Add remaining files in pruned_transducer_stateless2

egs/librispeech/ASR/pruned_transducer_stateless2/conformer.py

@@ -19,7 +19,7 @@ import copy
 import math
 import warnings
 from typing import Optional, Tuple, Sequence
-from subsampling import DoubleSwish, ActivationBalancer, BasicNorm, ScaledLinear, ScaledConv1d, ScaledConv2d
+from scaling import DoubleSwish, ActivationBalancer, BasicNorm, ScaledLinear, ScaledConv1d, ScaledConv2d
 
 import torch
 from torch import Tensor, nn

egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py

@@ -0,0 +1,241 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from typing import Optional
+from scaling import ScaledConv1d, ScaledLinear
+
+
+class Decoder(nn.Module):
+    """This class modifies the stateless decoder from the following paper:
+
+        RNN-transducer with stateless prediction network
+        https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
+
+    It removes the recurrent connection from the decoder, i.e., the prediction
+    network. Different from the above paper, it adds an extra Conv1d
+    right after the embedding layer.
+
+    TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        blank_id: int,
+        context_size: int,
+    ):
+        """
+        Args:
+          vocab_size:
+            Number of tokens of the modeling unit including blank.
+          embedding_dim:
+            Dimension of the input embedding.
+          blank_id:
+            The ID of the blank symbol.
+          context_size:
+            Number of previous words to use to predict the next word.
+            1 means bigram; 2 means trigram. n means (n+1)-gram.
+        """
+        super().__init__()
+        self.embedding = ScaledEmbedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+            padding_idx=blank_id,
+        )
+        self.blank_id = blank_id
+
+        assert context_size >= 1, context_size
+        self.context_size = context_size
+        if context_size > 1:
+            self.conv = ScaledConv1d(
+                in_channels=embedding_dim,
+                out_channels=embedding_dim,
+                kernel_size=context_size,
+                padding=0,
+                groups=embedding_dim,
+                bias=False,
+            )
+        self.output_linear = ScaledLinear(embedding_dim, vocab_size)
+
+
+    def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, U).
+          need_pad:
+            True to left pad the input. Should be True during training.
+            False to not pad the input. Should be False during inference.
+        Returns:
+          Return a tensor of shape (N, U, embedding_dim).
+        """
+        y = y.to(torch.int64)
+        embedding_out = self.embedding(y)
+        if self.context_size > 1:
+            embedding_out = embedding_out.permute(0, 2, 1)
+            if need_pad is True:
+                embedding_out = F.pad(
+                    embedding_out, pad=(self.context_size - 1, 0)
+                )
+            else:
+                # During inference time, there is no need to do extra padding
+                # as we only need one output
+                assert embedding_out.size(-1) == self.context_size
+            embedding_out = self.conv(embedding_out)
+            embedding_out = embedding_out.permute(0, 2, 1)
+        embedding_out = self.output_linear(F.relu(embedding_out))
+        return embedding_out
+
+
+
+class ScaledEmbedding(nn.Module):
+    r"""A simple lookup table that stores embeddings of a fixed dictionary and size.
+
+    This module is often used to store word embeddings and retrieve them using indices.
+    The input to the module is a list of indices, and the output is the corresponding
+    word embeddings.
+
+    Args:
+        num_embeddings (int): size of the dictionary of embeddings
+        embedding_dim (int): the size of each embedding vector
+        padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx`
+                                         (initialized to zeros) whenever it encounters the index.
+        max_norm (float, optional): If given, each embedding vector with norm larger than :attr:`max_norm`
+                                    is renormalized to have norm :attr:`max_norm`.
+        norm_type (float, optional): The p of the p-norm to compute for the :attr:`max_norm` option. Default ``2``.
+        scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of
+                                                the words in the mini-batch. Default ``False``.
+        sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor.
+                                 See Notes for more details regarding sparse gradients.
+
+    Attributes:
+        weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim)
+                         initialized from :math:`\mathcal{N}(0, 1)`
+
+    Shape:
+        - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract
+        - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}`
+
+    .. note::
+        Keep in mind that only a limited number of optimizers support
+        sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`),
+        :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`)
+
+    .. note::
+        With :attr:`padding_idx` set, the embedding vector at
+        :attr:`padding_idx` is initialized to all zeros. However, note that this
+        vector can be modified afterwards, e.g., using a customized
+        initialization method, and thus changing the vector used to pad the
+        output. The gradient for this vector from :class:`~torch.nn.Embedding`
+        is always zero.
+
+    Examples::
+
+        >>> # an Embedding module containing 10 tensors of size 3
+        >>> embedding = nn.Embedding(10, 3)
+        >>> # a batch of 2 samples of 4 indices each
+        >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]])
+        >>> embedding(input)
+        tensor([[[-0.0251, -1.6902,  0.7172],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 1.4970,  1.3448, -0.9685],
+                 [-0.3677, -2.7265, -0.1685]],
+
+                [[ 1.4970,  1.3448, -0.9685],
+                 [ 0.4362, -0.4004,  0.9400],
+                 [-0.6431,  0.0748,  0.6969],
+                 [ 0.9124, -2.3616,  1.1151]]])
+
+
+        >>> # example with padding_idx
+        >>> embedding = nn.Embedding(10, 3, padding_idx=0)
+        >>> input = torch.LongTensor([[0,2,0,5]])
+        >>> embedding(input)
+        tensor([[[ 0.0000,  0.0000,  0.0000],
+                 [ 0.1535, -2.0309,  0.9315],
+                 [ 0.0000,  0.0000,  0.0000],
+                 [-0.1655,  0.9897,  0.0635]]])
+    """
+    __constants__ = ['num_embeddings', 'embedding_dim', 'padding_idx',
+                     'scale_grad_by_freq', 'sparse']
+
+    num_embeddings: int
+    embedding_dim: int
+    padding_idx: int
+    scale_grad_by_freq: bool
+    weight: Tensor
+    sparse: bool
+
+    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None,
+                 scale_grad_by_freq: bool = False,
+                 sparse: bool = False,
+                 scale_speed: float = 5.0) -> None:
+        super(ScaledEmbedding, self).__init__()
+        self.num_embeddings = num_embeddings
+        self.embedding_dim = embedding_dim
+        if padding_idx is not None:
+            if padding_idx > 0:
+                assert padding_idx < self.num_embeddings, 'Padding_idx must be within num_embeddings'
+            elif padding_idx < 0:
+                assert padding_idx >= -self.num_embeddings, 'Padding_idx must be within num_embeddings'
+                padding_idx = self.num_embeddings + padding_idx
+        self.padding_idx = padding_idx
+        self.scale_grad_by_freq = scale_grad_by_freq
+
+        self.scale_speed = scale_speed
+        self.scale = nn.Parameter(torch.zeros(())) # see reset_parameters()
+        self.sparse = sparse
+
+        self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim))
+        self.reset_parameters()
+
+
+
+    def reset_parameters(self) -> None:
+        nn.init.normal_(self.weight, std=0.05)
+        nn.init.constant_(self.scale, torch.tensor(1.0/0.05).log() / self.scale_speed)
+
+        if self.padding_idx is not None:
+            with torch.no_grad():
+                self.weight[self.padding_idx].fill_(0)
+
+    def forward(self, input: Tensor) -> Tensor:
+        scale = (self.scale * self.scale_speed).exp()
+        if input.numel() < self.num_embeddings:
+            return F.embedding(
+                input, self.weight, self.padding_idx,
+                None, 2.0, # None, 2.0 relate to normalization
+                 self.scale_grad_by_freq, self.sparse) * scale
+        else:
+            return F.embedding(
+                input, self.weight * scale, self.padding_idx,
+                None, 2.0, # None, 2.0 relates to normalization
+                self.scale_grad_by_freq, self.sparse)
+
+    def extra_repr(self) -> str:
+        s = '{num_embeddings}, {embedding_dim}, scale_speed={scale_speed}, scale={scale}'
+        if self.padding_idx is not None:
+            s += ', padding_idx={padding_idx}'
+        if self.scale_grad_by_freq is not False:
+            s += ', scale_grad_by_freq={scale_grad_by_freq}'
+        if self.sparse is not False:
+            s += ', sparse=True'
+        return s.format(**self.__dict__)

egs/librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py

@@ -0,0 +1 @@
+../transducer_stateless/encoder_interface.py

egs/librispeech/ASR/pruned_transducer_stateless2/export.py

@@ -0,0 +1,182 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless/export.py \
+  --exp-dir ./pruned_transducer_stateless/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless/decode.py \
+        --exp-dir ./pruned_transducer_stateless/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    assert args.jit is False, "Support torchscript will be added later"
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()

egs/librispeech/ASR/pruned_transducer_stateless2/joiner.py

@@ -0,0 +1,50 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from scaling import ScaledLinear
+
+class Joiner(nn.Module):
+    def __init__(self, input_dim: int, inner_dim: int, output_dim: int):
+        super().__init__()
+
+        self.inner_linear = ScaledLinear(input_dim, inner_dim)
+        self.output_linear = ScaledLinear(inner_dim, output_dim)
+
+    def forward(
+        self, encoder_out: torch.Tensor, decoder_out: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            Output from the encoder. Its shape is (N, T, s_range, C).
+          decoder_out:
+            Output from the decoder. Its shape is (N, T, s_range, C).
+        Returns:
+          Return a tensor of shape (N, T, s_range, C).
+        """
+        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.shape == decoder_out.shape
+
+        logit = encoder_out + decoder_out
+
+        logit = self.inner_linear(torch.tanh(logit))
+
+        output = self.output_linear(F.relu(logit))
+
+        return output

egs/librispeech/ASR/pruned_transducer_stateless2/model.py

@@ -0,0 +1,170 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup_mode: bool = False
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, x_lens = self.encoder(x, x_lens, warmup_mode=warmup_mode)
+        assert torch.all(x_lens > 0)
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = self.decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, C]
+        decoder_out = self.decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (x.size(0), 4), dtype=torch.int64, device=x.device
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = x_lens
+
+        simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+            lm=decoder_out,
+            am=encoder_out,
+            symbols=y_padded,
+            termination_symbol=blank_id,
+            lm_only_scale=lm_scale,
+            am_only_scale=am_scale,
+            boundary=boundary,
+            reduction="sum",
+            return_grad=True,
+        )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, C]
+        # lm_pruned : [B, T, prune_range, C]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=encoder_out, lm=decoder_out, ranges=ranges
+        )
+
+        # logits : [B, T, prune_range, C]
+        logits = self.joiner(am_pruned, lm_pruned)
+
+        pruned_loss = k2.rnnt_loss_pruned(
+            logits=logits,
+            symbols=y_padded,
+            ranges=ranges,
+            termination_symbol=blank_id,
+            boundary=boundary,
+            reduction="sum",
+        )
+
+        return (simple_loss, pruned_loss)

egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py

@@ -0,0 +1,418 @@
+# Copyright    2022  Xiaomi Corp.        (authors: Daniel Povey)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+from typing import Tuple
+
+
+
+
+
+class ActivationBalancerFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor,
+                channel_dim: int,
+                min_positive: float, # e.g. 0.05
+                max_positive: float, # e.g. 0.95
+                max_factor: float, # e.g. 0.01
+                min_abs: float, # e.g. 0.2
+                max_abs: float, # e.g. 100.0
+    ) -> Tensor:
+        if x.requires_grad:
+            if channel_dim < 0:
+                channel_dim += x.ndim
+            sum_dims = [d for d in range(x.ndim) if d != channel_dim]
+            xgt0 = x > 0
+            proportion_positive = torch.mean(xgt0.to(x.dtype), dim=sum_dims, keepdim=True)
+            factor1 = ((min_positive - proportion_positive).relu() * (max_factor / min_positive)
+                       if min_positive != 0.0 else 0.0)
+            factor2 = ((proportion_positive - max_positive).relu() * (max_factor / (max_positive - 1.0))
+                       if max_positive != 1.0 else 0.0)
+            factor = factor1 + factor2
+            if isinstance(factor, float):
+                factor = torch.zeros_like(proportion_positive)
+
+            mean_abs = torch.mean(x.abs(), dim=sum_dims, keepdim=True)
+            below_threshold = (mean_abs < min_abs)
+            above_threshold = (mean_abs > max_abs)
+
+            ctx.save_for_backward(factor, xgt0, below_threshold, above_threshold)
+            ctx.max_factor = max_factor
+            ctx.sum_dims = sum_dims
+        return x
+
+    @staticmethod
+    def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None, None, None, None]:
+        factor, xgt0, below_threshold, above_threshold = ctx.saved_tensors
+        dtype = x_grad.dtype
+        scale_factor = ((below_threshold.to(dtype) - above_threshold.to(dtype)) *
+                        (xgt0.to(dtype) - 0.5) * (ctx.max_factor * 2.0))
+
+        neg_delta_grad = x_grad.abs() * (factor + scale_factor)
+        return x_grad - neg_delta_grad, None, None, None, None, None, None
+
+
+class BasicNorm(torch.nn.Module):
+    """
+    This is intended to be a simpler, and hopefully cheaper, replacement for
+    LayerNorm.  The observation this is based on, is that Transformer-type
+    networks, especially with pre-norm, sometimes seem to set one of the
+    feature dimensions to a large constant value (e.g. 50), which "defeats"
+    the LayerNorm because the output magnitude is then not strongly dependent
+    on the other (useful) features.  Presumably the weight and bias of the
+    LayerNorm are required to allow it to do this.
+
+    So the idea is to introduce this large constant value as an explicit
+    parameter, that takes the role of the "eps" in LayerNorm, so the network
+    doesn't have to do this trick.  We make the "eps" learnable.
+
+    Args:
+       num_channels: the number of channels, e.g. 512.
+      channel_dim: the axis/dimension corresponding to the channel,
+        interprted as an offset from the input's ndim if negative.
+        shis is NOT the num_channels; it should typically be one of
+        {-2, -1, 0, 1, 2, 3}.
+       eps: the initial "epsilon" that we add as ballast in:
+             scale = ((input_vec**2).mean() + epsilon)**-0.5
+          Note: our epsilon is actually large, but we keep the name
+          to indicate the connection with conventional LayerNorm.
+       learn_eps: if true, we learn epsilon; if false, we keep it
+         at the initial value.
+      eps_speed: a constant that determines how fast "eps" learns;
+          with Adam and variants, this should probably be >= 1,
+          e.g. 5.0.  For SGD and variants, probably a value less than one,
+          like 0.1, would be suitable, to prevent instability.
+    """
+    def __init__(self,
+                 num_channels: int,
+                 channel_dim: int = -1,  # CAUTION: see documentation.
+                 eps: float = 0.25,
+                 learn_eps: bool = True,
+                 eps_speed: float = 5.0):
+        super(BasicNorm, self).__init__()
+        self.num_channels = num_channels
+        self.channel_dim = channel_dim
+        self.eps_speed = eps_speed
+        if learn_eps:
+            self.eps = nn.Parameter((torch.tensor(eps).log() / eps_speed).detach())
+        else:
+            self.register_buffer('eps', (torch.tensor(eps).log() / eps_speed).detach())
+
+
+    def forward(self, x: Tensor) -> Tensor:
+        assert x.shape[self.channel_dim] == self.num_channels
+        scales = (torch.mean(x**2, dim=self.channel_dim, keepdim=True) +
+                  (self.eps * self.eps_speed).exp()) ** -0.5
+        return x * scales
+
+
+
+
+class ScaledLinear(nn.Linear):
+    """
+    A modified version of nn.Linear where the parameters are scaled before
+    use, via:
+         weight = self.weight * (self.weight_scale * self.scale_speed).exp()
+         bias = self.bias * (self.bias_scale * self.scale_speed).exp()
+
+    Args:
+        Accepts the standard args and kwargs that nn.Linear accepts
+        e.g. in_features, out_features, bias=False.
+
+        scale_speed: a factor that affects how fast the weight_scale
+           and bias_scale learn; this value is suitable for Adam-type
+           optimizers.
+        initial_scale: you can override this if you want to increase
+           or decrease the initial magnitude of the module's output
+           (affects the initialization of weight_scale and bias_scale).
+           Another option, if you want to do something like this, is
+           to re-initialize the parameters.
+
+       Note: it uses the default initialization for the weight and bias,
+       inherited from nn.Linear.  For modules with small fan-in, this
+       may be larger than optimal.
+    """
+    def __init__(self, *args,
+                 scale_speed: float = 5.0,
+                 initial_scale: float = 1.0,
+                 **kwargs):
+        super(ScaledLinear, self).__init__(*args, **kwargs)
+        initial_scale = (torch.tensor(initial_scale).log() / scale_speed)
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        self.scale_speed = scale_speed
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter('bias_scale', None)
+
+        self._reset_parameters()  # Overrides the reset_parameters in nn.Linear
+
+    def _reset_parameters(self):
+        std = 0.05
+        a = (3 ** 0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in ** -0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += (torch.tensor(scale / std).log() / self.scale_speed)
+
+    def get_weight(self):
+        return self.weight * (self.weight_scale * self.scale_speed).exp()
+
+    def get_bias(self):
+        return (None if self.bias is None else
+                self.bias * (self.bias_scale * self.scale_speed).exp())
+
+    def forward(self, input: Tensor) -> Tensor:
+        return torch.nn.functional.linear(input, self.get_weight(),
+                                          self.get_bias())
+
+
+class ScaledConv1d(nn.Conv1d):
+    def __init__(self, *args, scale_speed = 5.0,
+                 initial_scale=1.0, **kwargs):
+        super(ScaledConv1d, self).__init__(*args, **kwargs)
+        self.scale_speed = scale_speed
+        initial_scale = (torch.tensor(initial_scale).log() / scale_speed)
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter('bias_scale', None)
+        self._reset_parameters()  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self):
+        std = 0.05
+        a = (3 ** 0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in ** -0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += (torch.tensor(scale / std).log() / self.scale_speed)
+
+
+    def get_weight(self):
+        return self.weight * (self.weight_scale * self.scale_speed).exp()
+
+    def get_bias(self):
+        return (None if self.bias is None else
+                self.bias * (self.bias_scale * self.scale_speed).exp())
+
+    def forward(self, input: Tensor) -> Tensor:
+        F = torch.nn.functional
+        if self.padding_mode != 'zeros':
+            return F.conv1d(F.pad(input, self._reversed_padding_repeated_twice, mode=self.padding_mode),
+                            self.get_weight(), self.get_bias(), self.stride,
+                            _single(0), self.dilation, self.groups)
+        return F.conv1d(input, self.get_weight(), self.get_bias(), self.stride,
+                        self.padding, self.dilation, self.groups)
+
+
+
+class ScaledConv2d(nn.Conv2d):
+    def __init__(self, *args, scale_speed=5.0, initial_scale=1.0, **kwargs):
+        super(ScaledConv2d, self).__init__(*args, **kwargs)
+        self.scale_speed = scale_speed
+        initial_scale = (torch.tensor(initial_scale).log() / scale_speed)
+        self.weight_scale = nn.Parameter(initial_scale.clone().detach())
+        if self.bias is not None:
+            self.bias_scale = nn.Parameter(initial_scale.clone().detach())
+        else:
+            self.register_parameter('bias_scale', None)
+        self._reset_parameters()  # Overrides the reset_parameters in base class
+
+    def _reset_parameters(self):
+        std = 0.05
+        a = (3 ** 0.5) * std
+        nn.init.uniform_(self.weight, -a, a)
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0.0)
+        fan_in = self.weight.shape[1] * self.weight[0][0].numel()
+        scale = fan_in ** -0.5  # 1/sqrt(fan_in)
+        with torch.no_grad():
+            self.weight_scale += (torch.tensor(scale / std).log() / self.scale_speed)
+
+
+    def get_weight(self):
+        return self.weight * (self.weight_scale * self.scale_speed).exp()
+
+    def get_bias(self):
+        return (None if self.bias is None else
+                self.bias * (self.bias_scale * self.scale_speed).exp())
+
+    def _conv_forward(self, input, weight):
+        F = torch.nn.functional
+        if self.padding_mode != 'zeros':
+            return F.conv2d(F.pad(input, self._reversed_padding_repeated_twice, mode=self.padding_mode),
+                            weight, self.get_bias(), self.stride,
+                            _pair(0), self.dilation, self.groups)
+        return F.conv2d(input, weight, self.get_bias(), self.stride,
+                        self.padding, self.dilation, self.groups)
+
+    def forward(self, input: Tensor) -> Tensor:
+        return self._conv_forward(input, self.get_weight())
+
+
+
+
+class ActivationBalancer(torch.nn.Module):
+    """
+    Modifies the backpropped derivatives of a function to try to encourage, for
+    each channel, that it is positive at least a proportion `threshold` of the
+    time.  It does this by multiplying negative derivative values by up to
+    (1+max_factor), and positive derivative values by up to (1-max_factor),
+    interpolated from 1 at the threshold to those extremal values when none
+    of the inputs are positive.
+
+
+    Args:
+           channel_dim: the dimension/axis corresponding to the channel, e.g.
+               -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative.
+           min_positive: the minimum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_positive: the maximum, per channel, of the proportion of the time
+               that (x > 0), below which we start to modify the derivatives.
+           max_factor: the maximum factor by which we modify the derivatives for
+              either the sign constraint or the magnitude constraint;
+              e.g. with max_factor=0.02, the the derivatives would be multiplied by
+              values in the range [0.98..1.02].
+           min_abs:  the minimum average-absolute-value per channel, which
+              we allow, before we start to modify the derivatives to prevent
+              this.
+           max_abs:  the maximum average-absolute-value per channel, which
+               we allow, before we start to modify the derivatives to prevent
+               this.
+    """
+    def __init__(self, channel_dim: int,
+                 min_positive: float = 0.05,
+                 max_positive: float = 0.95,
+                 max_factor: float = 0.01,
+                 min_abs: float = 0.2,
+                 max_abs: float = 100.0):
+        super(ActivationBalancer, self).__init__()
+        self.channel_dim = channel_dim
+        self.min_positive = min_positive
+        self.max_positive = max_positive
+        self.max_factor = max_factor
+        self.min_abs = min_abs
+        self.max_abs = max_abs
+
+    def forward(self, x: Tensor) -> Tensor:
+        return ActivationBalancerFunction.apply(x, self.channel_dim,
+                                                self.min_positive, self.max_positive,
+                                                self.max_factor, self.min_abs,
+                                                self.max_abs)
+
+
+def _double_swish(x: Tensor) -> Tensor:
+    # double-swish, implemented/approximated as offset-swish
+    return x * torch.sigmoid(x - 1.0)
+
+class DoubleSwishFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x: Tensor) -> Tensor:
+        ctx.save_for_backward(x.detach())
+        return _double_swish(x)
+
+    @staticmethod
+    def backward(ctx, y_grad: Tensor) -> Tensor:
+        # TODO: can make this more efficient.
+        x, = ctx.saved_tensors
+        x.requires_grad = True
+        with torch.enable_grad():
+            y = _double_swish(x)
+            y.backward(gradient=y_grad)
+            return x.grad
+
+class DoubleSwish(torch.nn.Module):
+    def forward(self, x: Tensor) -> Tensor:
+        """Return double-swish activation function which is an approximation to Swish(Swish(x)),
+           that we approximate closely with x * sigmoid(x-1).
+        """
+        return DoubleSwishFunction.apply(x)
+
+
+
+def _test_activation_balancer_sign():
+    channel_dim = 0
+    probs = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = 1.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(channel_dim=0, min_positive=0.05, max_positive=0.95,
+                           max_factor=0.2, min_abs=0.0)
+
+    y_grad = torch.sign(torch.randn(probs.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_sign: x = ", x)
+    print("_test_activation_balancer_sign: y grad = ", y_grad)
+    print("_test_activation_balancer_sign: x grad = ", x.grad)
+
+def _test_activation_balancer_magnitude():
+    channel_dim = 0
+    magnitudes = torch.arange(0, 1, 0.01)
+    N = 1000
+    x = torch.sign(torch.randn(magnitudes.numel(), N))  * magnitudes.unsqueeze(-1)
+    x = x.detach()
+    x.requires_grad = True
+    m = ActivationBalancer(channel_dim=0,
+                           min_positive=0.0, max_positive=1.0,
+                           max_factor=0.2,
+                           min_abs=0.2, max_abs=0.8)
+
+    y_grad = torch.sign(torch.randn(magnitudes.numel(), N))
+
+    y = m(x)
+    y.backward(gradient=y_grad)
+    print("_test_activation_balancer_magnitude: x = ", x)
+    print("_test_activation_balancer_magnitude: y grad = ", y_grad)
+    print("_test_activation_balancer_magnitude: x grad = ", x.grad)
+
+
+def _test_basic_norm():
+    num_channels = 128
+    m = BasicNorm(num_channels=num_channels, channel_dim=1)
+
+    x = torch.randn(500, num_channels)
+
+    y = m(x)
+
+    assert y.shape == x.shape
+    x_rms = (x**2).mean().sqrt()
+    y_rms = (y**2).mean().sqrt()
+    print("x rms = ", x_rms)
+    print("y rms = ", y_rms)
+    assert y_rms < x_rms
+    assert y_rms > 0.5 * x_rms
+
+
+
+
+
+if __name__ == '__main__':
+    _test_activation_balancer_sign()
+    _test_activation_balancer_magnitude()
+    _test_basic_norm()

egs/librispeech/ASR/pruned_transducer_stateless2/subsampling.py

@@ -0,0 +1,176 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+from typing import Tuple
+from scaling import DoubleSwish, ActivationBalancer, BasicNorm, ScaledLinear, ScaledConv1d, ScaledConv2d
+
+class Conv2dSubsampling(nn.Module):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' == T//4
+
+    It is based on
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py  # noqa
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        assert idim >= 7
+        super().__init__()
+        self.conv = nn.Sequential(
+            ScaledConv2d(
+                in_channels=1, out_channels=odim, kernel_size=3, stride=2
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+            ScaledConv2d(
+                in_channels=odim, out_channels=odim, kernel_size=3, stride=2
+            ),
+            ActivationBalancer(channel_dim=1),
+            DoubleSwish(),
+        )
+        self.out = ScaledLinear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+        # set learn_eps=False because out_norm is preceded by `out`, and `out`
+        # itself has learned scale, so the extra degree of freedom is not
+        # needed.
+        self.out_norm = BasicNorm(odim, learn_eps=False)
+        # constrain median of output to be close to zero.
+        self.out_balancer = ActivationBalancer(channel_dim=-1,
+                                               min_positive=0.45,
+                                               max_positive=0.55)
+
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        # On entry, x is (N, T, idim)
+        x = x.unsqueeze(1)  # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W)
+        x = self.conv(x)
+        # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        # Now x is of shape (N, ((T-1)//2 - 1))//2, odim)
+        x = self.out_norm(x)
+        x = self.out_balancer(x)
+        return x
+
+
+class VggSubsampling(nn.Module):
+    """Trying to follow the setup described in the following paper:
+    https://arxiv.org/pdf/1910.09799.pdf
+
+    This paper is not 100% explicit so I am guessing to some extent,
+    and trying to compare with other VGG implementations.
+
+    Convert an input of shape (N, T, idim) to an output
+    with shape (N, T', odim), where
+    T' = ((T-1)//2 - 1)//2, which approximates T' = T//4
+    """
+
+    def __init__(self, idim: int, odim: int) -> None:
+        """Construct a VggSubsampling object.
+
+        This uses 2 VGG blocks with 2 Conv2d layers each,
+        subsampling its input by a factor of 4 in the time dimensions.
+
+        Args:
+          idim:
+            Input dim. The input shape is (N, T, idim).
+            Caution: It requires: T >=7, idim >=7
+          odim:
+            Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim)
+        """
+        super().__init__()
+
+        cur_channels = 1
+        layers = []
+        block_dims = [32, 64]
+
+        # The decision to use padding=1 for the 1st convolution, then padding=0
+        # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by
+        # a back-compatibility concern so that the number of frames at the
+        # output would be equal to:
+        #  (((T-1)//2)-1)//2.
+        # We can consider changing this by using padding=1 on the
+        # 2nd convolution, so the num-frames at the output would be T//4.
+        for block_dim in block_dims:
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=cur_channels,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=1,
+                    stride=1,
+                )
+            )
+            layers.append(torch.nn.ReLU())
+            layers.append(
+                torch.nn.Conv2d(
+                    in_channels=block_dim,
+                    out_channels=block_dim,
+                    kernel_size=3,
+                    padding=0,
+                    stride=1,
+                )
+            )
+            layers.append(
+                torch.nn.MaxPool2d(
+                    kernel_size=2, stride=2, padding=0, ceil_mode=True
+                )
+            )
+            cur_channels = block_dim
+
+        self.layers = nn.Sequential(*layers)
+
+        self.out = nn.Linear(
+            block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Subsample x.
+
+        Args:
+          x:
+            Its shape is (N, T, idim).
+
+        Returns:
+          Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim)
+        """
+        x = x.unsqueeze(1)
+        x = self.layers(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        return x

egs/librispeech/ASR/pruned_transducer_stateless2/train.py

@@ -0,0 +1,810 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+"""
+Usage:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --full-libri 1 \
+  --max-duration 300 \
+  --lr-factor 1.5
+"""
+
+
+import argparse
+import logging
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse.cut import Cut
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from torch import Tensor
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+from transformer import Noam
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall import diagnostics
+
+from icefall.utils import (
+    AttributeDict,
+    MetricsTracker,
+    setup_logger,
+    str2bool,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless/exp/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--lr-factor",
+        type=float,
+        default=5.0,
+        help="The lr_factor for Noam optimizer",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)"
+        "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - attention_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 3000,  # For the 100h subset, use 800
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "attention_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            "vgg_frontend": False,
+            # parameters for decoder
+            "embedding_dim": 512,
+            # parameters for Noam
+            "warm_step": 30000,  # For the 100h subset, use 8k
+            "model_warm_step": 3000, # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        output_dim=params.vocab_size,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.attention_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+        vgg_frontend=params.vgg_frontend,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        input_dim=params.vocab_size,
+        inner_dim=params.embedding_dim,
+        output_dim=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    batch: dict,
+    is_training: bool,
+    warmup_mode: bool = False
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+    """
+    device = model.device
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    texts = batch["supervisions"]["text"]
+    y = sp.encode(texts, out_type=int)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup_mode=warmup_mode,
+        )
+        loss = params.simple_loss_scale * simple_loss + pruned_loss
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    info["frames"] = (feature_lens // params.subsampling_factor).sum().item()
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            sp=sp,
+            batch=batch,
+            is_training=True,
+            warmup_mode=(params.batch_idx_train < params.model_warm_step)
+        )
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        # NOTE: We use reduction==sum and loss is computed over utterances
+        # in the batch and there is no normalization to it so far.
+        loss.backward()
+        optimizer.step()
+        optimizer.zero_grad()
+
+        if batch_idx % params.log_interval == 0:
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    if params.full_libri is False:
+        params.valid_interval = 800
+        params.warm_step = 8000
+
+    fix_random_seed(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    optimizer = Noam(
+        model.parameters(),
+        model_size=params.attention_dim,
+        factor=params.lr_factor,
+        warm_step=params.warm_step,
+    )
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+
+    if params.print_diagnostics:
+        diagnostic = diagnostics.attach_diagnostics(model)
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    train_cuts = librispeech.train_clean_100_cuts()
+    if params.full_libri:
+        train_cuts += librispeech.train_clean_360_cuts()
+        train_cuts += librispeech.train_other_500_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        return 1.0 <= c.duration <= 20.0
+
+    num_in_total = len(train_cuts)
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    num_left = len(train_cuts)
+    num_removed = num_in_total - num_left
+    removed_percent = num_removed / num_in_total * 100
+
+    logging.info(f"Before removing short and long utterances: {num_in_total}")
+    logging.info(f"After removing short and long utterances: {num_left}")
+    logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)")
+
+    train_dl = librispeech.train_dataloaders(train_cuts)
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
+
+    if not params.print_diagnostics:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_dl,
+            optimizer=optimizer,
+            sp=sp,
+            params=params,
+        )
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        fix_random_seed(params.seed + epoch)
+        train_dl.sampler.set_epoch(epoch)
+
+        cur_lr = optimizer._rate
+        if tb_writer is not None:
+            tb_writer.add_scalar(
+                "train/learning_rate", cur_lr, params.batch_idx_train
+            )
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        if rank == 0:
+            logging.info("epoch {}, learning rate {}".format(epoch, cur_lr))
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            sp=sp,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    sp: spm.SentencePieceProcessor,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 0 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            loss, _ = compute_loss(
+                params=params,
+                model=model,
+                sp=sp,
+                batch=batch,
+                is_training=True,
+                warmup_mode=True # may use slightly more memory
+            )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless2/transformer.py

@@ -0,0 +1,418 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# 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.
+
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling, ScaledLinear
+
+from icefall.utils import make_pad_mask
+
+
+class Transformer(EncoderInterface):
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        subsampling_factor: int = 4,
+        d_model: int = 256,
+        nhead: int = 4,
+        dim_feedforward: int = 2048,
+        num_encoder_layers: int = 12,
+        dropout: float = 0.1,
+        normalize_before: bool = True,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+            Must satisfy d_model // nhead == 0.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          dropout:
+            Dropout in encoder.
+          normalize_before:
+            If True, use pre-layer norm; False to use post-layer norm.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.num_features = num_features
+        self.output_dim = output_dim
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.encoder_pos = PositionalEncoding(d_model, dropout)
+
+        encoder_layer = TransformerEncoderLayer(
+            d_model=d_model,
+            nhead=nhead,
+            dim_feedforward=dim_feedforward,
+            dropout=dropout,
+            normalize_before=normalize_before,
+        )
+
+        if normalize_before:
+            encoder_norm = nn.LayerNorm(d_model)
+        else:
+            encoder_norm = None
+
+        self.encoder = nn.TransformerEncoder(
+            encoder_layer=encoder_layer,
+            num_layers=num_encoder_layers,
+            norm=encoder_norm,
+        )
+
+        # TODO(fangjun): remove dropout
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), ScaledLinear(d_model, output_dim)
+        )
+
+    def forward(
+        self, x: torch.Tensor, x_lens: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            The input tensor. Its shape is (batch_size, seq_len, feature_dim).
+          x_lens:
+            A tensor of shape (batch_size,) containing the number of frames in
+            `x` before padding.
+        Returns:
+          Return a tuple containing 2 tensors:
+            - logits, its shape is (batch_size, output_seq_len, output_dim)
+            - logit_lens, a tensor of shape (batch_size,) containing the number
+              of frames in `logits` before padding.
+        """
+        x = self.encoder_embed(x)
+        x = self.encoder_pos(x)
+        x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
+
+        # Caution: We assume the subsampling factor is 4!
+        lengths = ((x_lens - 1) // 2 - 1) // 2
+        assert x.size(0) == lengths.max().item()
+
+        mask = make_pad_mask(lengths)
+        x = self.encoder(x, src_key_padding_mask=mask)  # (T, N, C)
+
+        logits = self.encoder_output_layer(x)
+        logits = logits.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
+
+        return logits, lengths
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    Modified from torch.nn.TransformerEncoderLayer.
+    Add support of normalize_before,
+    i.e., use layer_norm before the first block.
+
+    Args:
+      d_model:
+        the number of expected features in the input (required).
+      nhead:
+        the number of heads in the multiheadattention models (required).
+      dim_feedforward:
+        the dimension of the feedforward network model (default=2048).
+      dropout:
+        the dropout value (default=0.1).
+      activation:
+        the activation function of intermediate layer, relu or
+        gelu (default=relu).
+      normalize_before:
+        whether to use layer_norm before the first block.
+
+    Examples::
+        >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8)
+        >>> src = torch.rand(10, 32, 512)
+        >>> out = encoder_layer(src)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int = 2048,
+        dropout: float = 0.1,
+        activation: str = "relu",
+        normalize_before: bool = True,
+    ) -> None:
+        super(TransformerEncoderLayer, self).__init__()
+        self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0)
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+
+        self.activation = _get_activation_fn(activation)
+
+        self.normalize_before = normalize_before
+
+    def __setstate__(self, state):
+        if "activation" not in state:
+            state["activation"] = nn.functional.relu
+        super(TransformerEncoderLayer, self).__setstate__(state)
+
+    def forward(
+        self,
+        src: torch.Tensor,
+        src_mask: Optional[torch.Tensor] = None,
+        src_key_padding_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Pass the input through the encoder layer.
+
+        Args:
+            src: the sequence to the encoder layer (required).
+            src_mask: the mask for the src sequence (optional).
+            src_key_padding_mask: the mask for the src keys per batch (optional)
+
+        Shape:
+            src: (S, N, E).
+            src_mask: (S, S).
+            src_key_padding_mask: (N, S).
+            S is the source sequence length, T is the target sequence length,
+            N is the batch size, E is the feature number
+        """
+        residual = src
+        if self.normalize_before:
+            src = self.norm1(src)
+        src2 = self.self_attn(
+            src,
+            src,
+            src,
+            attn_mask=src_mask,
+            key_padding_mask=src_key_padding_mask,
+        )[0]
+        src = residual + self.dropout1(src2)
+        if not self.normalize_before:
+            src = self.norm1(src)
+
+        residual = src
+        if self.normalize_before:
+            src = self.norm2(src)
+        src2 = self.linear2(self.dropout(self.activation(self.linear1(src))))
+        src = residual + self.dropout2(src2)
+        if not self.normalize_before:
+            src = self.norm2(src)
+        return src
+
+
+def _get_activation_fn(activation: str):
+    if activation == "relu":
+        return nn.functional.relu
+    elif activation == "gelu":
+        return nn.functional.gelu
+
+    raise RuntimeError(
+        "activation should be relu/gelu, not {}".format(activation)
+    )
+
+
+class PositionalEncoding(nn.Module):
+    """This class implements the positional encoding
+    proposed in the following paper:
+
+    - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf
+
+        PE(pos, 2i) = sin(pos / (10000^(2i/d_modle))
+        PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle))
+
+    Note::
+
+      1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model)))
+                               = exp(-1* 2i / d_model * log(100000))
+                               = exp(2i * -(log(10000) / d_model))
+    """
+
+    def __init__(self, d_model: int, dropout: float = 0.1) -> None:
+        """
+        Args:
+          d_model:
+            Embedding dimension.
+          dropout:
+            Dropout probability to be applied to the output of this module.
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout)
+        # not doing: self.pe = None because of errors thrown by torchscript
+        self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32)
+
+    def extend_pe(self, x: torch.Tensor) -> None:
+        """Extend the time t in the positional encoding if required.
+
+        The shape of `self.pe` is (1, T1, d_model). The shape of the input x
+        is (N, T, d_model). If T > T1, then we change the shape of self.pe
+        to (N, T, d_model). Otherwise, nothing is done.
+
+        Args:
+          x:
+            It is a tensor of shape (N, T, C).
+        Returns:
+          Return None.
+        """
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        # Now pe is of shape (1, T, d_model), where T is x.size(1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Add positional encoding.
+
+        Args:
+          x:
+            Its shape is (N, T, C)
+
+        Returns:
+          Return a tensor of shape (N, T, C)
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1), :]
+        return self.dropout(x)
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)

egs/librispeech/ASR/transducer_stateless/encoder_interface.py

@@ -22,7 +22,7 @@ import torch.nn as nn
 
 class EncoderInterface(nn.Module):
     def forward(
-      self, x: torch.Tensor, x_lens: torch.Tensor, warmup_mode: bool
+      self, x: torch.Tensor, x_lens: torch.Tensor
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         """
         Args:
@@ -32,8 +32,6 @@ class EncoderInterface(nn.Module):
           x_lens:
             A tensor of shape (batch_size,) containing the number of frames
             in `x` before padding.
-          warmup_mode: for training only, if true then train in
-            "warmup mode" (use this for the first few thousand minibatches).
         Returns:
           Return a tuple containing two tensors:
             - encoder_out, a tensor of (batch_size, out_seq_len, output_dim)