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.
"""
DistillNarrowTransformerAgentMixin.add_cmdline_args(argparser)
TransformerGeneratorAgent.add_cmdline_args(argparser)
return argparser
def add_cmdline_args(argparser: ParlaiParser):
"""
Add command-line arguments specifically for this agent.
"""
TransformerGeneratorAgent.add_cmdline_args(argparser)
parser = argparser.add_argument_group(
'RetNRef Transformer Generator Agent Args')
parser.add_argument(
'--use-knowledge',
type='bool',
default=False,
help='Use knowledge to generate response.',
)
parser.add_argument(
'--add-knowledge-to-history',
type='bool',
default=False,
help='using knowledge as temp history if True.',
)
parser.add_argument(
'--chosen-sentence',
type='bool',
default=True,
help='instead of using all knowledge, use gold'
'label, i.e. the chosen sentence',
)
def add_cmdline_args(
cls, parser: ParlaiParser, partial_opt: Optional[Opt] = None
) -> ParlaiParser:
"""
Add command-line arguments specifically for this agent.
"""
DistillTransformerAgentMixin.add_cmdline_args(parser, partial_opt=partial_opt)
TransformerGeneratorAgent.add_cmdline_args(parser, partial_opt=partial_opt)
return parser
def add_cmdline_args(argparser: ParlaiParser):
"""
Add command-line arguments specifically for this agent.
"""
TransformerGeneratorAgent.add_cmdline_args(argparser)
argparser.add_argument(
'--nidf',
type='int',
default=4,
help=
'nidf of weighted decoding , see https://arxiv.org/abs/1902.08654')
def add_cmdline_args(cls,
parser: ParlaiParser,
partial_opt: Optional[Opt] = None) -> ParlaiParser:
"""
Add command-line arguments specifically for this agent.
"""
TransformerGeneratorAgent.add_cmdline_args(parser,
partial_opt=partial_opt)
AbstractGeneratorRerankAgentMixin.add_cmdline_args(parser, partial_opt)
reranker_class = cls.get_reranker_class() or AbstractReranker
reranker_class.add_cmdline_args(parser, partial_opt=partial_opt)
return parser
def add_cmdline_args(argparser: ParlaiParser):
"""
Override to add init-fairseq-model arg.
"""
TransformerGeneratorAgent.add_cmdline_args(argparser)
group = argparser.add_argument_group('Bart Args')
group.add_argument(
'--init-fairseq-model',
type=str,
default=None,
help='fairseq checkpoint for bart',
)
argparser.set_defaults(dict_tokenizer='gpt2')
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 add_cmdline_args(cls,
parser: ParlaiParser,
partial_opt: Optional[Opt] = None) -> ParlaiParser:
TransformerGeneratorAgent.add_cmdline_args(
parser, partial_opt=partial_opt) # add transformer args
parser.add_argument(
'--n-positions-init',
type=int,
default=None,
hidden=True,
help='Number of positional embeddings from the init model. Defaults '
'to truncate or 1024 if not provided.',
)
return parser
def add_cmdline_args(argparser: ParlaiParser):
"""
Add command-line arguments specifically for this agent.
"""
TransformerGeneratorAgent.add_cmdline_args(argparser)
argparser.add_argument(
'--response-num',
type=int,
default=1,
help='response number of when act',
)
argparser.add_argument(
'--response-sep',
type=str,
default='__ressep__',
help='response resperator',
)
def add_cmdline_args(
cls, parser: ParlaiParser, partial_opt: Optional[Opt] = None
) -> ParlaiParser:
TransformerGeneratorAgent.add_cmdline_args(parser, partial_opt=partial_opt)
# Add transformer args.
parser.add_argument(
'--ladder-size',
type=int,
default=1,
help='Number of ladder steps, default is not to use ladder.',
)
parser.add_argument(
'--hash-size', type=int, default=32, help='Number of hash bins.'
)
parser.add_argument(
'--hash-layer',
type=int,
default=7,
help='Layer number the Hash Layer appears on.',
)
return parser
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 run_conversation(mturk_manager, opt, workers):
mturk_agent = workers[0]
model_agent = TransformerGeneratorAgent(opt=opt)
world = ModelEvaluatorWorld(
opt=opt,
model_agent=model_agent,
task_opt=task_opt,
mturk_agent=mturk_agent,
)
while not world.episode_done():
world.parley()
world.shutdown()
world.review_work()
def main():
argparser = ParlaiParser(False, False)
argparser.add_parlai_data_path()
argparser.add_mturk_args()
# The dialog model we want to evaluate
from parlai.agents.transformer.transformer import TransformerGeneratorAgent
from parlai.agents.ir_baseline.ir_baseline import IrBaselineAgent
TransformerGeneratorAgent.add_cmdline_args(argparser)
opt = argparser.parse_args()
opt['task'] = os.path.basename(os.path.dirname(os.path.abspath(__file__)))
opt.update(task_config)
# The task that we will evaluate the dialog model on
task_opt = {}
task_opt['datatype'] = 'test'
task_opt['datapath'] = opt['datapath']
task_opt['task'] = 'dailydialog:NoStart'
opt['dict_file'] = '/home/christian/developer/cs767hw4/models/test.dict'
opt['model_file'] = '/home/christian/developer/cs767hw4/models/test.checkpoint'
opt['truncate'] = 64
opt['inference'] = 'mmi'
opt['beam_size'] = 32
opt['no_cuda'] = True
mturk_agent_id = 'Worker'
mturk_manager = MTurkManager(opt=opt, mturk_agent_ids=[mturk_agent_id])
mturk_manager.setup_server() # task_name="dialo-mturk-eval", exists=True
try:
mturk_manager.start_new_run()
mturk_manager.create_hits()
def run_onboard(worker):
world = ModelEvaluatorOnboardWorld(opt=opt, mturk_agent=worker)
while not world.episode_done():
world.parley()
world.shutdown()
mturk_manager.set_onboard_function(onboard_function=run_onboard)
mturk_manager.ready_to_accept_workers()
def check_worker_eligibility(worker):
return True
def assign_worker_roles(worker):
worker[0].id = mturk_agent_id
global run_conversation
def run_conversation(mturk_manager, opt, workers):
mturk_agent = workers[0]
model_agent = TransformerGeneratorAgent(opt=opt)
world = ModelEvaluatorWorld(
opt=opt,
model_agent=model_agent,
task_opt=task_opt,
mturk_agent=mturk_agent,
)
while not world.episode_done():
world.parley()
world.shutdown()
world.review_work()
mturk_manager.start_task(
eligibility_function=check_worker_eligibility,
assign_role_function=assign_worker_roles,
task_function=run_conversation,
)
except BaseException:
raise
finally:
mturk_manager.expire_all_unassigned_hits()
mturk_manager.shutdown()
def add_cmdline_args(
cls, parser: ParlaiParser, partial_opt: Optional[Opt] = None
) -> ParlaiParser:
PolyencoderAgent.add_cmdline_args(parser, partial_opt)
TransformerGeneratorAgent.add_cmdline_args(parser, partial_opt)
MultiObjectiveAgentBase.add_cmdline_args(parser, partial_opt)
agent = parser.add_argument_group('Multi-Objective Generator Arguments')
agent.add_argument(
'--character-key',
type=str,
default='full_text',
help='key from which to extract characters',
)
agent.add_argument(
'--character-candidates-key',
type=str,
default='full_text',
help='key from which to extract character candidates (i.e., candidates for ranking)',
)
agent.add_argument(
'--multiobjective-latent-representation',
type=str,
default='decoder_final_layer',
choices=[
'encoder_final_layer',
'decoder_final_layer',
'encoder_and_decoder',
],
help='Which latent representation from the base model to pass to the extra layers',
)
agent.add_argument(
'--multiobjective-loss-ratio',
type=float,
default=0.5,
help='ratio of multiobjective loss to normal loss',
)
agent.add_argument(
'--multiobjective-loss',
type=str,
default='full',
help='Type of Multi-Objective loss to apply to this model, while training: '
'full ==> use the full teacher-forced output distributions as input to the classifier\n'
'partial ==> each partial output is classified as its own example. weights of the classifier '
'are kept constant\n'
'sliced ==> we arbitrarily slice the full output to compute partial-output losses\n'
'Can specify multiple via comma-separation.',
)
agent.add_argument(
'--multiobjective-full-loss-backprop',
type=str,
default='all',
choices=['all', 'extra_layers', 'base_model'],
help='Which parameters to train with the multiobjective full output loss',
)
agent.add_argument(
'--multiobjective-sliced-loss-backprop',
type=str,
default='all',
choices=['all', 'extra_layers', 'base_model'],
help='Which parameters to train with the multiobjective sliced output loss',
)
agent.add_argument(
'--multiobjective-partial-loss-backprop',
type=str,
default='base_model',
choices=['all', 'extra_layers', 'base_model'],
help='Which parameters to train with the multiobjective partial output loss',
)
agent.add_argument(
'--partial-output-loss-threshold',
type=float,
default=-1,
help='threshold of full loss at which point partial output loss can be computed'
'e.g., if set to 10, we dont compute partial loss until full loss reaches 10 or lower',
)
agent.add_argument(
'--sliced-output-loss-threshold',
type=float,
default=-1,
help='threshold of full loss at which point sliced output loss can be computed'
'e.g., if set to 10, we dont compute sliced loss until full loss reaches 10 or lower',
)
return parser
def add_cmdline_args(
cls, parser: ParlaiParser, partial_opt: Optional[Opt] = None
) -> ParlaiParser:
TransformerGeneratorAgent.add_cmdline_args(
parser, partial_opt=partial_opt
) # add transformer args
expanded_attn = parser.add_argument_group('Expanded Attention Arguments')
expanded_attn.add_argument(
'--expanded-attention-init-weights',
type=str,
default='random',
choices=['random', 'encoder_attention'],
help='how to initialize the expanded attention module.\n'
'random: initialize weights randomly\n'
'encoder_attention: initialize weights from encoder_attention weights\n',
)
expanded_attn.add_argument(
'--expanded-attention-share-weights',
type='bool',
default=True,
help='If true, share the weights from the normal encoder attention '
'with the extra attention (reduce network parameters).',
)
expanded_attn.add_argument(
'--expanded-attention-input-key',
type=str,
default='full_text',
help='Key in observation from which to extract the expanded attn input',
)
expanded_attn.add_argument(
'--expanded-attention-self-character-key',
type=str,
default='self_character',
help='Key in observation with model\'s character',
)
expanded_attn.add_argument(
'--expanded-attention-input-extractor-phrases',
type=str,
default='_self_name,_self_persona,_partner_name',
help='How to extract expanded attention input from the observation\n'
'Default are phrases from LIGHT',
)
expanded_attn.add_argument(
'--expanded-attention-type',
type=str,
choices=['profile', 'automated_classifier', 'automated_trainable_mask'],
default='profile',
help='Method of incorporating expanded attention. `profile` manually extracts from '
'provided context in `--expanded-attention-input-key`, while `automated` uses a separate '
'classifier to choose the tokens to which we re-attend.',
)
expanded_attn.add_argument(
'--expanded-attention-num-rounds',
type=int,
default=1,
help='how many rounds to re-apply expanded attention',
)
### Classifier Attention
expanded_attn.add_argument(
'--automated-expanded-attention-n-tokens',
type=int,
default=125,
help='If using automated attention, how many tokens do we re-attend to.',
)
expanded_attn.add_argument(
'--expanded-attention-classifier-model-file',
type=str,
default=None,
help='specified classifier model if using classifier attention to choose tokens',
)
expanded_attn.add_argument(
'--expanded-attention-classifier-truncate',
type=int,
default=510,
help='Truncation for the classifier.',
)
return parser
def main():
argparser = ParlaiParser(False, True)
argparser.add_parlai_data_path()
argparser.add_mturk_args()
# The dialog model we want to evaluate
from parlai.agents.transformer.transformer import TransformerGeneratorAgent as Agent
Agent.add_cmdline_args(argparser)
opt = argparser.parse_args()
opt['task'] = os.path.basename(os.path.dirname(os.path.abspath(__file__)))
opt.update(task_config)
opt['interactive_mode'] = True
opt['interactive_task'] = True
# The task that we will evaluate the dialog model on
task_opt = {}
task_opt['datatype'] = 'test'
task_opt['datapath'] = opt['datapath']
task_opt['task'] = 'interactive'
mturk_agent_id = 'Worker'
mturk_manager = MTurkManager(opt=opt, mturk_agent_ids=[mturk_agent_id])
mturk_manager.setup_server()
try:
mturk_manager.start_new_run()
mturk_manager.ready_to_accept_workers()
mturk_manager.create_hits()
def run_onboard(worker):
world = ModelEvaluatorOnboardWorld(opt=opt, mturk_agent=worker)
while not world.episode_done():
world.parley()
world.shutdown()
mturk_manager.set_onboard_function(onboard_function=run_onboard)
def check_worker_eligibility(worker):
return True
def assign_worker_roles(worker):
worker[0].id = mturk_agent_id
global run_conversation
def run_conversation(mturk_manager, opt, workers):
mturk_agent = workers[0]
model_agent = create_agent(opt, requireModelExists=True)
#model_agent = Agent(opt=opt)
world = ModelEvaluatorWorld(
opt=opt,
model_agent=model_agent,
task_opt=task_opt,
mturk_agent=mturk_agent,
)
while not world.episode_done():
world.parley()
world.shutdown()
world.review_work()
mturk_manager.start_task(
eligibility_function=check_worker_eligibility,
assign_role_function=assign_worker_roles,
task_function=run_conversation,
)
except BaseException:
raise
finally:
mturk_manager.expire_all_unassigned_hits()
mturk_manager.shutdown()
def add_cmdline_args(
cls, parser: ParlaiParser, partial_opt: Optional[Opt] = None
) -> ParlaiParser:
TransformerGeneratorAgent.add_cmdline_args(parser, partial_opt=partial_opt)
PacerAgentMixin.add_cmdline_args(parser, partial_opt=partial_opt)
return parser
