dtype=torch.float)
+            [0.0] + [-float('inf')] * (beam_size - 1), dtype=paddle.float)
         scores = scores.to(device).repeat([batch_size]).unsqueeze(1).to(
             device)  # (B*N, 1)
-        end_flag = torch.zeros_like(scores, dtype=torch.bool, device=device)
+        end_flag = paddle.zeros_like(scores, dtype=paddle.bool)  # (B*N, 1)
         cache: Optional[List[paddle.Tensor]] = None
         # 2. Decoder forward step by step
         for i in range(1, maxlen + 1):

deepspeech/modules/decoder.py

@@ -152,12 +152,12 @@ class TransformerDecoder(nn.Module):
             memory: encoded memory, float32  (batch, maxlen_in, feat)
             memory_mask: encoded memory mask, (batch, 1, maxlen_in)
             tgt: input token ids, int64 (batch, maxlen_out)
-            tgt_mask: input token mask,  (batch, maxlen_out)
+            tgt_mask: input token mask,  (batch, maxlen_out, maxlen_out)
                       dtype=paddle.bool
             cache: cached output list of (batch, max_time_out-1, size)
         Returns:
             y, cache: NN output value and cache per `self.decoders`.
-            y.shape` is (batch, maxlen_out, token)
+                y.shape` is (batch, token)
         """
         x, _ = self.embed(tgt)
         new_cache = []

deepspeech/modules/loss.py

@@ -88,7 +88,10 @@ class LabelSmoothingLoss(nn.Layer):
             size (int): the number of class
             padding_idx (int): padding class id which will be ignored for loss
             smoothing (float): smoothing rate (0.0 means the conventional CE)
-            normalize_length (bool): True, normalize loss by sequence length; False, normalize loss by batch size. Defaults to False.
+            normalize_length (bool): 
+                True, normalize loss by sequence length; 
+                False, normalize loss by batch size.
+                Defaults to False.
         """
         super().__init__()
         self.size = size
@@ -103,6 +106,7 @@ class LabelSmoothingLoss(nn.Layer):
         The model outputs and data labels tensors are flatten to
         (batch*seqlen, class) shape and a mask is applied to the
         padding part which should not be calculated for loss.
+        
         Args:
             x (paddle.Tensor): prediction (batch, seqlen, class)
             target (paddle.Tensor):

deepspeech/modules/mask.py

@@ -50,6 +50,52 @@ def sequence_mask(x_len, max_len=None, dtype='float32'):
     return mask
 
 
+def make_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
+    """Make mask tensor containing indices of padded part.
+    See description of make_non_pad_mask.
+    Args:
+        lengths (paddle.Tensor): Batch of lengths (B,).
+    Returns:
+        paddle.Tensor: Mask tensor containing indices of padded part.
+    Examples:
+        >>> lengths = [5, 3, 2]
+        >>> make_pad_mask(lengths)
+        masks = [[0, 0, 0, 0 ,0],
+                 [0, 0, 0, 1, 1],
+                 [0, 0, 1, 1, 1]]
+    """
+    batch_size = int(lengths.shape[0])
+    max_len = int(lengths.max())
+    seq_range = paddle.arange(0, max_len, dtype=paddle.int64)
+    seq_range_expand = seq_range.unsqueeze(0).expand([batch_size, max_len])
+    seq_length_expand = lengths.unsqueeze(-1)
+    mask = seq_range_expand >= seq_length_expand
+    return mask
+
+
+def make_non_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
+    """Make mask tensor containing indices of non-padded part.
+    The sequences in a batch may have different lengths. To enable
+    batch computing, padding is need to make all sequence in same
+    size. To avoid the padding part pass value to context dependent
+    block such as attention or convolution , this padding part is
+    masked.
+    This pad_mask is used in both encoder and decoder.
+    1 for non-padded part and 0 for padded part.
+    Args:
+        lengths (paddle.Tensor): Batch of lengths (B,).
+    Returns:
+        paddle.Tensor: mask tensor containing indices of padded part.
+    Examples:
+        >>> lengths = [5, 3, 2]
+        >>> make_non_pad_mask(lengths)
+        masks = [[1, 1, 1, 1 ,1],
+                 [1, 1, 1, 0, 0],
+                 [1, 1, 0, 0, 0]]
+    """
+    return ~make_pad_mask(lengths)
+
+
 def subsequent_mask(size: int) -> paddle.Tensor:
     """Create mask for subsequent steps (size, size).
     This mask is used only in decoder which works in an auto-regressive mode.
@@ -170,52 +216,6 @@ def add_optional_chunk_mask(xs: paddle.Tensor,
     return chunk_masks
 
 
-def make_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
-    """Make mask tensor containing indices of padded part.
-    See description of make_non_pad_mask.
-    Args:
-        lengths (paddle.Tensor): Batch of lengths (B,).
-    Returns:
-        paddle.Tensor: Mask tensor containing indices of padded part.
-    Examples:
-        >>> lengths = [5, 3, 2]
-        >>> make_pad_mask(lengths)
-        masks = [[0, 0, 0, 0 ,0],
-                 [0, 0, 0, 1, 1],
-                 [0, 0, 1, 1, 1]]
-    """
-    batch_size = int(lengths.shape[0])
-    max_len = int(lengths.max())
-    seq_range = paddle.arange(0, max_len, dtype=paddle.int64)
-    seq_range_expand = seq_range.unsqueeze(0).expand([batch_size, max_len])
-    seq_length_expand = lengths.unsqueeze(-1)
-    mask = seq_range_expand >= seq_length_expand
-    return mask
-
-
-def make_non_pad_mask(lengths: paddle.Tensor) -> paddle.Tensor:
-    """Make mask tensor containing indices of non-padded part.
-    The sequences in a batch may have different lengths. To enable
-    batch computing, padding is need to make all sequence in same
-    size. To avoid the padding part pass value to context dependent
-    block such as attention or convolution , this padding part is
-    masked.
-    This pad_mask is used in both encoder and decoder.
-    1 for non-padded part and 0 for padded part.
-    Args:
-        lengths (paddle.Tensor): Batch of lengths (B,).
-    Returns:
-        paddle.Tensor: mask tensor containing indices of padded part.
-    Examples:
-        >>> lengths = [5, 3, 2]
-        >>> make_non_pad_mask(lengths)
-        masks = [[1, 1, 1, 1 ,1],
-                 [1, 1, 1, 0, 0],
-                 [1, 1, 0, 0, 0]]
-    """
-    return ~make_pad_mask(lengths)
-
-
 def mask_finished_scores(score: paddle.Tensor,
                          flag: paddle.Tensor) -> paddle.Tensor:
     """

tests/network_test.py

@@ -46,7 +46,7 @@ if __name__ == '__main__':
         rnn_size=1024,
         use_gru=False,
         share_rnn_weights=False, )
-    logits, probs, logits_len = model(audio, text, audio_len, text_len)
+    logits, probs, logits_len = model(audio, audio_len, text, text_len)
     print('probs.shape', probs.shape)
     print("-----------------")
 
@@ -58,7 +58,7 @@ if __name__ == '__main__':
         rnn_size=1024,
         use_gru=True,
         share_rnn_weights=False, )
-    logits, probs, logits_len = model2(audio, text, audio_len, text_len)
+    logits, probs, logits_len = model2(audio, audio_len, text, text_len)
     print('probs.shape', probs.shape)
     print("-----------------")
 
@@ -70,7 +70,7 @@ if __name__ == '__main__':
         rnn_size=1024,
         use_gru=False,
         share_rnn_weights=True, )
-    logits, probs, logits_len = model3(audio, text, audio_len, text_len)
+    logits, probs, logits_len = model3(audio, audio_len, text, text_len)
     print('probs.shape', probs.shape)
     print("-----------------")
 
@@ -82,7 +82,7 @@ if __name__ == '__main__':
         rnn_size=1024,
         use_gru=True,
         share_rnn_weights=True, )
-    logits, probs, logits_len = model4(audio, text, audio_len, text_len)
+    logits, probs, logits_len = model4(audio, audio_len, text, text_len)
     print('probs.shape', probs.shape)
     print("-----------------")
 
@@ -94,6 +94,6 @@ if __name__ == '__main__':
         rnn_size=1024,
         use_gru=False,
         share_rnn_weights=False, )
-    logits, probs, logits_len = model5(audio, text, audio_len, text_len)
+    logits, probs, logits_len = model5(audio, audio_len, text, text_len)
     print('probs.shape', probs.shape)
     print("-----------------")