add layer for transformer

deepspeech/exps/deepspeech2/model.py

@@ -28,7 +28,7 @@ from paddle import distributed as dist
 from paddle.io import DataLoader
 
 from deepspeech.training import Trainer
-from deepspeech.training.gradclip import MyClipGradByGlobalNorm
+from deepspeech.training.gradclip import ClipGradByGlobalNormWithLog
 
 from deepspeech.utils import mp_tools
 from deepspeech.utils import layer_tools
@@ -125,7 +125,8 @@ class DeepSpeech2Trainer(Trainer):
 
         layer_tools.print_params(model, self.logger.info)
 
-        grad_clip = MyClipGradByGlobalNorm(config.training.global_grad_clip)
+        grad_clip = ClipGradByGlobalNormWithLog(
+            config.training.global_grad_clip)
         lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
             learning_rate=config.training.lr,
             gamma=config.training.lr_decay,

deepspeech/modules/activation.py

@@ -12,9 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Union
 import logging
 import numpy as np
 import math
+from collections import OrderedDict
 
 import paddle
 from paddle import nn
@@ -23,7 +25,7 @@ from paddle.nn import initializer as I
 
 logger = logging.getLogger(__name__)
 
-__all__ = ['brelu', "softplus", "gelu_accurate", "gelu", 'Swish']
+__all__ = ['brelu', "glu"]
 
 
 def brelu(x, t_min=0.0, t_max=24.0, name=None):
@@ -33,36 +35,180 @@ def brelu(x, t_min=0.0, t_max=24.0, name=None):
     return x.maximum(t_min).minimum(t_max)
 
 
-def softplus(x):
-    """Softplus function."""
-    if hasattr(paddle.nn.functional, 'softplus'):
-        #return paddle.nn.functional.softplus(x.float()).type_as(x)
-        return paddle.nn.functional.softplus(x)
-    else:
-        raise NotImplementedError
+# def softplus(x):
+#     """Softplus function."""
+#     if hasattr(paddle.nn.functional, 'softplus'):
+#         #return paddle.nn.functional.softplus(x.float()).type_as(x)
+#         return paddle.nn.functional.softplus(x)
+#     else:
+#         raise NotImplementedError
 
+# def gelu_accurate(x):
+#     """Gaussian Error Linear Units (GELU) activation."""
+#     # [reference] https://github.com/pytorch/fairseq/blob/e75cff5f2c1d62f12dc911e0bf420025eb1a4e33/fairseq/modules/gelu.py
+#     if not hasattr(gelu_accurate, "_a"):
+#         gelu_accurate._a = math.sqrt(2 / math.pi)
+#     return 0.5 * x * (1 + paddle.tanh(gelu_accurate._a *
+#                                       (x + 0.044715 * paddle.pow(x, 3))))
 
-def gelu_accurate(x):
-    """Gaussian Error Linear Units (GELU) activation."""
-    # [reference] https://github.com/pytorch/fairseq/blob/e75cff5f2c1d62f12dc911e0bf420025eb1a4e33/fairseq/modules/gelu.py
-    if not hasattr(gelu_accurate, "_a"):
-        gelu_accurate._a = math.sqrt(2 / math.pi)
-    return 0.5 * x * (1 + paddle.tanh(gelu_accurate._a *
-                                      (x + 0.044715 * paddle.pow(x, 3))))
+# def gelu(x):
+#     """Gaussian Error Linear Units (GELU) activation."""
+#     if hasattr(nn.functional, 'gelu'):
+#         #return nn.functional.gelu(x.float()).type_as(x)
+#         return nn.functional.gelu(x)
+#     else:
+#         return x * 0.5 * (1.0 + paddle.erf(x / math.sqrt(2.0)))
 
 
-def gelu(x):
-    """Gaussian Error Linear Units (GELU) activation."""
-    if hasattr(torch.nn.functional, 'gelu'):
-        #return torch.nn.functional.gelu(x.float()).type_as(x)
-        return torch.nn.functional.gelu(x)
+# TODO(Hui Zhang): remove this activation
+def glu(x, dim=-1):
+    """The gated linear unit (GLU) activation."""
+    if hasattr(nn.functional, 'glu'):
+        return nn.functional.glu(x)
     else:
-        return x * 0.5 * (1.0 + paddle.erf(x / math.sqrt(2.0)))
+        a, b = x.split(2, axis=dim)
+        act_b = F.sigmoid(b)
+        return a * act_b
+
+
+# TODO(Hui Zhang): remove this activation
+if not hasattr(nn.functional, 'glu'):
+    setattr(nn.functional, 'glu', glu)
+
+
+# TODO(Hui Zhang): remove this activation
+class GLU(nn.Layer):
+    """Gated Linear Units (GLU) Layer"""
+
+    def __init__(self, dim: int=-1):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, xs):
+        return glu(xs, dim=self.dim)
+
+
+class LinearGLUBlock(nn.Layer):
+    """A linear Gated Linear Units (GLU) block."""
+
+    def __init__(self, idim: int):
+        """ GLU.
+        Args:
+            idim (int): input and output dimension
+        """
+        super().__init__()
+        self.fc = nn.Linear(idim, idim * 2)
+
+    def forward(self, xs):
+        return glu(self.fc(xs), dim=-1)
+
+
+# TODO(Hui Zhang): remove this Layer
+class ConstantPad2d(nn.Layer):
+    """Pads the input tensor boundaries with a constant value.
+    For N-dimensional padding, use paddle.nn.functional.pad().
+    """
+
+    def __init__(self, padding: Union[tuple, list, int], value: float):
+        """
+        Args:
+            paddle ([tuple]): the size of the padding. 
+                If is int, uses the same padding in all boundaries. 
+                If a 4-tuple, uses (padding_left, padding_right, padding_top, padding_bottom)
+            value ([flaot]): pad value
+        """
+        self.padding = padding if isinstance(padding,
+                                             [tuple, list]) else [padding] * 4
+        self.value = value
+
+    def forward(self, xs: paddle.Tensor) -> paddle.Tensor:
+        return nn.functional.pad(
+            xs,
+            self.padding,
+            mode='constant',
+            value=self.value,
+            data_format='NCHW')
+
+
+class ConvGLUBlock(nn.Layer):
+    def __init__(self, kernel_size, in_ch, out_ch, bottlececk_dim=0,
+                 dropout=0.):
+        """A convolutional Gated Linear Units (GLU) block.
+
+        Args:
+            kernel_size (int): kernel size
+            in_ch (int): number of input channels
+            out_ch (int): number of output channels
+            bottlececk_dim (int): dimension of the bottleneck layers for computational efficiency. Defaults to 0.
+            dropout (float): dropout probability. Defaults to 0..
+        """
+
+        super().__init__()
+
+        self.conv_residual = None
+        if in_ch != out_ch:
+            self.conv_residual = nn.utils.weight_norm(
+                nn.Conv2D(
+                    in_channels=in_ch, out_channels=out_ch, kernel_size=(1, 1)),
+                name='weight',
+                dim=0)
+            self.dropout_residual = nn.Dropout(p=dropout)
+
+        self.pad_left = ConstantPad2d((0, 0, kernel_size - 1, 0), 0)
+
+        layers = OrderedDict()
+        if bottlececk_dim == 0:
+            layers['conv'] = nn.utils.weight_norm(
+                nn.Conv2D(
+                    in_channels=in_ch,
+                    out_channels=out_ch * 2,
+                    kernel_size=(kernel_size, 1)),
+                name='weight',
+                dim=0)
+            # TODO(hirofumi0810): padding?
+            layers['dropout'] = nn.Dropout(p=dropout)
+            layers['glu'] = GLU()
 
+        elif bottlececk_dim > 0:
+            layers['conv_in'] = nn.utils.weight_norm(
+                nn.Conv2D(
+                    in_channels=in_ch,
+                    out_channels=bottlececk_dim,
+                    kernel_size=(1, 1)),
+                name='weight',
+                dim=0)
+            layers['dropout_in'] = nn.Dropout(p=dropout)
+            layers['conv_bottleneck'] = nn.utils.weight_norm(
+                nn.Conv2D(
+                    in_channels=bottlececk_dim,
+                    out_channels=bottlececk_dim,
+                    kernel_size=(kernel_size, 1)),
+                name='weight',
+                dim=0)
+            layers['dropout'] = nn.Dropout(p=dropout)
+            layers['glu'] = GLU()
+            layers['conv_out'] = nn.utils.weight_norm(
+                nn.Conv2D(
+                    in_channels=bottlececk_dim,
+                    out_channels=out_ch * 2,
+                    kernel_size=(1, 1)),
+                name='weight',
+                dim=0)
+            layers['dropout_out'] = nn.Dropout(p=dropout)
 
-class Swish(nn.Layer):
-    """Construct an Swish object."""
+        self.layers = nn.Sequential(layers)
 
-    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
-        """Return Swish activation function."""
-        return x * F.sigmoid(x)
+    def forward(self, xs):
+        """Forward pass.
+        Args:
+            xs (FloatTensor): `[B, in_ch, T, feat_dim]`
+        Returns:
+            out (FloatTensor): `[B, out_ch, T, feat_dim]`
+        """
+        residual = xs
+        if self.conv_residual is not None:
+            residual = self.dropout_residual(self.conv_residual(residual))
+        xs = self.pad_left(xs)  # `[B, embed_dim, T+kernel-1, 1]`
+        xs = self.layers(xs)  # `[B, out_ch * 2, T ,1]`
+        xs = xs + residual
+        return xs

deepspeech/modules/conformer_convolution.py

@@ -0,0 +1,149 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""ConvolutionModule definition."""
+from typing import Optional, Tuple
+from typeguard import check_argument_types
+
+import logging
+
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+from paddle.nn import initializer as I
+
+# init F.glu func
+# TODO(Hui Zhang): remove this line
+import deepspeech.modules.activation
+
+logger = logging.getLogger(__name__)
+
+__all__ = ['ConvolutionModule']
+
+
+class ConvolutionModule(nn.Layer):
+    """ConvolutionModule in Conformer model."""
+
+    def __init__(self,
+                 channels: int,
+                 kernel_size: int=15,
+                 activation: nn.Layer=nn.ReLU(),
+                 norm: str="batch_norm",
+                 causal: bool=False,
+                 bias: bool=True):
+        """Construct an ConvolutionModule object.
+        Args:
+            channels (int): The number of channels of conv layers.
+            kernel_size (int): Kernel size of conv layers.
+            activation (nn.Layer): Activation Layer.
+            norm (str): Normalization type, 'batch_norm' or 'layer_norm'
+            causal (bool): Whether use causal convolution or not
+            bias (bool): Whether Conv with bias or not
+        """
+        assert check_argument_types()
+        super().__init__()
+        self.pointwise_conv1 = nn.Conv1D(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=None if bias else False,  # None for True as default
+        )
+
+        # self.lorder is used to distinguish if it's a causal convolution,
+        # if self.lorder > 0: 
+        #    it's a causal convolution, the input will be padded with 
+        #    `self.lorder` frames on the left in forward (causal conv impl).
+        # else: it's a symmetrical convolution
+        if causal:
+            padding = 0
+            self.lorder = kernel_size - 1
+        else:
+            # kernel_size should be an odd number for none causal convolution
+            assert (kernel_size - 1) % 2 == 0
+            padding = (kernel_size - 1) // 2
+            self.lorder = 0
+
+        self.depthwise_conv = nn.Conv1D(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=None if bias else False,  # None for True as default
+        )
+
+        assert norm in ['batch_norm', 'layer_norm']
+        if norm == "batch_norm":
+            self.use_layer_norm = False
+            self.norm = nn.BatchNorm1D(channels)
+        else:
+            self.use_layer_norm = True
+            self.norm = nn.LayerNorm(channels)
+
+        self.pointwise_conv2 = nn.Conv1D(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=None if bias else False,  # None for True as default
+        )
+        self.activation = activation
+
+    def forward(self, x: paddle.Tensor, cache: Optional[paddle.Tensor]=None
+                ) -> Tuple[paddle.Tensor, paddle.Tensor]:
+        """Compute convolution module.
+        Args:
+            x (paddle.Tensor): Input tensor (#batch, time, channels).
+            cache (paddle.Tensor): left context cache, it is only
+                used in causal convolution. (#batch, channels, time)
+        Returns:
+            paddle.Tensor: Output tensor (#batch, time, channels).
+            paddle.Tensor: Output cache tensor (#batch, channels, time)
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.transpose([0, 2, 1])  # [B, C, T]
+        if self.lorder > 0:
+            if cache is None:
+                x = nn.functional.pad(
+                    x, (self.lorder, 0), 'constant', 0.0, data_format='NCL')
+            else:
+                assert cache.shape[0] == x.shape[0]  # B
+                assert cache.shape[1] == x.shape[1]  # C
+                x = paddle.concat((cache, x), axis=2)
+
+            assert (x.shape[2] > self.lorder)
+            new_cache = x[:, :, -self.lorder:]  #[B, C, T]
+        else:
+            # It's better we just return None if no cache is requried,
+            # However, for JIT export, here we just fake one tensor instead of
+            # None.
+            new_cache = paddle.to_tensor([0.0], dtype=x.dtype, place=x.place)
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
+        x = nn.functional.glu(x, dim=1)  # (batch, channel, dim)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+        if self.use_layer_norm:
+            x = x.transpose([0, 2, 1])  # [B, T, C]
+        x = self.activation(self.norm(x))
+        if self.use_layer_norm:
+            x = x.transpose([0, 2, 1])  # [B, C, T]
+        x = self.pointwise_conv2(x)
+        x = x.transpose([0, 2, 1])  # [B, T, C]
+        return x, new_cache

deepspeech/modules/conv.py

@@ -24,7 +24,32 @@ from deepspeech.modules.activation import brelu
 
 logger = logging.getLogger(__name__)
 
-__all__ = ['ConvStack']
+__all__ = ['ConvStack', "conv_output_size"]
+
+
+def conv_output_size(I, F, P, S):
+    # https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-convolutional-neural-networks#hyperparameters
+    # Output size after Conv:
+    #   By noting I the length of the input volume size, 
+    #   F the length of the filter, 
+    #   P the amount of zero padding, 
+    #   S the stride,
+    #   then the output size O of the feature map along that dimension is given by:
+    #       O = (I - F + Pstart + Pend) // S + 1
+    #   When Pstart == Pend == P, we can replace Pstart + Pend by 2P.
+    #   When Pstart == Pend == 0
+    #       O = (I - F - S) // S
+    # https://iq.opengenus.org/output-size-of-convolution/
+    # Output height = (Input height + padding height top + padding height bottom - kernel height) / (stride height) + 1
+    # Output width = (Output width + padding width right + padding width left - kernel width) / (stride width) + 1
+    return (I - F + 2 * P - S) // S
+
+
+# receptive field calculator
+# https://fomoro.com/research/article/receptive-field-calculator
+# https://stanford.edu/~shervine/teaching/cs-230/cheatsheet-convolutional-neural-networks#hyperparameters
+# https://distill.pub/2019/computing-receptive-fields/
+# Rl-1 = Sl * Rl + (Kl - Sl) 
 
 
 class ConvBn(nn.Layer):

deepspeech/modules/encoder_layer.py

@@ -0,0 +1,277 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Encoder self-attention layer definition."""
+from typing import Optional, Tuple
+import logging
+
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+from paddle.nn import initializer as I
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["TransformerEncoderLayer", "ConformerEncoderLayer"]
+
+
+class TransformerEncoderLayer(nn.Layer):
+    """Encoder layer module."""
+
+    def __init__(
+            self,
+            size: int,
+            self_attn: nn.Layer,
+            feed_forward: nn.Layer,
+            dropout_rate: float,
+            normalize_before: bool=True,
+            concat_after: bool=False, ):
+        """Construct an EncoderLayer object.
+        
+        Args:
+            size (int): Input dimension.
+            self_attn (nn.Layer): Self-attention module instance.
+                `MultiHeadedAttention` or `RelPositionMultiHeadedAttention`
+                instance can be used as the argument.
+            feed_forward (nn.Layer): Feed-forward module instance.
+                `PositionwiseFeedForward`, instance can be used as the argument.
+            dropout_rate (float): Dropout rate.
+            normalize_before (bool):
+                True: use layer_norm before each sub-block.
+                False: to use layer_norm after each sub-block.
+            concat_after (bool): Whether to concat attention layer's input and
+                output.
+                True: x -> x + linear(concat(x, att(x)))
+                False: x -> x + att(x)
+        """
+        super().__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.norm1 = nn.LayerNorm(size, epsilon=1e-12)
+        self.norm2 = nn.LayerNorm(size, epsilon=1e-12)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+        self.concat_after = concat_after
+        # concat_linear may be not used in forward fuction,
+        # but will be saved in the *.pt
+        self.concat_linear = nn.Linear(size + size, size)
+
+    def forward(
+            self,
+            x: paddle.Tensor,
+            mask: paddle.Tensor,
+            pos_emb: paddle.Tensor,
+            output_cache: Optional[paddle.Tensor]=None,
+            cnn_cache: Optional[paddle.Tensor]=None,
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Compute encoded features.
+        Args:
+            x (paddle.Tensor): Input tensor (#batch, time, size).
+            mask (paddle.Tensor): Mask tensor for the input (#batch, time).
+            pos_emb (paddle.Tensor): just for interface compatibility
+                to ConformerEncoderLayer
+            output_cache (paddle.Tensor): Cache tensor of the output
+                (#batch, time2, size), time2 < time in x.
+            cnn_cache (paddle.Tensor): not used here, it's for interface
+                compatibility to ConformerEncoderLayer
+        Returns:
+            paddle.Tensor: Output tensor (#batch, time, size).
+            paddle.Tensor: Mask tensor (#batch, time).
+        """
+        residual = x
+        if self.normalize_before:
+            x = self.norm1(x)
+
+        if output_cache is None:
+            x_q = x
+        else:
+            assert output_cache.shape[0] == x.shape[0]
+            assert output_cache.shape[1] < x.shape[1]
+            assert output_cache.shape[2] == self.size
+            chunk = x.shape[1] - output_cache.shape[1]
+            x_q = x[:, -chunk:, :]
+            residual = residual[:, -chunk:, :]
+            mask = mask[:, -chunk:, :]
+
+        if self.concat_after:
+            x_concat = paddle.concat(
+                (x, self.self_attn(x_q, x, x, mask)), axis=-1)
+            x = residual + self.concat_linear(x_concat)
+        else:
+            x = residual + self.dropout(self.self_attn(x_q, x, x, mask))
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        x = residual + self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        if output_cache is not None:
+            x = paddle.concat([output_cache, x], axis=1)
+
+        fake_cnn_cache = paddle.to_tensor([0.0], dtype=x.dtype, place=x.place)
+        return x, mask, fake_cnn_cache
+
+
+class ConformerEncoderLayer(nn.Layer):
+    """Encoder layer module."""
+
+    def __init__(
+            self,
+            size: int,
+            self_attn: int,
+            feed_forward: Optional[nn.Layer]=None,
+            feed_forward_macaron: Optional[nn.Layer]=None,
+            conv_module: Optional[nn.Layer]=None,
+            dropout_rate: float=0.1,
+            normalize_before: bool=True,
+            concat_after: bool=False, ):
+        """Construct an EncoderLayer object.
+        
+        Args:
+            size (int): Input dimension.
+            self_attn (nn.Layer): Self-attention module instance.
+                `MultiHeadedAttention` or `RelPositionMultiHeadedAttention`
+                instance can be used as the argument.
+            feed_forward (nn.Layer): Feed-forward module instance.
+                `PositionwiseFeedForward` instance can be used as the argument.
+            feed_forward_macaron (nn.Layer): Additional feed-forward module
+                instance.
+                `PositionwiseFeedForward` instance can be used as the argument.
+            conv_module (nn.Layer): Convolution module instance.
+                `ConvlutionModule` instance can be used as the argument.
+            dropout_rate (float): Dropout rate.
+            normalize_before (bool):
+                True: use layer_norm before each sub-block.
+                False: use layer_norm after each sub-block.
+            concat_after (bool): Whether to concat attention layer's input and
+                output.
+                True: x -> x + linear(concat(x, att(x)))
+                False: x -> x + att(x)
+        """
+        super().__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.feed_forward_macaron = feed_forward_macaron
+        self.conv_module = conv_module
+        self.norm_ff = nn.LayerNorm(size, epsilon=1e-12)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(size, epsilon=1e-12)  # for the MHA module
+        if feed_forward_macaron is not None:
+            self.norm_ff_macaron = nn.LayerNorm(size, epsilon=1e-12)
+            self.ff_scale = 0.5
+        else:
+            self.ff_scale = 1.0
+        if self.conv_module is not None:
+            self.norm_conv = nn.LayerNorm(
+                size, epsilon=1e-12)  # for the CNN module
+            self.norm_final = nn.LayerNorm(
+                size, epsilon=1e-12)  # for the final output of the block
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+        self.concat_after = concat_after
+        self.concat_linear = nn.Linear(size + size, size)
+
+    def forward(
+            self,
+            x: paddle.Tensor,
+            mask: paddle.Tensor,
+            pos_emb: paddle.Tensor,
+            output_cache: Optional[paddle.Tensor]=None,
+            cnn_cache: Optional[paddle.Tensor]=None,
+    ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Compute encoded features.
+        Args:
+            x (paddle.Tensor): (#batch, time, size)
+            mask (paddle.Tensor): Mask tensor for the input (#batch, time，time).
+            pos_emb (paddle.Tensor): positional encoding, must not be None
+                for ConformerEncoderLayer.
+            output_cache (paddle.Tensor): Cache tensor of the output
+                (#batch, time2, size), time2 < time in x.
+            cnn_cache (paddle.Tensor): Convolution cache in conformer layer
+        Returns:
+            paddle.Tensor: Output tensor (#batch, time, size).
+            paddle.Tensor: Mask tensor (#batch, time).
+        """
+        # whether to use macaron style FFN
+        if self.feed_forward_macaron is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_ff_macaron(x)
+            x = residual + self.ff_scale * self.dropout(
+                self.feed_forward_macaron(x))
+            if not self.normalize_before:
+                x = self.norm_ff_macaron(x)
+
+        # multi-headed self-attention module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_mha(x)
+
+        if output_cache is None:
+            x_q = x
+        else:
+            assert output_cache.shape[0] == x.shape[0]
+            assert output_cache.shape[1] < x.shape[1]
+            assert output_cache.shape[2] == self.size
+            chunk = x.shape[1] - output_cache.shape[1]
+            x_q = x[:, -chunk:, :]
+            residual = residual[:, -chunk:, :]
+            mask = mask[:, -chunk:, :]
+
+        x_att = self.self_attn(x_q, x, x, pos_emb, mask)
+
+        if self.concat_after:
+            x_concat = paddle.concat((x, x_att), axis=-1)
+            x = residual + self.concat_linear(x_concat)
+        else:
+            x = residual + self.dropout(x_att)
+
+        if not self.normalize_before:
+            x = self.norm_mha(x)
+
+        # convolution module
+        # Fake new cnn cache here, and then change it in conv_module
+        new_cnn_cache = paddle.to_tensor([0.0], dtype=x.dtype, place=x.place)
+        if self.conv_module is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_conv(x)
+
+            x, new_cnn_cache = self.conv_module(x, cnn_cache)
+            x = residual + self.dropout(x)
+
+            if not self.normalize_before:
+                x = self.norm_conv(x)
+
+        # feed forward module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_ff(x)
+
+        x = residual + self.ff_scale * self.dropout(self.feed_forward(x))
+
+        if not self.normalize_before:
+            x = self.norm_ff(x)
+
+        if self.conv_module is not None:
+            x = self.norm_final(x)
+
+        if output_cache is not None:
+            x = paddle.concat([output_cache, x], axis=1)
+
+        return x, mask, new_cnn_cache

deepspeech/modules/positionwise_feed_forward.py

@@ -0,0 +1,59 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Positionwise feed forward layer definition."""
+import logging
+
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+from paddle.nn import initializer as I
+
+logger = logging.getLogger(__name__)
+
+__all__ = ["PositionwiseFeedForward"]
+
+
+class PositionwiseFeedForward(nn.Layer):
+    """Positionwise feed forward layer."""
+
+    def __init__(self,
+                 idim: int,
+                 hidden_units: int,
+                 dropout_rate: float,
+                 activation: nn.Layer=nn.ReLU()):
+        """Construct a PositionwiseFeedForward object.
+
+        FeedForward are appied on each position of the sequence.
+        The output dim is same with the input dim.
+
+        Args:
+            idim (int): Input dimenstion.
+            hidden_units (int): The number of hidden units.
+            dropout_rate (float): Dropout rate.
+            activation (paddle.nn.Layer): Activation function
+        """
+        super().__init__()
+        self.w_1 = nn.Linear(idim, hidden_units)
+        self.activation = activation
+        self.dropout = nn.Dropout(dropout_rate)
+        self.w_2 = nn.Linear(hidden_units, idim)
+
+    def forward(self, xs: paddle.Tensor) -> paddle.Tensor:
+        """Forward function.
+        Args:
+            xs: input tensor (B, Lmax, D)
+        Returns:
+            output tensor, (B, Lmax, D)
+        """
+        return self.w_2(self.dropout(self.activation(self.w_1(xs))))

deepspeech/modules/subsampling.py

@@ -0,0 +1,235 @@
+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Subsampling layer definition."""
+
+from typing import Tuple
+import logging
+
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+from paddle.nn import initializer as I
+
+from deepspeech.modules.embedding import PositionalEncoding
+
+logger = logging.getLogger(__name__)
+
+__all__ = [
+    "LinearNoSubsampling", "Conv2dSubsampling4", "Conv2dSubsampling6",
+    "Conv2dSubsampling8"
+]
+
+
+class BaseSubsampling(nn.Layer):
+    def __init__(self, pos_enc_class: PositionalEncoding):
+        super().__init__()
+        self.pos_enc = pos_enc_class
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def position_encoding(self, offset: int, size: int) -> paddle.Tensor:
+        return self.pos_enc.position_encoding(offset, size)
+
+
+class LinearNoSubsampling(BaseSubsampling):
+    """Linear transform the input without subsampling."""
+
+    def __init__(self,
+                 idim: int,
+                 odim: int,
+                 dropout_rate: float,
+                 pos_enc_class: PositionalEncoding):
+        """Construct an linear object.
+        Args:
+            idim (int): Input dimension.
+            odim (int): Output dimension.
+            dropout_rate (float): Dropout rate.
+            pos_enc_class (PositionalEncoding): position encoding class
+        """
+        super().__init__(pos_enc_class)
+        self.out = nn.Sequential(
+            nn.Linear(idim, odim),
+            nn.LayerNorm(odim, epsilon=1e-12),
+            nn.Dropout(dropout_rate), )
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def forward(self, x: paddle.Tensor, x_mask: paddle.Tensor, offset: int=0
+                ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Input x.
+        Args:
+            x (paddle.Tensor): Input tensor (#batch, time, idim).
+            x_mask (paddle.Tensor): Input mask (#batch, 1, time).
+        Returns:
+            paddle.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            paddle.Tensor: positional encoding
+            paddle.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+        """
+        x = self.out(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask
+
+
+class Conv2dSubsampling4(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/4 length)."""
+
+    def __init__(self,
+                 idim: int,
+                 odim: int,
+                 dropout_rate: float,
+                 pos_enc_class: PositionalEncoding):
+        """Construct an Conv2dSubsampling4 object.
+        
+        Args:
+            idim (int): Input dimension.
+            odim (int): Output dimension.
+            dropout_rate (float): Dropout rate.
+        """
+        super().__init__(pos_enc_class)
+        self.conv = nn.Sequential(
+            nn.Conv2D(1, odim, 3, 2),
+            nn.ReLU(),
+            nn.Conv2D(odim, odim, 3, 2),
+            nn.ReLU(), )
+        self.linear = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+        self.subsampling_rate = 4
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) / 2 * stride  * frame_rate_of_this_layer
+        # 6 = (3 - 1) / 2 * 2 * 1 + (3 - 1) / 2 * 2 * 2
+        self.right_context = 6
+
+    def forward(self, x: paddle.Tensor, x_mask: paddle.Tensor, offset: int=0
+                ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Subsample x.
+        Args:
+            x (paddle.Tensor): Input tensor (#batch, time, idim).
+            x_mask (paddle.Tensor): Input mask (#batch, 1, time).
+        Returns:
+            paddle.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 4.
+            paddle.Tensor: positional encoding
+            paddle.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 4.
+        """
+        x = x.unsqueeze(1)  # (b, c=1, t, f)
+        x = self.conv(x)
+        b, c, t, f = paddle.shape(x)
+        x = self.linear(x.transpose([0, 1, 2, 3]).reshape([b, t, c * f]))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, :-2:2][:, :, :-2:2]
+
+
+class Conv2dSubsampling6(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/6 length)."""
+
+    def __init__(self,
+                 idim: int,
+                 odim: int,
+                 dropout_rate: float,
+                 pos_enc_class: PositionalEncoding):
+        """Construct an Conv2dSubsampling6 object.
+        
+        Args:
+            idim (int): Input dimension.
+            odim (int): Output dimension.
+            dropout_rate (float): Dropout rate.
+            pos_enc (PositionalEncoding): Custom position encoding layer.
+        """
+        super().__init__(pos_enc_class)
+        self.conv = nn.Sequential(
+            nn.Conv2D(1, odim, 3, 2),
+            nn.ReLU(),
+            nn.Conv2D(odim, odim, 5, 3),
+            nn.ReLU(), )
+        # O = (I - F + Pstart + Pend) // S + 1
+        # when Padding == 0, O = (I - F - S) // S
+        self.linear = nn.Linear(odim * (((idim - 1) // 2 - 2) // 3), odim)
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) / 2 * stride  * frame_rate_of_this_layer
+        # 14 = (3 - 1) / 2 * 2 * 1 + (5 - 1) / 2 * 3 * 2
+        self.subsampling_rate = 6
+        self.right_context = 14
+
+    def forward(self, x: paddle.Tensor, x_mask: paddle.Tensor, offset: int=0
+                ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Subsample x.
+        Args:
+            x (paddle.Tensor): Input tensor (#batch, time, idim).
+            x_mask (paddle.Tensor): Input mask (#batch, 1, time).
+        Returns:
+            paddle.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 6.
+            paddle.Tensor: positional encoding
+            paddle.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 6.
+        """
+        x = x.unsqueeze(1)  # (b, c, t, f)
+        x = self.conv(x)
+        b, c, t, f = paddle.shape(x)
+        x = self.linear(x.transpose([0, 1, 2, 3]).reshape([b, t, c * f]))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, :-2:2][:, :, :-4:3]
+
+
+class Conv2dSubsampling8(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/8 length)."""
+
+    def __init__(self,
+                 idim: int,
+                 odim: int,
+                 dropout_rate: float,
+                 pos_enc_class: PositionalEncoding):
+        """Construct an Conv2dSubsampling8 object.
+        
+        Args:
+            idim (int): Input dimension.
+            odim (int): Output dimension.
+            dropout_rate (float): Dropout rate.
+        """
+        super().__init__(pos_enc_class)
+        self.conv = nn.Sequential(
+            nn.Conv2D(1, odim, 3, 2),
+            nn.ReLU(),
+            nn.Conv2D(odim, odim, 3, 2),
+            nn.ReLU(),
+            nn.Conv2D(odim, odim, 3, 2),
+            nn.ReLU(), )
+        self.linear = nn.Linear(odim * ((((idim - 1) // 2 - 1) // 2 - 1) // 2),
+                                odim)
+        self.subsampling_rate = 8
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) / 2 * stride  * frame_rate_of_this_layer
+        # 14 = (3 - 1) / 2 * 2 * 1 + (3 - 1) / 2 * 2 * 2 + (3 - 1) / 2 * 2 * 4
+        self.right_context = 14
+
+    def forward(self, x: paddle.Tensor, x_mask: paddle.Tensor, offset: int=0
+                ) -> Tuple[paddle.Tensor, paddle.Tensor, paddle.Tensor]:
+        """Subsample x.
+        Args:
+            x (paddle.Tensor): Input tensor (#batch, time, idim).
+            x_mask (paddle.Tensor): Input mask (#batch, 1, time).
+        Returns:
+            paddle.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 8.
+            paddle.Tensor: positional encoding
+            paddle.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 8.
+        """
+        x = x.unsqueeze(1)  # (b, c, t, f)
+        x = self.conv(x)
+        x = self.linear(x.transpose([0, 1, 2, 3]).reshape([b, t, c * f]))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, :-2:2][:, :, :-2:2][:, :, :-2:2]

deepspeech/training/gradclip.py

@@ -21,8 +21,10 @@ from paddle.fluid import core
 
 logger = logging.getLogger(__name__)
 
+__all__ = ["ClipGradByGlobalNormWithLog"]
 
-class MyClipGradByGlobalNorm(paddle.nn.ClipGradByGlobalNorm):
+
+class ClipGradByGlobalNormWithLog(paddle.nn.ClipGradByGlobalNorm):
     def __init__(self, clip_norm):
         super().__init__(clip_norm)