#!/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: ./conv_emformer_transducer_stateless2/export.py \ --exp-dir ./conv_emformer_transducer_stateless2/exp \ --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 10 \ --use-averaged-model=True \ --num-encoder-layers 12 \ --chunk-length 32 \ --cnn-module-kernel 31 \ --left-context-length 32 \ --right-context-length 8 \ --memory-size 32 \ --jit False It will generate a file exp_dir/pretrained.pt To use the generated file with `conv_emformer_transducer_stateless2/decode.py`, you can do: cd /path/to/exp_dir ln -s pretrained.pt epoch-9999.pt cd /path/to/egs/librispeech/ASR ./conv_emformer_transducer_stateless2/decode.py \ --exp-dir ./conv_emformer_transducer_stateless2/exp \ --epoch 9999 \ --avg 1 \ --max-duration 100 \ --bpe-model data/lang_bpe_500/bpe.model \ --use-averaged-model=False \ --num-encoder-layers 12 \ --chunk-length 32 \ --cnn-module-kernel 31 \ --left-context-length 32 \ --right-context-length 8 \ --memory-size 32 """ import argparse import logging from pathlib import Path import k2 import torch 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 num_tokens, 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( "--exp-dir", type=str, default="pruned_transducer_stateless2/exp", help="""It specifies the directory where all training related files, e.g., checkpoints, log, etc, are saved """, ) parser.add_argument( "--tokens", type=str, required=True, 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. """, ) 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( "--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. ", ) add_model_arguments(parser) return parser 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") 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) 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.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 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()