Support exporting LSTM with projection to ONNX (#621)

* Support exporting LSTM with projection to ONNX

* Add missing files

* small fixes

.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh

@@ -42,7 +42,7 @@ for sym in 1 2 3; do
     --lang-dir $repo/data/lang_char \
     $repo/test_wavs/BAC009S0764W0121.wav \
     $repo/test_wavs/BAC009S0764W0122.wav \
-    $rep/test_wavs/BAC009S0764W0123.wav
+    $repo/test_wavs/BAC009S0764W0123.wav
 done
 
 for method in modified_beam_search beam_search fast_beam_search; do
@@ -55,7 +55,7 @@ for method in modified_beam_search beam_search fast_beam_search; do
     --lang-dir $repo/data/lang_char \
     $repo/test_wavs/BAC009S0764W0121.wav \
     $repo/test_wavs/BAC009S0764W0122.wav \
-    $rep/test_wavs/BAC009S0764W0123.wav
+    $repo/test_wavs/BAC009S0764W0123.wav
 done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"

.github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml

@@ -105,6 +105,47 @@ log "Decode with models exported by torch.jit.trace()"
   $repo/test_wavs/1221-135766-0001.wav \
   $repo/test_wavs/1221-135766-0002.wav
 
+log "Test exporting to ONNX"
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir $repo/exp \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --onnx 1
+
+log "Decode with ONNX models "
+
+./lstm_transducer_stateless2/streaming-onnx-decode.py \
+  --bpe-model-filename $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo//exp/encoder.onnx \
+  --decoder-model-filename $repo/exp/decoder.onnx \
+  --joiner-model-filename $repo/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename $repo/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename $repo/exp/joiner_decoder_proj.onnx \
+ $repo/test_wavs/1089-134686-0001.wav
+
+./lstm_transducer_stateless2/streaming-onnx-decode.py \
+  --bpe-model-filename $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo//exp/encoder.onnx \
+  --decoder-model-filename $repo/exp/decoder.onnx \
+  --joiner-model-filename $repo/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename $repo/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename $repo/exp/joiner_decoder_proj.onnx \
+ $repo/test_wavs/1221-135766-0001.wav
+
+./lstm_transducer_stateless2/streaming-onnx-decode.py \
+  --bpe-model-filename $repo/data/lang_bpe_500/bpe.model \
+  --encoder-model-filename $repo//exp/encoder.onnx \
+  --decoder-model-filename $repo/exp/decoder.onnx \
+  --joiner-model-filename $repo/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename $repo/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename $repo/exp/joiner_decoder_proj.onnx \
+ $repo/test_wavs/1221-135766-0002.wav
+
+
+
 for sym in 1 2 3; do
   log "Greedy search with --max-sym-per-frame $sym"
 
@@ -133,7 +174,7 @@ done
 
 echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
 echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
-if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"ncnn"  ]]; then
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
   mkdir -p lstm_transducer_stateless2/exp
   ln -s $PWD/$repo/exp/pretrained.pt lstm_transducer_stateless2/exp/epoch-999.pt
   ln -s $PWD/$repo/data/lang_bpe_500 data/

.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh

@@ -13,10 +13,14 @@ cd egs/librispeech/ASR
 repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless2-2022-04-29
 
 log "Downloading pre-trained model from $repo_url"
-git lfs install
-git clone $repo_url
+GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
 repo=$(basename $repo_url)
 
+pushd $repo
+git lfs pull --include "data/lang_bpe_500/bpe.model"
+git lfs pull --include "exp/pretrained-epoch-38-avg-10.pt"
+popd
+
 log "Display test files"
 tree $repo/
 soxi $repo/test_wavs/*.wav

.github/workflows/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml

@@ -1,4 +1,4 @@
-name: run-librispeech-lstm-transducer-2022-09-03
+name: run-librispeech-lstm-transducer2-2022-09-03
 
 on:
   push:
@@ -17,8 +17,8 @@ on:
     - cron: "50 15 * * *"
 
 jobs:
-  run_librispeech_pruned_transducer_stateless3_2022_05_13:
-    if: github.event.label.name == 'ncnn' || github.event_name == 'push' || github.event_name == 'schedule'
+  run_librispeech_lstm_transducer_stateless2_2022_09_03:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'ncnn' || github.event.label.name == 'onnx' || github.event_name == 'push' || github.event_name == 'schedule'
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
@@ -110,7 +110,7 @@ jobs:
           .github/scripts/run-librispeech-lstm-transducer-stateless2-2022-09-03.yml
 
       - name: Display decoding results for lstm_transducer_stateless2
-        if: github.event_name == 'schedule' || github.event.label.name == 'ncnn'
+        if: github.event_name == 'schedule'
         shell: bash
         run: |
           cd egs/librispeech/ASR
@@ -130,7 +130,7 @@ jobs:
 
       - name: Upload decoding results for lstm_transducer_stateless2
         uses: actions/upload-artifact@v2
-        if: github.event_name == 'schedule' || github.event.label.name == 'ncnn'
+        if: github.event_name == 'schedule'
         with:
           name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-lstm_transducer_stateless2-2022-09-03
           path: egs/librispeech/ASR/lstm_transducer_stateless2/exp/

egs/librispeech/ASR/lstm_transducer_stateless/lstmp.py

@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py

egs/librispeech/ASR/lstm_transducer_stateless2/export.py

@@ -74,6 +74,29 @@ with the following commands:
     git lfs install
     git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
     # You will find the pre-trained models in icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp
+
+(3) Export to ONNX format
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir ./lstm_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --onnx 1
+
+It will generate the following files in the given `exp_dir`.
+
+    - encoder.onnx
+    - decoder.onnx
+    - joiner.onnx
+    - joiner_encoder_proj.onnx
+    - joiner_decoder_proj.onnx
+
+Please see ./streaming-onnx-decode.py for usage of the generated files
+
+Check
+https://github.com/k2-fsa/sherpa-onnx
+for how to use the exported models outside of icefall.
 """
 
 import argparse
@@ -181,6 +204,23 @@ def get_parser():
         """,
     )
 
+    parser.add_argument(
+        "--onnx",
+        type=str2bool,
+        default=False,
+        help="""If True, --jit and --pnnx are ignored and it exports the model
+        to onnx format. It will generate the following files:
+
+            - encoder.onnx
+            - decoder.onnx
+            - joiner.onnx
+            - joiner_encoder_proj.onnx
+            - joiner_decoder_proj.onnx
+
+        Refer to ./onnx_check.py and ./onnx_pretrained.py for how to use them.
+        """,
+    )
+
     parser.add_argument(
         "--context-size",
         type=int,
@@ -266,6 +306,215 @@ def export_joiner_model_jit_trace(
     logging.info(f"Saved to {joiner_filename}")
 
 
+def export_encoder_model_onnx(
+    encoder_model: nn.Module,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has 3 inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+        - states: a tuple containing:
+            - h0: a tensor of shape (num_layers, N, proj_size)
+            - c0: a tensor of shape (num_layers, N, hidden_size)
+
+    and it has 3 outputs:
+
+        - encoder_out, a tensor of shape (N, T, C)
+        - encoder_out_lens, a tensor of shape (N,)
+        - states: a tuple containing:
+            - next_h0: a tensor of shape (num_layers, N, proj_size)
+            - next_c0: a tensor of shape (num_layers, N, hidden_size)
+
+    Note: The warmup argument is fixed to 1.
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    N = 1
+    x = torch.zeros(N, 9, 80, dtype=torch.float32)
+    x_lens = torch.tensor([9], dtype=torch.int64)
+    h = torch.rand(encoder_model.num_encoder_layers, N, encoder_model.d_model)
+    c = torch.rand(
+        encoder_model.num_encoder_layers, N, encoder_model.rnn_hidden_size
+    )
+
+    warmup = 1.0
+    torch.onnx.export(
+        encoder_model,  # use torch.jit.trace() internally
+        (x, x_lens, (h, c), warmup),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens", "h", "c", "warmup"],
+        output_names=["encoder_out", "encoder_out_lens", "next_h", "next_c"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "h": {1: "N"},
+            "c": {1: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+            "next_h": {1: "N"},
+            "next_c": {1: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_filename}")
+
+
+def export_decoder_model_onnx(
+    decoder_model: nn.Module,
+    decoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the decoder model to ONNX format.
+
+    The exported model has one input:
+
+        - y: a torch.int64 tensor of shape (N, decoder_model.context_size)
+
+    and has one output:
+
+        - decoder_out: a torch.float32 tensor of shape (N, 1, C)
+
+    Note: The argument need_pad is fixed to False.
+
+    Args:
+      decoder_model:
+        The decoder model to be exported.
+      decoder_filename:
+        Filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64)
+    need_pad = False  # Always False, so we can use torch.jit.trace() here
+    # Note(fangjun): torch.jit.trace() is more efficient than torch.jit.script()
+    # in this case
+    torch.onnx.export(
+        decoder_model,
+        (y, need_pad),
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y", "need_pad"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_filename}")
+
+
+def export_joiner_model_onnx(
+    joiner_model: nn.Module,
+    joiner_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the joiner model to ONNX format.
+    The exported joiner model has two inputs:
+
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+
+    The exported encoder_proj model has one input:
+
+        - encoder_out: a tensor of shape (N, encoder_out_dim)
+
+    and produces one output:
+
+        - projected_encoder_out: a tensor of shape (N, joiner_dim)
+
+    The exported decoder_proj model has one input:
+
+        - decoder_out: a tensor of shape (N, decoder_out_dim)
+
+    and produces one output:
+
+        - projected_decoder_out: a tensor of shape (N, joiner_dim)
+    """
+    encoder_proj_filename = str(joiner_filename).replace(
+        ".onnx", "_encoder_proj.onnx"
+    )
+
+    decoder_proj_filename = str(joiner_filename).replace(
+        ".onnx", "_decoder_proj.onnx"
+    )
+
+    encoder_out_dim = joiner_model.encoder_proj.weight.shape[1]
+    decoder_out_dim = joiner_model.decoder_proj.weight.shape[1]
+    joiner_dim = joiner_model.decoder_proj.weight.shape[0]
+
+    projected_encoder_out = torch.rand(1, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(1, joiner_dim, dtype=torch.float32)
+
+    project_input = False
+    # Note: It uses torch.jit.trace() internally
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out, project_input),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "projected_encoder_out",
+            "projected_decoder_out",
+            "project_input",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "projected_encoder_out": {0: "N"},
+            "projected_decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {joiner_filename}")
+
+    encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.encoder_proj,
+        encoder_out,
+        encoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["encoder_out"],
+        output_names=["projected_encoder_out"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "projected_encoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {encoder_proj_filename}")
+
+    decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32)
+    torch.onnx.export(
+        joiner_model.decoder_proj,
+        decoder_out,
+        decoder_proj_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["decoder_out"],
+        output_names=["projected_decoder_out"],
+        dynamic_axes={
+            "decoder_out": {0: "N"},
+            "projected_decoder_out": {0: "N"},
+        },
+    )
+    logging.info(f"Saved to {decoder_proj_filename}")
+
+
 @torch.no_grad()
 def main():
     args = get_parser().parse_args()
@@ -387,7 +636,33 @@ def main():
     model.to("cpu")
     model.eval()
 
-    if params.pnnx:
+    if params.onnx:
+        logging.info("Export model to ONNX format")
+        convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True)
+
+        opset_version = 11
+        encoder_filename = params.exp_dir / "encoder.onnx"
+        export_encoder_model_onnx(
+            model.encoder,
+            encoder_filename,
+            opset_version=opset_version,
+        )
+
+        decoder_filename = params.exp_dir / "decoder.onnx"
+        export_decoder_model_onnx(
+            model.decoder,
+            decoder_filename,
+            opset_version=opset_version,
+        )
+
+        joiner_filename = params.exp_dir / "joiner.onnx"
+        export_joiner_model_onnx(
+            model.joiner,
+            joiner_filename,
+            opset_version=opset_version,
+        )
+
+    elif params.pnnx:
         convert_scaled_to_non_scaled(model, inplace=True)
         logging.info("Using torch.jit.trace()")
         encoder_filename = params.exp_dir / "encoder_jit_trace-pnnx.pt"

egs/librispeech/ASR/lstm_transducer_stateless2/lstmp.py

@@ -0,0 +1,102 @@
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class LSTMP(nn.Module):
+    """LSTM with projection.
+
+    PyTorch does not support exporting LSTM with projection to ONNX.
+    This class reimplements LSTM with projection using basic matrix-matrix
+    and matrix-vector operations. It is not intended for training.
+    """
+
+    def __init__(self, lstm: nn.LSTM):
+        """
+        Args:
+          lstm:
+            LSTM with proj_size. We support only uni-directional,
+            1-layer LSTM with projection at present.
+        """
+        super().__init__()
+        assert lstm.bidirectional is False, lstm.bidirectional
+        assert lstm.num_layers == 1, lstm.num_layers
+        assert 0 < lstm.proj_size < lstm.hidden_size, (
+            lstm.proj_size,
+            lstm.hidden_size,
+        )
+
+        assert lstm.batch_first is False, lstm.batch_first
+
+        state_dict = lstm.state_dict()
+
+        w_ih = state_dict["weight_ih_l0"]
+        w_hh = state_dict["weight_hh_l0"]
+
+        b_ih = state_dict["bias_ih_l0"]
+        b_hh = state_dict["bias_hh_l0"]
+
+        w_hr = state_dict["weight_hr_l0"]
+        self.input_size = lstm.input_size
+        self.proj_size = lstm.proj_size
+        self.hidden_size = lstm.hidden_size
+
+        self.w_ih = w_ih
+        self.w_hh = w_hh
+        self.b = b_ih + b_hh
+        self.w_hr = w_hr
+
+    def forward(
+        self,
+        input: torch.Tensor,
+        hx: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+          input:
+            A tensor of shape [T, N, hidden_size]
+          hx:
+            A tuple containing:
+              - h0: a tensor of shape (1, N, proj_size)
+              - c0: a tensor of shape (1, N, hidden_size)
+        Returns:
+          Return a tuple containing:
+            - output: a tensor of shape (T, N, proj_size).
+            - A tuple containing:
+               - h: a tensor of shape (1, N, proj_size)
+               - c: a tensor of shape (1, N, hidden_size)
+
+        """
+        x_list = input.unbind(dim=0)  # We use batch_first=False
+
+        if hx is not None:
+            h0, c0 = hx
+        else:
+            h0 = torch.zeros(1, input.size(1), self.proj_size)
+            c0 = torch.zeros(1, input.size(1), self.hidden_size)
+        h0 = h0.squeeze(0)
+        c0 = c0.squeeze(0)
+        y_list = []
+        for x in x_list:
+            gates = F.linear(x, self.w_ih, self.b) + F.linear(h0, self.w_hh)
+            i, f, g, o = gates.chunk(4, dim=1)
+
+            i = i.sigmoid()
+            f = f.sigmoid()
+            g = g.tanh()
+            o = o.sigmoid()
+
+            c = f * c0 + i * g
+            h = o * c.tanh()
+
+            h = F.linear(h, self.w_hr)
+            y_list.append(h)
+
+            c0 = c
+            h0 = h
+
+        y = torch.stack(y_list, dim=0)
+
+        return y, (h0.unsqueeze(0), c0.unsqueeze(0))

egs/librispeech/ASR/lstm_transducer_stateless2/streaming-ncnn-decode.py

@@ -233,13 +233,12 @@ def greedy_search(
             hyp, dtype=torch.int32
         )  # (1, context_size)
         decoder_out = model.run_decoder(decoder_input).squeeze(0)
-
     else:
         assert decoder_out.ndim == 1
         assert hyp is not None, hyp
 
     joiner_out = model.run_joiner(encoder_out, decoder_out)
-    y = joiner_out.argmax(dim=0).tolist()
+    y = joiner_out.argmax(dim=0).item()
     if y != blank_id:
         hyp.append(y)
         decoder_input = hyp[-context_size:]

egs/librispeech/ASR/lstm_transducer_stateless2/streaming-onnx-decode.py

@@ -0,0 +1,478 @@
+#!/usr/bin/env python3
+# Copyright      2022  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.
+"""
+This script loads ONNX models and uses them to decode waves.
+You can use the following command to get the exported models:
+
+./lstm_transducer_stateless2/export.py \
+  --exp-dir ./lstm_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10 \
+  --onnx 1
+
+Usage of this script:
+
+./lstm_transducer_stateless2/onnx-streaming-decode.py \
+  --encoder-model-filename ./lstm_transducer_stateless2/exp/encoder.onnx \
+  --decoder-model-filename ./lstm_transducer_stateless2/exp/decoder.onnx \
+  --joiner-model-filename ./lstm_transducer_stateless2/exp/joiner.onnx \
+  --joiner-encoder-proj-model-filename ./lstm_transducer_stateless2/exp/joiner_encoder_proj.onnx \
+  --joiner-decoder-proj-model-filename ./lstm_transducer_stateless2/exp/joiner_decoder_proj.onnx \
+  --bpe-model ./data/lang_bpe_500/bpe.model \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+from typing import List, Optional, Tuple
+
+import onnxruntime as ort
+import sentencepiece as spm
+import torch
+import torchaudio
+from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature
+
+
+def get_args():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--bpe-model-filename",
+        type=str,
+        help="Path to bpe.model",
+    )
+
+    parser.add_argument(
+        "--encoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the encoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--decoder-model-filename",
+        type=str,
+        required=True,
+        help="Path to the decoder onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-encoder-proj-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner encoder_proj onnx model. ",
+    )
+
+    parser.add_argument(
+        "--joiner-decoder-proj-model-filename",
+        type=str,
+        required=True,
+        help="Path to the joiner decoder_proj onnx model. ",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="""Path to bpe.model.""",
+    )
+
+    parser.add_argument(
+        "sound_filename",
+        type=str,
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="Context size of the decoder model",
+    )
+
+    return parser.parse_args()
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+class Model:
+    def __init__(self, args):
+        session_opts = ort.SessionOptions()
+        session_opts.inter_op_num_threads = 5
+        session_opts.intra_op_num_threads = 5
+        self.session_opts = session_opts
+
+        self.init_encoder(args)
+        self.init_decoder(args)
+        self.init_joiner(args)
+        self.init_joiner_encoder_proj(args)
+        self.init_joiner_decoder_proj(args)
+
+    def init_encoder(self, args):
+        self.encoder = ort.InferenceSession(
+            args.encoder_model_filename,
+            sess_options=self.session_opts,
+        )
+
+    def init_decoder(self, args):
+        self.decoder = ort.InferenceSession(
+            args.decoder_model_filename,
+            sess_options=self.session_opts,
+        )
+
+    def init_joiner(self, args):
+        self.joiner = ort.InferenceSession(
+            args.joiner_model_filename,
+            sess_options=self.session_opts,
+        )
+
+    def init_joiner_encoder_proj(self, args):
+        self.joiner_encoder_proj = ort.InferenceSession(
+            args.joiner_encoder_proj_model_filename,
+            sess_options=self.session_opts,
+        )
+
+    def init_joiner_decoder_proj(self, args):
+        self.joiner_decoder_proj = ort.InferenceSession(
+            args.joiner_decoder_proj_model_filename,
+            sess_options=self.session_opts,
+        )
+
+    def run_encoder(
+        self, x, h0, c0
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            A tensor of shape (N, T, C)
+          h0:
+            A tensor of shape (num_layers, N, proj_size)
+          c0:
+            A tensor of shape (num_layers, N, hidden_size)
+        Returns:
+          Return a tuple containing:
+            - encoder_out: A tensor of shape (N, T', C')
+            - next_h0: A tensor of shape (num_layers, N, proj_size)
+            - next_c0: A tensor of shape (num_layers, N, hidden_size)
+        """
+        encoder_input_nodes = self.encoder.get_inputs()
+        encoder_out_nodes = self.encoder.get_outputs()
+        x_lens = torch.tensor([x.size(1)], dtype=torch.int64)
+
+        encoder_out, encoder_out_lens, next_h0, next_c0 = self.encoder.run(
+            [
+                encoder_out_nodes[0].name,
+                encoder_out_nodes[1].name,
+                encoder_out_nodes[2].name,
+                encoder_out_nodes[3].name,
+            ],
+            {
+                encoder_input_nodes[0].name: x.numpy(),
+                encoder_input_nodes[1].name: x_lens.numpy(),
+                encoder_input_nodes[2].name: h0.numpy(),
+                encoder_input_nodes[3].name: c0.numpy(),
+            },
+        )
+        return (
+            torch.from_numpy(encoder_out),
+            torch.from_numpy(next_h0),
+            torch.from_numpy(next_c0),
+        )
+
+    def run_decoder(self, decoder_input: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          decoder_input:
+            A tensor of shape (N, context_size). Its dtype is torch.int64.
+        Returns:
+          Return a tensor of shape (N, 1, decoder_out_dim).
+        """
+        decoder_input_nodes = self.decoder.get_inputs()
+        decoder_output_nodes = self.decoder.get_outputs()
+
+        decoder_out = self.decoder.run(
+            [decoder_output_nodes[0].name],
+            {
+                decoder_input_nodes[0].name: decoder_input.numpy(),
+            },
+        )[0]
+
+        return self.run_joiner_decoder_proj(
+            torch.from_numpy(decoder_out).squeeze(1)
+        )
+
+    def run_joiner(
+        self,
+        projected_encoder_out: torch.Tensor,
+        projected_decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          projected_encoder_out:
+            A tensor of shape (N, joiner_dim)
+          projected_decoder_out:
+            A tensor of shape (N, joiner_dim)
+        Returns:
+          Return a tensor of shape (N, vocab_size)
+        """
+        joiner_input_nodes = self.joiner.get_inputs()
+        joiner_output_nodes = self.joiner.get_outputs()
+
+        logits = self.joiner.run(
+            [joiner_output_nodes[0].name],
+            {
+                joiner_input_nodes[0].name: projected_encoder_out.numpy(),
+                joiner_input_nodes[1].name: projected_decoder_out.numpy(),
+            },
+        )[0]
+
+        return torch.from_numpy(logits)
+
+    def run_joiner_encoder_proj(
+        self,
+        encoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A tensor of shape (N, encoder_out_dim)
+        Returns:
+            A tensor of shape (N, joiner_dim)
+        """
+
+        projected_encoder_out = self.joiner_encoder_proj.run(
+            [self.joiner_encoder_proj.get_outputs()[0].name],
+            {
+                self.joiner_encoder_proj.get_inputs()[
+                    0
+                ].name: encoder_out.numpy()
+            },
+        )[0]
+
+        return torch.from_numpy(projected_encoder_out)
+
+    def run_joiner_decoder_proj(
+        self,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          decoder_out:
+            A tensor of shape (N, decoder_out_dim)
+        Returns:
+            A tensor of shape (N, joiner_dim)
+        """
+
+        projected_decoder_out = self.joiner_decoder_proj.run(
+            [self.joiner_decoder_proj.get_outputs()[0].name],
+            {
+                self.joiner_decoder_proj.get_inputs()[
+                    0
+                ].name: decoder_out.numpy()
+            },
+        )[0]
+
+        return torch.from_numpy(projected_decoder_out)
+
+
+def create_streaming_feature_extractor() -> OnlineFeature:
+    """Create a CPU streaming feature extractor.
+
+    At present, we assume it returns a fbank feature extractor with
+    fixed options. In the future, we will support passing in the options
+    from outside.
+
+    Returns:
+      Return a CPU streaming feature extractor.
+    """
+    opts = FbankOptions()
+    opts.device = "cpu"
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = 16000
+    opts.mel_opts.num_bins = 80
+    return OnlineFbank(opts)
+
+
+def greedy_search(
+    model: Model,
+    encoder_out: torch.Tensor,
+    decoder_out: Optional[torch.Tensor] = None,
+    hyp: Optional[List[int]] = None,
+):
+    assert encoder_out.ndim == 2
+    assert encoder_out.shape[0] == 1, "TODO: support batch_size > 1"
+    context_size = 2
+    blank_id = 0
+
+    if decoder_out is None:
+        assert hyp is None, hyp
+        hyp = [blank_id] * context_size
+        decoder_input = torch.tensor(
+            [hyp], dtype=torch.int64
+        )  # (1, context_size)
+        decoder_out = model.run_decoder(decoder_input)
+    else:
+        assert decoder_out.shape[0] == 1
+        assert hyp is not None, hyp
+
+    projected_encoder_out = model.run_joiner_encoder_proj(encoder_out)
+
+    joiner_out = model.run_joiner(projected_encoder_out, decoder_out)
+    y = joiner_out.squeeze(0).argmax(dim=0).item()
+
+    if y != blank_id:
+        hyp.append(y)
+        decoder_input = hyp[-context_size:]
+        decoder_input = torch.tensor([decoder_input], dtype=torch.int64)
+        decoder_out = model.run_decoder(decoder_input)
+
+    return hyp, decoder_out
+
+
+def main():
+    args = get_args()
+    logging.info(vars(args))
+
+    model = Model(args)
+
+    sound_file = args.sound_filename
+    sample_rate = 16000
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model_filename)
+
+    logging.info("Constructing Fbank computer")
+    online_fbank = create_streaming_feature_extractor()
+
+    logging.info(f"Reading sound files: {sound_file}")
+    wave_samples = read_sound_files(
+        filenames=[sound_file],
+        expected_sample_rate=sample_rate,
+    )[0]
+    logging.info(wave_samples.shape)
+
+    num_encoder_layers = 12
+    batch_size = 1
+    d_model = 512
+    rnn_hidden_size = 1024
+
+    h0 = torch.zeros(num_encoder_layers, batch_size, d_model)
+    c0 = torch.zeros(num_encoder_layers, batch_size, rnn_hidden_size)
+
+    hyp = None
+    decoder_out = None
+
+    num_processed_frames = 0
+    segment = 9
+    offset = 4
+
+    chunk = 3200  # 0.2 second
+
+    start = 0
+    while start < wave_samples.numel():
+        end = min(start + chunk, wave_samples.numel())
+        samples = wave_samples[start:end]
+        start += chunk
+
+        online_fbank.accept_waveform(
+            sampling_rate=sample_rate,
+            waveform=samples,
+        )
+        while online_fbank.num_frames_ready - num_processed_frames >= segment:
+            frames = []
+            for i in range(segment):
+                frames.append(online_fbank.get_frame(num_processed_frames + i))
+
+            num_processed_frames += offset
+            frames = torch.cat(frames, dim=0).unsqueeze(0)
+            encoder_out, h0, c0 = model.run_encoder(frames, h0, c0)
+            hyp, decoder_out = greedy_search(
+                model, encoder_out.squeeze(0), decoder_out, hyp
+            )
+    online_fbank.accept_waveform(
+        sampling_rate=sample_rate, waveform=torch.zeros(5000, dtype=torch.float)
+    )
+
+    online_fbank.input_finished()
+    while online_fbank.num_frames_ready - num_processed_frames >= segment:
+        frames = []
+        for i in range(segment):
+            frames.append(online_fbank.get_frame(num_processed_frames + i))
+        num_processed_frames += offset
+        frames = torch.cat(frames, dim=0).unsqueeze(0)
+        encoder_out, h0, c0 = model.run_encoder(frames, h0, c0)
+        hyp, decoder_out = greedy_search(
+            model, encoder_out.squeeze(0), decoder_out, hyp
+        )
+
+    context_size = 2
+
+    logging.info(sound_file)
+    logging.info(sp.decode(hyp[context_size:]))
+
+
+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/lstm_transducer_stateless2/test_lstmp.py

@@ -0,0 +1,70 @@
+#!/usr/bin/env python3
+
+import torch
+import torch.nn as nn
+from lstmp import LSTMP
+
+
+def test():
+    input_size = torch.randint(low=10, high=1024, size=(1,)).item()
+    hidden_size = torch.randint(low=10, high=1024, size=(1,)).item()
+    proj_size = hidden_size - 1
+    lstm = nn.LSTM(
+        input_size=input_size,
+        hidden_size=hidden_size,
+        num_layers=1,
+        bias=True,
+        proj_size=proj_size,
+    )
+    lstmp = LSTMP(lstm)
+
+    N = torch.randint(low=1, high=10, size=(1,)).item()
+    T = torch.randint(low=1, high=20, size=(1,)).item()
+    x = torch.rand(T, N, input_size)
+    h0 = torch.rand(1, N, proj_size)
+    c0 = torch.rand(1, N, hidden_size)
+
+    y1, (h1, c1) = lstm(x, (h0, c0))
+    y2, (h2, c2) = lstmp(x, (h0, c0))
+
+    assert torch.allclose(y1, y2, atol=1e-5), (y1 - y2).abs().max()
+    assert torch.allclose(h1, h2, atol=1e-5), (h1 - h2).abs().max()
+    assert torch.allclose(c1, c2, atol=1e-5), (c1 - c2).abs().max()
+
+    #  lstm_script = torch.jit.script(lstm) # pytorch does not support it
+    lstm_script = lstm
+    lstmp_script = torch.jit.script(lstmp)
+
+    y3, (h3, c3) = lstm_script(x, (h0, c0))
+    y4, (h4, c4) = lstmp_script(x, (h0, c0))
+
+    assert torch.allclose(y3, y4, atol=1e-5), (y3 - y4).abs().max()
+    assert torch.allclose(h3, h4, atol=1e-5), (h3 - h4).abs().max()
+    assert torch.allclose(c3, c4, atol=1e-5), (c3 - c4).abs().max()
+
+    assert torch.allclose(y3, y1, atol=1e-5), (y3 - y1).abs().max()
+    assert torch.allclose(h3, h1, atol=1e-5), (h3 - h1).abs().max()
+    assert torch.allclose(c3, c1, atol=1e-5), (c3 - c1).abs().max()
+
+    lstm_trace = torch.jit.trace(lstm, (x, (h0, c0)))
+    lstmp_trace = torch.jit.trace(lstmp, (x, (h0, c0)))
+
+    y5, (h5, c5) = lstm_trace(x, (h0, c0))
+    y6, (h6, c6) = lstmp_trace(x, (h0, c0))
+
+    assert torch.allclose(y5, y6, atol=1e-5), (y5 - y6).abs().max()
+    assert torch.allclose(h5, h6, atol=1e-5), (h5 - h6).abs().max()
+    assert torch.allclose(c5, c6, atol=1e-5), (c5 - c6).abs().max()
+
+    assert torch.allclose(y5, y1, atol=1e-5), (y5 - y1).abs().max()
+    assert torch.allclose(h5, h1, atol=1e-5), (h5 - h1).abs().max()
+    assert torch.allclose(c5, c1, atol=1e-5), (c5 - c1).abs().max()
+
+
+@torch.no_grad()
+def main():
+    test()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/lstm_transducer_stateless3/lstmp.py

@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py

egs/librispeech/ASR/pruned_transducer_stateless3/lstmp.py

@@ -0,0 +1 @@
+../lstm_transducer_stateless2/lstmp.py

egs/librispeech/ASR/pruned_transducer_stateless3/scaling_converter.py

@@ -29,6 +29,7 @@ from typing import List
 
 import torch
 import torch.nn as nn
+from lstmp import LSTMP
 from scaling import (
     ActivationBalancer,
     BasicNorm,
@@ -259,7 +260,11 @@ def get_submodule(model, target):
     return mod
 
 
-def convert_scaled_to_non_scaled(model: nn.Module, inplace: bool = False):
+def convert_scaled_to_non_scaled(
+    model: nn.Module,
+    inplace: bool = False,
+    is_onnx: bool = False,
+):
     """Convert `ScaledLinear`, `ScaledConv1d`, and `ScaledConv2d`
     in the given modle to their unscaled version `nn.Linear`, `nn.Conv1d`,
     and `nn.Conv2d`.
@@ -270,6 +275,9 @@ def convert_scaled_to_non_scaled(model: nn.Module, inplace: bool = False):
       inplace:
         If True, the input model is modified inplace.
         If False, the input model is copied and we modify the copied version.
+      is_onnx:
+        If True, we are going to export the model to ONNX. In this case,
+        we will convert nn.LSTM with proj_size to LSTMP.
     Return:
       Return a model without scaled layers.
     """
@@ -294,7 +302,13 @@ def convert_scaled_to_non_scaled(model: nn.Module, inplace: bool = False):
         elif isinstance(m, BasicNorm):
             d[name] = convert_basic_norm(m)
         elif isinstance(m, ScaledLSTM):
-            d[name] = scaled_lstm_to_lstm(m)
+            if is_onnx:
+                d[name] = LSTMP(scaled_lstm_to_lstm(m))
+                # See
+                # https://github.com/pytorch/pytorch/issues/47887
+                #  d[name] = torch.jit.script(LSTMP(scaled_lstm_to_lstm(m)))
+            else:
+                d[name] = scaled_lstm_to_lstm(m)
         elif isinstance(m, ActivationBalancer):
             d[name] = nn.Identity()
 

egs/wenetspeech/ASR/pruned_transducer_stateless2/lstmp.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/lstm_transducer_stateless2/lstmp.py