def observe(self, observation, increment_turn: bool = True):
"""
Need to protect the observe also with a semaphore for composed models where an
act() may be called within an observe()
"""
logging.info(
f'{self.__class__.__name__}: In observe() before semaphore, self.turn_idx is {self.turn_idx} and observation is {observation}'
)
new_ob = copy.deepcopy(observation)
if self.semaphore:
with self.semaphore:
self.model_agent.observe(new_ob)
else:
self.model_agent.observe(new_ob)
if 'text' not in new_ob:
logging.warning(
f'{self.__class__.__name__}: In observe() AFTER semaphore, self.turn_idx: {self.turn_idx}, and observation is missing a "text" field: {new_ob}'
)
else:
logging.info(
f'{self.__class__.__name__}: In observe() AFTER semaphore, self.turn_idx: {self.turn_idx}, observation["text"]: {new_ob["text"]}'
)
if increment_turn:
self.turn_idx += 1
def __init__(self, opt, shared=None):
mutators = '+'.join(
['do_generate_search_query_mutator', 'skip_retrieval_mutator'])
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
opt['mutators'] = mutators
super().__init__(opt, shared)
self.id = 'NQSearchDecisionTeacher'
def __init__(self, opt: Opt):
"""
Initialize IVFPQ FAISS Indexer.
The IVFPQ Indexer is a great way to reduce memory footprint of dense embeddings.
"""
super().__init__(opt)
self.dim = opt['retriever_embedding_size']
self.use_gpu_train = (not opt['no_cuda'] and torch.cuda.is_available()
and opt['compressed_indexer_gpu_train'])
self.hnsw_ef_search = opt['hnsw_ef_search']
self.index_factory = opt['compressed_indexer_factory']
if self.index_factory:
logging.warning(
f'Creating Index from Index Factory: {self.index_factory}')
self.is_ivf_index = 'IVF' in self.index_factory
self.index = self.faiss.index_factory(
self.dim, self.index_factory, self.faiss.METRIC_INNER_PRODUCT)
else:
self.is_ivf_index = True
quantizer = self.faiss.IndexHNSWFlat(
self.dim, opt['hnsw_indexer_store_n'],
self.faiss.METRIC_INNER_PRODUCT)
quantizer.hnsw.efConstruction = opt['hnsw_ef_construction']
quantizer.hnsw.efSearch = opt['hnsw_ef_search']
ivf_index = self.faiss.IndexIVFPQ(quantizer, self.dim, 4096, 128,
8,
self.faiss.METRIC_INNER_PRODUCT)
ivf_index.nprobe = opt['compressed_indexer_nprobe']
self.index = ivf_index
if self.is_ivf_index:
self.index_ivf = self.faiss.extract_index_ivf(self.index)
self.index_ivf.metric_type = self.faiss.METRIC_INNER_PRODUCT
self.nlist = self.index_ivf.nlist
self.index_ivf.verbose = True
self.downcast_quantizer = self.faiss.downcast_index(
self.index_ivf.quantizer)
self.downcast_quantizer.verbose = True
self.downcast_quantizer.metric_type = self.faiss.METRIC_INNER_PRODUCT
if hasattr(self.downcast_quantizer, 'hnsw'):
self.downcast_quantizer.hnsw.efSearch = opt['hnsw_ef_search']
self.downcast_quantizer.hnsw.efConstruction = opt[
'hnsw_ef_construction']
self.downcast_quantizer.hnsw.metric_type = (
self.faiss.METRIC_INNER_PRODUCT)
self.setup_gpu_train()
self.index.nprobe = opt['compressed_indexer_nprobe']
self.nprobe = opt['compressed_indexer_nprobe']
self.span = 5 # arbitrarily chosen, from prior evidence
self.random = random.Random(42)
def setup_gpu_train(self):
"""
Setup training on the gpu.
"""
if self.use_gpu_train:
logging.warning('Will train index on GPU')
try:
clustering_index = self.faiss.index_cpu_to_all_gpus(
self.faiss.IndexFlatIP(self.index_ivf.d))
self.index.clustering_index = clustering_index
except NameError:
logging.warning(
'GPU training not supported; switching to CPU.')
def get_dialogue_task_mutators(opt: Opt) -> str:
"""
Set the mutators appropriately for the dialogue tasks.
"""
mutators = '+'.join([
'flatten',
'skip_retrieval_mutator',
'bst_tasks_maybe_generate_search_query_mutator',
])
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
return mutators
def __init__(self, opt, shared=None):
mutators = '+'.join([
'flatten',
'wow_maybe_generate_search_query_mutator',
'skip_retrieval_mutator',
])
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
opt['mutators'] = mutators
opt['add_missing_turns'] = 'all'
super().__init__(opt, shared)
self.id = 'WowSearchDecisionTeacher'
def __init__(self, opt, shared=None):
mutators = '+'.join([
'flatten',
'woi_dropout_retrieved_docs',
'woi_maybe_generate_search_query_mutator',
'skip_retrieval_mutator',
])
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
opt['mutators'] = mutators
super().__init__(opt, shared)
self.id = 'WoiSearchDecisionTeacher'
def __init__(self, opt, shared=None):
mutators = '+'.join([
'ms_marco_filter_has_answer',
'ms_marco_create_fid_docs',
'ms_marco_find_selected_sentence_for_response',
'woi_pop_documents_mutator',
'skip_retrieval_mutator',
])
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
opt['mutators'] = mutators
super().__init__(opt, shared)
self.id = "MsMarcoDialogueTeacher"
def __init__(self, opt, shared=None):
mutators = '+'.join([
'flatten',
'woi_pop_documents_mutator',
'woi_filter_no_passage_used',
'woi_add_checked_sentence_to_input',
'skip_retrieval_mutator',
] + opt.get('mutators', '').split('+'))
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
opt['mutators'] = mutators
super().__init__(opt, shared)
self.id = "WoiDialogueTeacher"