def add_args(parser, args):
group = parser.add_argument_group("RNN encoder")
parser = ConvEncoder.add_args(parser, args)
group.add_argument('--enc_n_units',
type=int,
default=512,
help='number of units in each encoder RNN layer')
group.add_argument(
'--enc_n_projs',
type=int,
default=0,
help=
'number of units in the projection layer after each encoder RNN layer'
)
group.add_argument(
'--bidirectional_sum_fwd_bwd',
type=strtobool,
default=False,
help='sum forward and backward RNN outputs for dimension reduction'
)
# streaming
group.add_argument(
'--lc_chunk_size_left',
type=str,
default="0",
help='left chunk size for latency-controlled RNN encoder')
group.add_argument(
'--lc_chunk_size_right',
type=str,
default="0",
help='right chunk size for latency-controlled RNN encoder')
return parser
def add_args(parser, args):
"""Add arguments."""
group = parser.add_argument_group("Transformer encoder")
if 'conv' in args.enc_type:
parser = ConvEncoder.add_args(parser, args)
# Transformer common
if not hasattr(args, 'transformer_layer_norm_eps'):
group.add_argument('--transformer_ffn_bottleneck_dim', type=int, default=0,
help='bottleneck dimension in the FFN layer')
group.add_argument('--transformer_input_bottleneck_dim', type=int, default=0,
help='bottleneck dimension in the FFN layer')
group.add_argument('--transformer_layer_norm_eps', type=float, default=1e-12,
help='epsilon value for layer normalization')
group.add_argument('--transformer_ffn_activation', type=str, default='relu',
choices=['relu', 'gelu', 'gelu_accurate', 'glu', 'swish'],
help='nonlinear activation for the FFN layer')
group.add_argument('--transformer_param_init', type=str, default='xavier_uniform',
choices=['xavier_uniform', 'pytorch'],
help='parameter initialization')
# Transformer encoder specific
group.add_argument('--transformer_enc_d_model', type=int, default=256,
help='number of units in the MHA layer for Transformer encoder')
group.add_argument('--transformer_enc_d_ff', type=int, default=2048,
help='number of units in the FFN layer for Transformer encoder')
group.add_argument('--transformer_enc_n_heads', type=int, default=4,
help='number of heads in the MHA layer for Transformer encoder')
group.add_argument('--transformer_enc_pe_type', type=str, default='add',
choices=['add', 'none', 'relative', 'relative_xl'],
help='type of positional encoding for Transformer encoder')
group.add_argument('--dropout_enc_layer', type=float, default=0.0,
help='LayerDrop probability for Transformer encoder layers')
group.add_argument('--transformer_enc_clamp_len', type=int, default=-1,
help='maximum length for relative positional encoding. -1 means infinite length.')
# streaming
group.add_argument('--transformer_enc_lookaheads', type=str, default="0_0_0_0_0_0_0_0_0_0_0_0",
help='lookahead frames per layer for unidirectional Transformer encoder')
group.add_argument('--lc_chunk_size_left', type=str, default="0",
help='left chunk size for latency-controlled Transformer encoder')
group.add_argument('--lc_chunk_size_current', type=str, default="0",
help='current chunk size (and hop size) for latency-controlled Transformer encoder')
group.add_argument('--lc_chunk_size_right', type=str, default="0",
help='right chunk size for latency-controlled Transformer encoder')
group.add_argument('--lc_type', type=str, default='reshape',
choices=['reshape', 'mask'],
help='implementation methods of latency-controlled Transformer encoder')
return parser
def add_args(parser, args):
"""Add arguments."""
group = parser.add_argument_group("Transformer encoder")
if 'conv' in args.enc_type:
parser = ConvEncoder.add_args(parser, args)
# Transformer common
if not hasattr(args, 'transformer_d_model'):
group.add_argument('--transformer_d_model', type=int, default=256,
help='number of units in the MHA layer')
if not hasattr(args, 'transformer_d_ff'):
group.add_argument('--transformer_d_ff', type=int, default=2048,
help='number of units in the FFN layer')
if not hasattr(args, 'transformer_d_ff_bottleneck_dim'):
group.add_argument('--transformer_d_ff_bottleneck_dim', type=int, default=0,
help='bottleneck dimension in the FFN layer')
if not hasattr(args, 'transformer_n_heads'):
group.add_argument('--transformer_n_heads', type=int, default=4,
help='number of heads in the MHA layer')
if not hasattr(args, 'transformer_layer_norm_eps'):
group.add_argument('--transformer_layer_norm_eps', type=float, default=1e-12,
help='epsilon value for layer normalization')
if not hasattr(args, 'transformer_ffn_activation'):
group.add_argument('--transformer_ffn_activation', type=str, default='relu',
choices=['relu', 'gelu', 'gelu_accurate', 'glu', 'swish'],
help='nonlinear activation for the FFN layer')
if not hasattr(args, 'transformer_param_init'):
group.add_argument('--transformer_param_init', type=str, default='xavier_uniform',
choices=['xavier_uniform', 'pytorch'],
help='parameter initializatin')
# NOTE: These checks are important to avoid conflict with args in Transformer decoder
# Conformer encoder specific
group.add_argument('--transformer_enc_pe_type', type=str, default='relative',
choices=['relative'],
help='type of positional encoding for the Transformer encoder')
group.add_argument('--conformer_kernel_size', type=int, default=32,
help='kernel size for depthwise convolution in convolution module for Conformer encoder layers')
group.add_argument('--dropout_enc_layer', type=float, default=0.0,
help='LayerDrop probability for Conformer encoder layers')
# streaming
group.add_argument('--lc_chunk_size_left', type=int, default=0,
help='left chunk size for latency-controlled Conformer encoder')
group.add_argument('--lc_chunk_size_current', type=int, default=0,
help='current chunk size (and hop size) for latency-controlled Conformer encoder')
group.add_argument('--lc_chunk_size_right', type=int, default=0,
help='right chunk size for latency-controlled Conformer encoder')
return parser
def add_args(parser, args):
group = parser.add_argument_group("RNN encoder")
parser = ConvEncoder.add_args(parser, args)
group.add_argument('--enc_n_units',
type=int,
default=512,
help='number of units in each encoder RNN layer')
group.add_argument(
'--enc_n_projs',
type=int,
default=0,
help=
'number of units in the projection layer after each encoder RNN layer'
)
group.add_argument(
'--bidirectional_sum_fwd_bwd',
type=strtobool,
default=False,
help='sum forward and backward RNN outputs for dimension reduction'
)
# streaming
group.add_argument(
'--lc_chunk_size_left',
type=str,
default="-1",
help='current chunk size for latency-controlled RNN encoder')
group.add_argument(
'--lc_chunk_size_right',
type=str,
default="0",
help='right chunk size for latency-controlled RNN encoder')
group.add_argument('--cnn_lookahead',
type=strtobool,
default=True,
help='disable lookahead frames in CNN layers')
group.add_argument('--rsp_prob_enc',
type=float,
default=0.0,
help='probability for Random State Passing (RSP)')
return parser
def add_args(parser, args):
# group = parser.add_argument_group("TDS encoder")
parser = ConvEncoder.add_args(parser, args)
return parser