def cl_build_ref_agent(self):
ref_model_file = self.opt['ref_model_file']
if ref_model_file is None or ref_model_file.lower() == "none":
raise RuntimeError("CL training requires reference model!")
else:
from parlai.core.agents import create_agent_from_opt_file
ref_agent = create_agent_from_opt_file(
Opt({'model_file': ref_model_file}))
eval_ref_agent = create_agent_from_opt_file(
Opt({'model_file': ref_model_file}))
if ref_agent is None:
raise RuntimeError(
"Build reference model failed! check your `ref_model_file`:{}!"
.format(ref_model_file))
if self.id == ref_agent.id and dict_same(self, ref_agent):
self.use_external_ref_model = False
else:
self.use_external_ref_model = True
# No need to do this
# # check dict
# if self.dict.tok2ind != ref_agent.dict.tok2ind or self.dict.ind2tok != ref_agent.dict.ind2tok:
# raise RuntimeError("Reference model is using different dict!")
self.eval_ref_agent = eval_ref_agent
self.ref_agent = ref_agent
def __init__(self, model: str, device: str, maxlen: int) -> None:
model = self.check_agent(model)
maxlen = maxlen if maxlen > 0 else self.default_maxlen()
if "xxlarge" in model:
size = "9B"
elif "xlarge" in model:
size = "3B"
elif "large" in model:
size = "1Bdistill"
elif "medium" in model:
size = "400Mdistill"
elif "small" in model:
size = "90M"
else:
raise Exception("wrong model")
option = self.set_options(
name=f"zoo:blender/blender_{size}/model",
device=device,
)
super().__init__(
name=model,
suffix="\n",
device=device,
maxlen=maxlen,
model=create_agent_from_opt_file(option),
)
def test_init_from_from_checkpoint(self):
with testing_utils.tempdir() as temp_dir:
opt_from_file = {
'datapath': 'dummy_path',
'model': 'repeat_label',
'init_model': os.path.join(temp_dir, 'something'),
'model_file': os.path.join(temp_dir, 'something_else'),
}
opt = Opt({
'datapath':
'dummy_path',
'model':
'repeat_label',
'init_model':
os.path.join(temp_dir, 'something_else.checkpoint'),
'model_file':
os.path.join(temp_dir, 'something_else'),
'load_from_checkpoint':
True,
})
with open(os.path.join(temp_dir, 'something_else.opt'), 'w') as f:
f.write(json.dumps(opt_from_file))
agent = create_agent_from_opt_file(opt)
init_model = agent.opt['init_model']
# assert that the model was loaded with the correct checkpoitn
assert '.checkpoint' in init_model
def __init__(self, model, device, maxlen):
option = self.set_options(
name="zoo:sensitive_topics_classifier/model",
device=device,
)
super(SensitiveAgent, self).__init__(
device=device,
maxlen=maxlen,
model=create_agent_from_opt_file(option),
suffix="",
name=model,
)
def __init__(self, model, device, maxlen):
model = self.check_agent(model)
maxlen = maxlen if maxlen > 0 else self.default_maxlen()
model = model + "_ft" if model != "all_tasks_mt" else model
name = f"zoo:dodecadialogue/{model.split('.')[-1]}/model"
option = self.set_options(name, device)
super().__init__(
name=model,
suffix="\n",
device=device,
maxlen=maxlen,
model=create_agent_from_opt_file(option),
)
if "wizard_of_wikipedia" in name:
inherit(self, (WizardOfWikipediaAgent, Seq2SeqLM))
self.build_wizard_of_wikipedia()
elif "convai2" in name:
inherit(self, (ConvAI2Agent, Seq2SeqLM))
