import os
import sys
import json
import random
import configargparse
import numpy as np
from trainer import Trainer
os.environ[“KMP_DUPLICATE_LIB_OK”] = “TRUE”
def get_parser(parser=None, required=True):
if parser is None:
parser = configargparse.ArgumentParser(
description=”Train an automatic speech recognition (ASR) model on one CPU, ”
“one or multiple GPUs”,
config_file_parser_class=configargparse.YAMLConfigFileParser,
formatter_class=configargparse.ArgumentDefaultsHelpFormatter,
)
parser.add(“–config”, is_config_file=True, help=”config file path”,default=”conf/base.yaml”)
## General utils
parser.add_argument(
“–tag”,
type=str,
help=”Experiment Tag for storing logs,models”
)
parser.add_argument(
“–seed”,
default=1,
type=int,
help=”Random seed”
)
parser.add_argument(
“–text_pad”,
default=-1,
type=int,
help=”Padding Index for Text Labels”
)
parser.add_argument(
“–audio_pad”,
default=0,
type=int,
help=”Padding Index for Audio features”
)
## I/O parameters
parser.add_argument(
“–train-json”,
type=str,
default=”asr_data/train_si284/data.json”,
help=”Filename of train label data (json)”,
)
parser.add_argument(
“–valid-json”,
type=str,
default=”asr_data/test_dev93/data.json”,
help=”Filename of validation label data (json)”,
)
parser.add_argument(
“–dict”,
type=str,
default=”asr_data/train_si284_units.txt”,
help=”Filename of the dictionary file”,
)
parser.add_argument(
“–out_dir”,
type=str,
default=”exp/base”,
help=”Output Data Directory/Experiment Directory”
)
## Model Parameters
parser.add_argument(
“–idim”,
type=int,
default=83,
help=”Input Feature Size”
)
## Encoder
parser.add_argument(
“–ehiddens”,
type=int,
default=320,
help=”Encoder Hidden Size”
)
parser.add_argument(
“–eprojs”,
type=int,
default=320,
help=”Encoder Projection Output Size”
)
parser.add_argument(
“–etype”,
type=str,
default=”blstm”,
help=”Encoder Type”
)
parser.add_argument(
“–elayers”,
type=int,
default=3,
help=”Number of encoder layers”
)
parser.add_argument(
“–pfactor”,
type=int,
default=2,
help=”Pyramidal Downsampling factor”
)
## Decoder
parser.add_argument(
“–dtype”,
type=str,
default=”lstm”,
help=”Decoder Type”
)
parser.add_argument(
“–dhiddens”,
type=int,
default=300,
help=”Decoder Hidden Size”
)
parser.add_argument(
“–demb_dim”,
type=int,
default=100,
help=”Decoder Embedding Size”
)
parser.add_argument(
“–dlayers”,
type=int,
default=2,
help=”Number of decoder layers”
)
parser.add_argument(
“–ddropout”,
type=float,
default=0.0,
help=”Decoder Dropout Ratio”
)
parser.add_argument(
“–ssprob”,
type=float,
default=0.0,
help=”Probability of Scheduled Sampling”
)
## Attention
parser.add_argument(
“–att_dim”,
type=int,
default=320,
help=”Attention Projection Dimension for Location Based Attention”
)
parser.add_argument(
“–aconv_chans”,
type=int,
default=10,
help=”Attention Convolution Channels for Location Based Attention”
)
parser.add_argument(
“–aconv_filts”,
type=int,
default=100,
help=”Attention Convolution Filter Size for Location Based Attention”
)
## Batching
parser.add_argument(
“–batch_bins”,
type=int,
default=800000
)
parser.add_argument(
“–nworkers”,
dest=”nworkers”,
type=int,
default=0
)
## Optimization Parameters
parser.add_argument(
“–opt”,
default=”adadelta”,
type=str,
choices=[“adadelta”, “adam”],
help=”Optimizer”,
)
parser.add_argument(
“–lr”,
type=float,
default=1e-3
)
parser.add_argument(
“–lr_decay”,
type=float,
default=0.5
)
parser.add_argument(
“–opt_patience”,
type=int,
default=5
)
parser.add_argument(
“–grad_clip”,
type=float,
default=5
)
parser.add_argument(
“–wdecay”,
type=float,
default=0,
help=”Weight Decay Parameter”
)
parser.add_argument(
“–accum-grad”,
default=1,
type=int,
help=”Number of gradient accumuration”
)
parser.add_argument(
“–eps”,
default=1e-8,
type=float,
help=”Epsilon constant for optimizer”
)
parser.add_argument(
“–eps-decay”,
default=0.01,
type=float,
help=”Decaying ratio of epsilon”
)
parser.add_argument(
“–patience”,
default=5,
type=int,
nargs=”?”,
help=”Number of epochs to wait without improvement before stopping the training”,
)
## Model Initialization
parser.add_argument(
“–resume”,
type=str,
default=””,
help=”Path of model and optimizer to be loaded”
)
## Training Config
parser.add_argument(
“–nepochs”,
type=int,
default=30,
help=”Number of GPUs used for training”
)
parser.add_argument(
“–ngpu”,
default=1,
type=int,
help=”Number of GPUs. If not given, use all visible devices”,
)
parser.add_argument(
“–num-batch-save-attention”,
default=1,
type=int,
help=”Number of samples of attention to be saved”,
)
return parser
def main(cmd_args):
## Return the arguments from parser
parser = get_parser()
args, _ = parser.parse_known_args(cmd_args)
## Set Random Seed for Reproducibility
random.seed(args.seed)
np.random.seed(args.seed)
## Set all directories
expdir = os.path.join(“exp”,”train_”+args.tag)
graph_dir = os.path.join(expdir,”plots”)
model_dir = os.path.join(expdir,”models”)
log_dir = os.path.join(expdir,”logs”)
tb_dir = os.path.join(“tensorboard”,”train_”+args.tag)
args.graph_dir = graph_dir
args.model_dir = model_dir
args.log_dir = log_dir
args.tb_dir = tb_dir
args.expdir = expdir
for x in [“exp”,”tensorboard”,expdir,graph_dir,model_dir,log_dir,tb_dir]:
if not os.path.isdir(x):
os.mkdir(x)
## Load the Character List, add
## We use the same token to represent
char_list = [“
with open(args.dict,”r”,encoding=”utf-8″) as f:
char_dict = [line.strip().split(‘ ‘)[0] for line in f.readlines()]
char_list = char_list + char_dict + [“
args.char_list = char_list
args.odim = len(char_list)
## Start training
trainer = Trainer(args)
trainer.train()
if __name__ == “__main__”:
main(sys.argv[1:])