def add_model_args(parser):
group = parser.add_argument_group('Model configuration')
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
group.add_argument(
'--arch', '-a', default='fconv', metavar='ARCH', required=True,
choices=ARCH_MODEL_REGISTRY.keys(),
help='model architecture: {} (default: fconv)'.format(
', '.join(ARCH_MODEL_REGISTRY.keys())),
)
# Criterion definitions can be found under fairseq/criterions/
group.add_argument(
'--criterion', default='cross_entropy', metavar='CRIT',
choices=CRITERION_REGISTRY.keys(),
help='training criterion: {} (default: cross_entropy)'.format(
', '.join(CRITERION_REGISTRY.keys())),
)
return group
def add_model_args(parser):
group = parser.add_argument_group('Model configuration')
# fmt: off
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
from fairseq.models import ARCH_MODEL_REGISTRY
group.add_argument('--arch', '-a', default='fconv', metavar='ARCH', required=True,
choices=ARCH_MODEL_REGISTRY.keys(),
help='Model Architecture')
# fmt: on
group.add_argument('--init-type', default='default', type=str, choices=['adaptive', 'adaptive-profiling', \
'default', 'looklinear', 'rezero', 'rezero_postln'])
group.add_argument('--plot_variance', action='store_true')
group.add_argument('--plot_gradient', action='store_true')
group.add_argument('--plot_stability', action='store_true')
group.add_argument('--gradient_as_delta', action='store_true')
group.add_argument('--mixed_precision', action='store_true')
return group
def add_model_args(parser):
group = parser.add_argument_group("Model configuration")
# fmt: off
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
group.add_argument(
'--ngrams',
default=5,
type=int,
metavar='N',
help='read this many sentences into a buffer before processing them')
from fairseq.models import ARCH_MODEL_REGISTRY
group.add_argument('--arch',
'-a',
default='fconv',
metavar='ARCH',
choices=ARCH_MODEL_REGISTRY.keys(),
help='Model Architecture')
# fmt: on
return group
def add_model_args(parser):
group = parser.add_argument_group("Model configuration")
# fmt: off
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
from fairseq.models import ARCH_MODEL_REGISTRY
group.add_argument('--arch',
'-a',
metavar='ARCH',
choices=ARCH_MODEL_REGISTRY.keys(),
help='model architecture')
group.add_argument('--z-size',
default=64,
type=int,
help='latent hidden size for cvae')
group.add_argument('--init-w',
default=0.02,
type=float,
help='init weight for the Variation')
# fmt: on
return group
def add_model_args(parser):
group = parser.add_argument_group("Model configuration")
# fmt: off
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
from fairseq.models import ARCH_MODEL_REGISTRY
group.add_argument('--arch',
'-a',
default='fconv',
metavar='ARCH',
choices=ARCH_MODEL_REGISTRY.keys(),
help='Model Architecture')
# Some of the algorithm-specific options below
group.add_argument(
'--use-is-obj',
type=int,
default=1,
choices=[0, 1],
help='use importance sampling objective') # set to 0 for mle
group.add_argument('--load-path-mle', type=str, default=None)
group.add_argument(
'--q-baseline',
type=float,
default=-10.0,
help='subtracted baseline for q function') # per-step b in Algo 1
group.add_argument('--reward-type',
type=str,
default='logp',
choices=['sump',
'logp']) # logp is GOLD-p; sump is GOLD-s
group.add_argument('--trunc-min', type=float, default=1.0) # c in Algo 1
group.add_argument('--iw-min', type=float, default=0.0) # u in Algo 1
group.add_argument('--policy-update-per-k-epoch', type=int,
default=5000) # k in Algo 1
# Below: not used for now; could be helpful for further experiments; ablation studies / more tuning
group.add_argument('--suffix-num',
type=int,
default=5,
choices=[0, 1, 2, 3, 4, 5],
help='number of suffix tokens to consider'
) # not used now; will be useful for ablation/tuning
group.add_argument(
'--gamma', type=float, default=1.0, help='discount rate'
) # not used now; will be useful for full trajectory returns
group.add_argument('--p40', type=int, default=0,
choices=[0, 1]) # not used now
# fmt: on
return group
def add_model_args(parser):
group = parser.add_argument_group('Model configuration')
# fmt: off
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
from fairseq.models import ARCH_MODEL_REGISTRY
group.add_argument('--arch', '-a', default='fconv', metavar='ARCH', required=True,
choices=ARCH_MODEL_REGISTRY.keys(),
help='Model Architecture')
# fmt: on
return group
def add_model_args(parser):
group = parser.add_argument_group('Model configuration')
# Model definitions can be found under fairseq/models/
#
# The model architecture can be specified in several ways.
# In increasing order of priority:
# 1) model defaults (lowest priority)
# 2) --arch argument
# 3) --encoder/decoder-* arguments (highest priority)
group.add_argument(
'--arch', '-a', default='fconv', metavar='ARCH', required=True,
choices=ARCH_MODEL_REGISTRY.keys(),
help='Model Architecture',
)
# Criterion definitions can be found under fairseq/criterions/
group.add_argument(
'--criterion', default='cross_entropy', metavar='CRIT',
choices=CRITERION_REGISTRY.keys(),
help='Training Criterion',
)
return group