Add onnx export support for pruned_transducer_stateless5 (#883)

.github/scripts/test-onnx-export.sh

@@ -120,3 +120,75 @@ log "Run onnx_pretrained.py"
   $repo/test_wavs/1089-134686-0001.wav \
   $repo/test_wavs/1221-135766-0001.wav \
   $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"
+
+
+log "=========================================================================="
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+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-39-avg-7.pt"
+
+cd exp
+ln -s pretrained-epoch-39-avg-7.pt epoch-99.pt
+popd
+
+log "Export via torch.jit.script()"
+
+./pruned_transducer_stateless5/export.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --jit 1
+
+log "Test exporting to ONNX format"
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+ls -lh $repo/exp
+
+log "Run onnx_check.py"
+
+./pruned_transducer_stateless5/onnx_check.py \
+  --jit-filename $repo/exp/cpu_jit.pt \
+  --onnx-encoder-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --onnx-decoder-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --onnx-joiner-filename $repo/exp/joiner-epoch-99-avg-1.onnx
+
+log "Run onnx_pretrained.py"
+
+./pruned_transducer_stateless5/onnx_pretrained.py \
+  --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \
+  --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \
+  --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \
+  --tokens $repo/data/lang_bpe_500/tokens.txt \
+  $repo/test_wavs/1089-134686-0001.wav \
+  $repo/test_wavs/1221-135766-0001.wav \
+  $repo/test_wavs/1221-135766-0002.wav
+
+rm -rf $repo
+log "--------------------------------------------------------------------------"

egs/librispeech/ASR/pruned_transducer_stateless3/export-onnx.py

@@ -6,7 +6,7 @@
 This script exports a transducer model from PyTorch to ONNX.
 
 We use the pre-trained model from
-https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-streaming-2022-12-29
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-05-13
 as an example to show how to use this file.
 
 1. Download the pre-trained model

egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py

@@ -32,7 +32,7 @@ from scaling import (
 )
 from torch import Tensor, nn
 
-from icefall.utils import make_pad_mask, subsequent_chunk_mask
+from icefall.utils import is_jit_tracing, make_pad_mask, subsequent_chunk_mask
 
 
 class Conformer(EncoderInterface):
@@ -1012,15 +1012,28 @@ class RelPositionMultiheadAttention(nn.Module):
             n == left_context + 2 * time1 - 1
         ), f"{n} == {left_context} + 2 * {time1} - 1"
         # Note: TorchScript requires explicit arg for stride()
-        batch_stride = x.stride(0)
-        head_stride = x.stride(1)
-        time1_stride = x.stride(2)
-        n_stride = x.stride(3)
-        return x.as_strided(
-            (batch_size, num_heads, time1, time2),
-            (batch_stride, head_stride, time1_stride - n_stride, n_stride),
-            storage_offset=n_stride * (time1 - 1),
-        )
+
+        if is_jit_tracing():
+            rows = torch.arange(start=time1 - 1, end=-1, step=-1)
+            cols = torch.arange(time2)
+            rows = rows.repeat(batch_size * num_heads).unsqueeze(-1)
+            indexes = rows + cols
+
+            x = x.reshape(-1, n)
+            x = torch.gather(x, dim=1, index=indexes)
+            x = x.reshape(batch_size, num_heads, time1, time2)
+            return x
+        else:
+            # Note: TorchScript requires explicit arg for stride()
+            batch_stride = x.stride(0)
+            head_stride = x.stride(1)
+            time1_stride = x.stride(2)
+            n_stride = x.stride(3)
+            return x.as_strided(
+                (batch_size, num_heads, time1, time2),
+                (batch_stride, head_stride, time1_stride - n_stride, n_stride),
+                storage_offset=n_stride * (time1 - 1),
+            )
 
     def multi_head_attention_forward(
         self,

egs/librispeech/ASR/pruned_transducer_stateless5/export-onnx.py

@@ -0,0 +1,565 @@
+#!/usr/bin/env python3
+#
+# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang)
+
+"""
+This script exports a transducer model from PyTorch to ONNX.
+
+We use the pre-trained model from
+https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+as an example to show how to use this file.
+
+1. Download the pre-trained model
+
+cd egs/librispeech/ASR
+
+repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless5-2022-05-13
+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-39-avg-7.pt"
+
+cd exp
+ln -s pretrained-epoch-39-avg-7.pt epoch-99.pt
+popd
+
+2. Export the model to ONNX
+
+./pruned_transducer_stateless5/export-onnx.py \
+  --bpe-model $repo/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --exp-dir $repo/exp \
+  --num-encoder-layers 18 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --decoder-dim 512 \
+  --joiner-dim 512
+
+It will generate the following 3 files inside $repo/exp:
+
+  - encoder-epoch-99-avg-1.onnx
+  - decoder-epoch-99-avg-1.onnx
+  - joiner-epoch-99-avg-1.onnx
+
+See ./onnx_pretrained.py and ./onnx_check.py for how to
+use the exported ONNX models.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+from typing import Dict, Tuple
+
+import onnx
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from conformer import Conformer
+from decoder import Decoder
+from scaling_converter import convert_scaled_to_non_scaled
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import setup_logger, 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 averaging.
+        Note: Epoch counts from 0.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless5/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(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def add_meta_data(filename: str, meta_data: Dict[str, str]):
+    """Add meta data to an ONNX model. It is changed in-place.
+
+    Args:
+      filename:
+        Filename of the ONNX model to be changed.
+      meta_data:
+        Key-value pairs.
+    """
+    model = onnx.load(filename)
+    for key, value in meta_data.items():
+        meta = model.metadata_props.add()
+        meta.key = key
+        meta.value = value
+
+    onnx.save(model, filename)
+
+
+class OnnxEncoder(nn.Module):
+    """A wrapper for Conformer and the encoder_proj from the joiner"""
+
+    def __init__(self, encoder: Conformer, encoder_proj: nn.Linear):
+        """
+        Args:
+          encoder:
+            A Conformer encoder.
+          encoder_proj:
+            The projection layer for encoder from the joiner.
+        """
+        super().__init__()
+        self.encoder = encoder
+        self.encoder_proj = encoder_proj
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Please see the help information of Conformer.forward
+
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C)
+          x_lens:
+            A 1-D tensor of shape (N,). Its dtype is torch.int64
+        Returns:
+          Return a tuple containing:
+            - encoder_out, A 3-D tensor of shape (N, T', joiner_dim)
+            - encoder_out_lens, A 1-D tensor of shape (N,)
+        """
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens)
+
+        encoder_out = self.encoder_proj(encoder_out)
+        # Now encoder_out is of shape (N, T, joiner_dim)
+
+        return encoder_out, encoder_out_lens
+
+
+class OnnxDecoder(nn.Module):
+    """A wrapper for Decoder and the decoder_proj from the joiner"""
+
+    def __init__(self, decoder: Decoder, decoder_proj: nn.Linear):
+        super().__init__()
+        self.decoder = decoder
+        self.decoder_proj = decoder_proj
+
+    def forward(self, y: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+          y:
+            A 2-D tensor of shape (N, context_size).
+        Returns
+          Return a 2-D tensor of shape (N, joiner_dim)
+        """
+        need_pad = False
+        decoder_output = self.decoder(y, need_pad=need_pad)
+        decoder_output = decoder_output.squeeze(1)
+        output = self.decoder_proj(decoder_output)
+
+        return output
+
+
+class OnnxJoiner(nn.Module):
+    """A wrapper for the joiner"""
+
+    def __init__(self, output_linear: nn.Linear):
+        super().__init__()
+        self.output_linear = output_linear
+
+    def forward(
+        self,
+        encoder_out: torch.Tensor,
+        decoder_out: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Args:
+          encoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+          decoder_out:
+            A 2-D tensor of shape (N, joiner_dim)
+        Returns:
+          Return a 2-D tensor of shape (N, vocab_size)
+        """
+        logit = encoder_out + decoder_out
+        logit = self.output_linear(torch.tanh(logit))
+        return logit
+
+
+def export_encoder_model_onnx(
+    encoder_model: OnnxEncoder,
+    encoder_filename: str,
+    opset_version: int = 11,
+) -> None:
+    """Export the given encoder model to ONNX format.
+    The exported model has two inputs:
+
+        - x, a tensor of shape (N, T, C); dtype is torch.float32
+        - x_lens, a tensor of shape (N,); dtype is torch.int64
+
+    and it has two outputs:
+
+        - encoder_out, a tensor of shape (N, T', joiner_dim)
+        - encoder_out_lens, a tensor of shape (N,)
+
+    Args:
+      encoder_model:
+        The input encoder model
+      encoder_filename:
+        The filename to save the exported ONNX model.
+      opset_version:
+        The opset version to use.
+    """
+    x = torch.zeros(1, 100, 80, dtype=torch.float32)
+    x_lens = torch.tensor([100], dtype=torch.int64)
+
+    torch.onnx.export(
+        encoder_model,
+        (x, x_lens),
+        encoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["x", "x_lens"],
+        output_names=["encoder_out", "encoder_out_lens"],
+        dynamic_axes={
+            "x": {0: "N", 1: "T"},
+            "x_lens": {0: "N"},
+            "encoder_out": {0: "N", 1: "T"},
+            "encoder_out_lens": {0: "N"},
+        },
+    )
+
+
+def export_decoder_model_onnx(
+    decoder_model: OnnxDecoder,
+    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, joiner_dim)
+
+    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.
+    """
+    context_size = decoder_model.decoder.context_size
+    vocab_size = decoder_model.decoder.vocab_size
+
+    y = torch.zeros(10, context_size, dtype=torch.int64)
+    torch.onnx.export(
+        decoder_model,
+        y,
+        decoder_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=["y"],
+        output_names=["decoder_out"],
+        dynamic_axes={
+            "y": {0: "N"},
+            "decoder_out": {0: "N"},
+        },
+    )
+
+    meta_data = {
+        "context_size": str(context_size),
+        "vocab_size": str(vocab_size),
+    }
+    add_meta_data(filename=decoder_filename, meta_data=meta_data)
+
+
+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:
+
+        - encoder_out: a tensor of shape (N, joiner_dim)
+        - decoder_out: a tensor of shape (N, joiner_dim)
+
+    and produces one output:
+
+        - logit: a tensor of shape (N, vocab_size)
+    """
+    joiner_dim = joiner_model.output_linear.weight.shape[1]
+    logging.info(f"joiner dim: {joiner_dim}")
+
+    projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+    projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32)
+
+    torch.onnx.export(
+        joiner_model,
+        (projected_encoder_out, projected_decoder_out),
+        joiner_filename,
+        verbose=False,
+        opset_version=opset_version,
+        input_names=[
+            "encoder_out",
+            "decoder_out",
+        ],
+        output_names=["logit"],
+        dynamic_axes={
+            "encoder_out": {0: "N"},
+            "decoder_out": {0: "N"},
+            "logit": {0: "N"},
+        },
+    )
+    meta_data = {
+        "joiner_dim": str(joiner_dim),
+    }
+    add_meta_data(filename=joiner_filename, meta_data=meta_data)
+
+
+@torch.no_grad()
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    setup_logger(f"{params.exp_dir}/log-export/log-export-onnx")
+
+    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 not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif 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 i >= 1:
+                    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))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
+                : params.avg + 1
+            ]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    convert_scaled_to_non_scaled(model, inplace=True)
+
+    encoder = OnnxEncoder(
+        encoder=model.encoder,
+        encoder_proj=model.joiner.encoder_proj,
+    )
+
+    decoder = OnnxDecoder(
+        decoder=model.decoder,
+        decoder_proj=model.joiner.decoder_proj,
+    )
+
+    joiner = OnnxJoiner(output_linear=model.joiner.output_linear)
+
+    encoder_num_param = sum([p.numel() for p in encoder.parameters()])
+    decoder_num_param = sum([p.numel() for p in decoder.parameters()])
+    joiner_num_param = sum([p.numel() for p in joiner.parameters()])
+    total_num_param = encoder_num_param + decoder_num_param + joiner_num_param
+    logging.info(f"encoder parameters: {encoder_num_param}")
+    logging.info(f"decoder parameters: {decoder_num_param}")
+    logging.info(f"joiner parameters: {joiner_num_param}")
+    logging.info(f"total parameters: {total_num_param}")
+
+    if params.iter > 0:
+        suffix = f"iter-{params.iter}"
+    else:
+        suffix = f"epoch-{params.epoch}"
+
+    suffix += f"-avg-{params.avg}"
+
+    opset_version = 13
+
+    logging.info("Exporting encoder")
+    encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx"
+    export_encoder_model_onnx(
+        encoder,
+        encoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported encoder to {encoder_filename}")
+
+    logging.info("Exporting decoder")
+    decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx"
+    export_decoder_model_onnx(
+        decoder,
+        decoder_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported decoder to {decoder_filename}")
+
+    logging.info("Exporting joiner")
+    joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx"
+    export_joiner_model_onnx(
+        joiner,
+        joiner_filename,
+        opset_version=opset_version,
+    )
+    logging.info(f"Exported joiner to {joiner_filename}")
+
+
+if __name__ == "__main__":
+    formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+
+    main()

egs/librispeech/ASR/pruned_transducer_stateless5/onnx_check.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_check.py

egs/librispeech/ASR/pruned_transducer_stateless5/onnx_pretrained.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless3/onnx_pretrained.py