def save_image_contexts(task_opt: Opt):
"""
Save a JSON of images and associated contexts for the model image chat task.
Note that each image will have BST-style context information saved with it, such as
persona strings and a pair of lines of dialogue from another dataset.
TODO: perhaps have the image chat task make use of this context information
"""
print('Creating teacher to loop over images.')
agent = RepeatLabelAgent(task_opt)
world = create_task(task_opt, agent)
num_examples = task_opt['num_examples']
print('Creating context generator.')
context_generator = get_context_generator()
print(
f'Looping over {num_examples:d} images and pulling a context for each one.'
)
image_contexts = []
unique_image_srcs = set()
while len(image_contexts) < num_examples:
# Get the next teacher act
world.parley()
teacher_act = world.get_acts()[0]
image_src = get_image_src(image=teacher_act['image'])
if image_src in unique_image_srcs:
# Skip over non-unique images, such as from the later turns of an episode
print('\tSkipping non-unique image.')
else:
unique_image_srcs.add(image_src)
image_context = {
'image_act': teacher_act,
'context_info': context_generator.get_context(),
}
image_contexts.append(image_context)
if len(image_contexts) % 5 == 0:
print(f'Collected {len(image_contexts):d} images.')
print(f'{len(image_contexts):d} image contexts created.')
# Save
with open(task_opt['image_context_path'], 'wb') as f:
pickle.dump(image_contexts, f)
def __init__(self, task_run: "TaskRun", args: "DictConfig",
shared_state: "SharedTaskState"):
conversations_needed = self._process_conversations_needed(args)
self.conversations_needed = conversations_needed
shared_state.conversations_needed = conversations_needed
args.blueprint.num_conversations = sum(conversations_needed.values())
super().__init__(task_run=task_run,
args=args,
shared_state=shared_state)
if args.blueprint.get("annotations_config_path", "") != "":
# We are going to do annotations, so load the onboarding data file that will
# be used to onboard users into knowing how to do the annotations properly
onboard_task_data_path = os.path.expanduser(
args.blueprint.onboard_task_data_path)
with open(onboard_task_data_path, "r") as onboard_task_data_file:
self.onboard_task_data = json.load(onboard_task_data_file)
else:
self.onboard_task_data = None
run_statistics = {r: 0 for (r, v) in self.conversations_needed.items()}
shared_state.run_statistics = run_statistics
context_generator: Optional[ContextGenerator] = None
if (args.blueprint.include_persona
or args.blueprint.conversation_start_mode == 'bst'):
context_generator = get_context_generator(
args.blueprint.override_opt)
shared_state.context_generator = context_generator
# Lock for editing run statistics between threads
statistics_condition = Condition()
# Move shared state into the world and onboarding opts, such that these
# can be used by the worlds
shared_state.onboarding_world_opt.update({
'onboard_statistics':
shared_state.onboard_statistics,
'statistics_condition':
statistics_condition,
'max_onboard_time':
args.blueprint.max_onboard_time,
'onboard_task_data':
self.onboard_task_data,
'onboarding_qualification':
args.blueprint.onboarding_qualification,
})
shared_state.world_opt.update({
'conversations_needed':
conversations_needed,
'run_statistics':
shared_state.run_statistics,
'context_generator':
context_generator,
'statistics_condition':
statistics_condition,
'conversation_start_mode':
args.blueprint.conversation_start_mode,
'include_persona':
args.blueprint.include_persona,
})