diff --git a/.github/scripts/test-onnx-export.sh b/.github/scripts/test-onnx-export.sh index 1f5fbdc71..327e31942 100755 --- a/.github/scripts/test-onnx-export.sh +++ b/.github/scripts/test-onnx-export.sh @@ -18,7 +18,6 @@ GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url repo=$(basename $repo_url) pushd $repo -git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/pretrained.pt" cd exp ln -s pretrained.pt epoch-99.pt @@ -84,7 +83,6 @@ GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url repo=$(basename $repo_url) pushd $repo -git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/pretrained.pt" cd exp @@ -322,7 +320,6 @@ GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url repo=$(basename $repo_url) pushd $repo -git lfs pull --include "data/lang_bpe_500/tokens.txt" git lfs pull --include "exp/pretrained.pt" cd exp diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/export2.py b/egs/aishell/ASR/pruned_transducer_stateless7/export2.py deleted file mode 100755 index 14c826521..000000000 --- a/egs/aishell/ASR/pruned_transducer_stateless7/export2.py +++ /dev/null @@ -1,322 +0,0 @@ -#!/usr/bin/env python3 -# -# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang -# Zengrui Jin) -# -# 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: - -(1) Export to torchscript model using torch.jit.script() - -./pruned_transducer_stateless7/export.py \ - --exp-dir ./pruned_transducer_stateless7/exp \ - --tokens data/lang_bpe_500/tokens.txt \ - --epoch 30 \ - --avg 9 \ - --jit 1 - -It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later -load it by `torch.jit.load("cpu_jit.pt")`. - -Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python -are on CPU. You can use `to("cuda")` to move them to a CUDA device. - -Check -https://github.com/k2-fsa/sherpa -for how to use the exported models outside of icefall. - -(2) Export `model.state_dict()` - -./pruned_transducer_stateless7/export.py \ - --exp-dir ./pruned_transducer_stateless7/exp \ - --tokens data/lang_char/tokens.txt \ - --epoch 20 \ - --avg 10 - -It will generate a file `pretrained.pt` in the given `exp_dir`. You can later -load it by `icefall.checkpoint.load_checkpoint()`. - -To use the generated file with `pruned_transducer_stateless7/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_stateless7/decode.py \ - --exp-dir ./pruned_transducer_stateless7/exp \ - --epoch 9999 \ - --avg 1 \ - --max-duration 600 \ - --decoding-method greedy_search \ - --tokens data/lang_char/tokens.txt - -Check ./pretrained.py for its usage. - -Note: If you don't want to train a model from scratch, we have -provided one for you. You can get it at - -https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11 - -with the following commands: - - sudo apt-get install git-lfs - git lfs install - git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11 - # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp -""" - -import argparse -import logging -from pathlib import Path -import re - -import k2 -import torch -import torch.nn as nn -from scaling_converter import convert_scaled_to_non_scaled -from train2 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 str2bool, num_tokens - - -def get_parser(): - parser = argparse.ArgumentParser( - formatter_class=argparse.ArgumentDefaultsHelpFormatter - ) - - parser.add_argument( - "--epoch", - type=int, - default=30, - help="""It specifies the checkpoint to use for decoding. - Note: Epoch counts from 1. - 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=9, - 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_stateless7/exp", - help="""It specifies the directory where all training related - files, e.g., checkpoints, log, etc, are saved - """, - ) - - parser.add_argument( - "--tokens", - type=str, - default="data/lang_bpe_500/tokens.txt", - help="Path to the tokens.txt.", - ) - - parser.add_argument( - "--jit", - type=str2bool, - default=False, - help="""True to save a model after applying torch.jit.script. - It will generate a file named cpu_jit.pt - - Check ./jit_pretrained.py for how to use it. - """, - ) - - 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 - - -@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) - - logging.info(f"device: {device}") - - # Load tokens.txt here - token_table = k2.SymbolTable.from_file(params.tokens) - - # Load id of the token and the vocab size - params.blank_id = token_table[""] - params.vocab_size = num_tokens(token_table) + 1 # +1 for - - 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() - - if params.jit: - convert_scaled_to_non_scaled(model, inplace=True) - # We won't use the forward() method of the model in C++, so just ignore - # it here. - # Otherwise, one of its arguments is a ragged tensor and is not - # torch scriptabe. - model.__class__.forward = torch.jit.ignore(model.__class__.forward) - 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 torchscript. Export model.state_dict()") - # 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() diff --git a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py index 8d4e003d4..f5a04e7f5 100755 --- a/egs/librispeech/ASR/zipformer/export-onnx-streaming.py +++ b/egs/librispeech/ASR/zipformer/export-onnx-streaming.py @@ -756,7 +756,7 @@ def main(): quantize_dynamic( model_input=decoder_filename, model_output=decoder_filename_int8, - op_types_to_quantize=["MatMul"], + op_types_to_quantize=["MatMul", "Gather"], weight_type=QuantType.QInt8, ) diff --git a/egs/librispeech/ASR/zipformer/export-onnx.py b/egs/librispeech/ASR/zipformer/export-onnx.py index cb7f8f0df..adc50453e 100755 --- a/egs/librispeech/ASR/zipformer/export-onnx.py +++ b/egs/librispeech/ASR/zipformer/export-onnx.py @@ -601,7 +601,7 @@ def main(): quantize_dynamic( model_input=decoder_filename, model_output=decoder_filename_int8, - op_types_to_quantize=["MatMul"], + op_types_to_quantize=["MatMul", "Gather"], weight_type=QuantType.QInt8, )