def add_cmdline_args(cls,
parser: ParlaiParser,
partial_opt: Optional[Opt] = None) -> ParlaiParser:
"""
Override to add one arg.
"""
TransformerGeneratorAgent.add_cmdline_args(parser,
partial_opt=partial_opt)
TorchImageAgent.add_cmdline_args(parser, partial_opt=partial_opt)
group = parser.add_argument_group('Image Encoder Args')
group.add_argument(
'--include-image-token',
type='bool',
default=True,
recommended=True,
help=
'if true, include image token (or no image token) for each example',
)
group.add_argument(
'--image-fusion-type',
type=str,
default='late',
choices=[f.value for f in FusionType],
help='which fusion type to use',
)
return group
def add_cmdline_args(cls, argparser):
"""
Add command-line arguments specifically for this agent.
"""
PolyencoderAgent.add_cmdline_args(argparser)
TorchImageAgent.add_cmdline_args(argparser)
agent = argparser.add_argument_group('ImagePolyencoder Args')
agent.add_argument(
'--image-combination-mode',
type=str,
default='prepend',
choices=['add', 'append', 'prepend'],
help='How to combine image embedding (if used) with context embedding',
)
# TODO: more thoroughly test out whether one of these choices is best and add a
# 'recommended' arg here. 'add' and 'prepend' seem to be roughly similar in
# performance
agent.add_argument(
'--n-image-tokens',
type=int,
default=1,
help=(
'Number of tokens that the image encoding will consist of (when adding '
'or prepending)'
),
)
agent.set_defaults(reduction_type=None)
# This agent doesn't support any encoder output reductions
return agent
def add_cmdline_args(cls, argparser):
"""
Override to add one arg.
"""
TransformerGeneratorAgent.add_cmdline_args(argparser)
TorchImageAgent.add_cmdline_args(argparser)
group = argparser.add_argument_group('Image Encoder Args')
group.add_argument(
'--include-image-token',
type='bool',
default=True,
recommended=True,
help='if true, include image token (or no image token) for each example',
)
def _process_image_features(self, features: torch.Tensor) -> torch.Tensor:
"""
Format shape and type of input image-feature tensor.
Override TorchImageAgent._process_image_features to handle multi-dimensional
images.
"""
features = features.view(-1, self.image_features_dim)
return torch.stack([
TorchImageAgent._process_image_features(self, features[i])
for i in range(features.size(0))
])
def add_cmdline_args(cls, argparser):
"""
Override to add one arg.
"""
TransformerGeneratorAgent.add_cmdline_args(argparser)
TorchImageAgent.add_cmdline_args(argparser)
group = argparser.add_argument_group('Image Encoder Args')
group.add_argument(
'--include-image-token',
type='bool',
default=True,
recommended=True,
help=
'if true, include image token (or no image token) for each example',
)
group.add_argument(
'--image-fusion-type',
type=str,
default='late',
choices=[f.value for f in FusionType],
help='which fusion type to use',
)
def add_cmdline_args(cls,
parser: ParlaiParser,
partial_opt: Optional[Opt] = None) -> ParlaiParser:
"""
Add command-line arguments specifically for this agent.
"""
PolyencoderAgent.add_cmdline_args(parser, partial_opt=partial_opt)
TorchImageAgent.add_cmdline_args(parser, partial_opt=partial_opt)
agent = parser.add_argument_group('ImagePolyencoder Args')
agent.add_argument(
'--image-combination-mode',
type=str,
default='prepend',
choices=['add', 'append', 'prepend'],
help=
'How to combine image embedding (if used) with context embedding',
)
# TODO: more thoroughly test out whether one of these choices is best and add a
# 'recommended' arg here. 'add' and 'prepend' seem to be roughly similar in
# performance
agent.set_defaults(reduction_type=None)
# This agent doesn't support any encoder output reductions
return agent