def _load_known_opts(self, optfile, parsed):
"""
Pull in CLI args for proper models/tasks/etc.
Called before args are parsed; ``_load_opts`` is used for actually overriding
opts after they are parsed.
"""
new_opt = Opt.load(optfile)
for key, value in new_opt.items():
# existing command line parameters take priority.
if key not in parsed or parsed[key] is None:
parsed[key] = value
def test_nodatafile(self):
for dt in [
'train:ordered',
'train:stream:ordered',
'valid',
'test',
'valid:stream',
'test:stream',
]:
opt = Opt({'datatype': dt, 'datapath': '/tmp', 'task': 'test'})
with self.assertRaises(KeyError):
NoDatafileTeacher(opt)
def _load_opts(self, opt):
optfile = opt.get('init_opt')
new_opt = Opt.load(optfile)
for key, value in new_opt.items():
# existing command line parameters take priority.
if key not in opt:
raise RuntimeError(
'Trying to set opt from file that does not exist: ' +
str(key))
if key not in opt['override']:
opt[key] = value
opt['override'][key] = value
def _initialize_bart(self, opt: Opt) -> Opt:
"""
Download and convert BART pre-trained models.
Additionally, convert `init-fairseq-model` if necessary.
:param opt:
ParlAI-parsed options
:return opt:
return opt with BART-specific args.
"""
if not opt.get('converting'):
download(opt['datapath'])
opt['init_model'] = os.path.join(opt['datapath'],
'models/bart/bart_large/model')
if opt.get('init_fairseq_model'):
opt = self._convert_model(opt)
opt.update(BART_ARGS)
compare_init_model_opts(opt, opt)
return opt
def get_dialogue_task_mutators(opt: Opt) -> str:
"""
Set the mutators appropriately for the dialogue tasks.
"""
mutators = '+'.join([
'flatten', 'extract_entity_for_response_model',
'skip_retrieval_mutator'
])
if opt.get('mutators'):
mutators = '+'.join([mutators, opt['mutators']])
logging.warning(f'overriding mutators to {mutators}')
return mutators
def __init__(self, opt: Opt, shared=None):
"""
Initializes reranker.
"""
self.predictor_model_file = modelzoo_path(opt['datapath'],
opt['predictor_model_file'])
self.reranker_strategy = opt['reranker_strategy']
self.normalize_candidates = opt['normalize_candidates']
self.delimiter = opt.get('delimiter', '\n')
self.include_context = True
self.include_label_cand_only = False
self.init_predictor(opt, shared)
def test_multitask(self):
"""
Test that model correctly handles multiple inputs.
Random chance is 10%, so this should be able to get much better than that very
quickly.
"""
args = Opt({**self.base_args, **self.multitask_args})
valid, test = testing_utils.train_model(args)
assert (
valid['accuracy'] > 0.2
), f'ImagePolyencoderAgent val-set accuracy on a simple task was {valid["accuracy"].value():0.2f}.'
def create_task(opt: Opt, user_agents, default_world=None):
"""
Create a world + task_agents (aka a task).
Assuming ``opt['task']="task_dir:teacher_class:options"`` e.g. ``"babi:Task1k:1"``
or ``"#babi-1k"`` or ``"#QA"``, see ``parlai/tasks/tasks.py`` and see
``parlai/tasks/task_list.py`` for list of tasks.
"""
task = opt.get('task')
if not task:
raise RuntimeError(
'No task specified. Please select a task with ' + '--task {task_name}.'
)
if type(user_agents) != list:
user_agents = [user_agents]
# Convert any hashtag task labels to task directory path names.
# (e.g. "#QA" to the list of tasks that are QA tasks).
opt = copy.deepcopy(opt)
opt['task'] = ids_to_tasks(opt['task'])
print('[creating task(s): ' + opt['task'] + ']')
# check if single or multithreaded, and single-example or batched examples
if ',' not in opt['task']:
# Single task
world = create_task_world(opt, user_agents, default_world=default_world)
else:
# Multitask teacher/agent
# TODO: remove and replace with multiteachers only?
world = MultiWorld(opt, user_agents, default_world=default_world)
if opt.get('numthreads', 1) > 1:
# use hogwild world if more than one thread requested
# hogwild world will create sub batch worlds as well if bsz > 1
world = HogwildWorld(opt, world)
elif opt.get('batchsize', 1) > 1:
# otherwise check if should use batchworld
world = BatchWorld(opt, world)
return world
def __init__(self, opt: Opt, agents=None, shared=None, default_world=None):
super().__init__(opt)
self.worlds: List[World] = []
for index, k in enumerate(opt['task'].split(',')):
k = k.strip()
if k:
if shared:
# Create worlds based on shared data.
s = shared['worlds'][index]
self.worlds.append(s['world_class'](s['opt'], None, s))
else:
# Agents are already specified.
opt_singletask = copy.deepcopy(opt)
opt_singletask['task'] = k
self.worlds.append(
create_task_world(
opt_singletask, agents, default_world=default_world
)
)
self.world_idx = -1
self.new_world = True
self.parleys = -1
# Check to see if we are training
self.is_training = DatatypeHelper.is_training(opt.get('datatype'))
# Make multi-task task probabilities.
self.cum_task_weights = [1] * len(self.worlds)
self.task_choices = range(len(self.worlds))
weights = self.opt.get('multitask_weights', [1])
if weights == 'stochastic':
weights = [w.num_episodes() for w in self.worlds]
sum = 0
for i in self.task_choices:
if len(weights) > i:
weight = weights[i]
else:
weight = 1
self.cum_task_weights[i] = weight + sum
sum += weight
task_ids: Dict[str, Teacher] = {}
# Having overlap in teacher ids will cause issues for metrics aggregation.
for each_world in self.worlds:
world_id = each_world.getID()
if world_id in task_ids:
raise AssertionError(
'{} and {} teachers have overlap in id {}.'.format(
task_ids[world_id],
each_world.get_agents()[0].__class__,
world_id,
)
)
else:
task_ids[world_id] = each_world.get_task_agent()
def bpe_factory(opt: Opt, shared: TShared) -> 'BPEHelper':
"""
BPE Helper Factory.
Returns the appropriate BPE helper given the opt
as well as available libraries.
:param opt:
options
:param shared:
shared dict
:return BPEHelper:
returns the appropriate BPEHelper object
"""
from parlai.core.dict import DictionaryAgent
tokenizer = opt.get('dict_tokenizer', DictionaryAgent.default_tok)
bpe_helper: Optional[BPEHelper] = None
if tokenizer == 'bytelevelbpe':
# Attempt to instantiate HF tokenizer
try:
bpe_helper = HuggingFaceBpeHelper(opt, shared)
except ImportError:
if opt['dict_loaded']:
warn_once(
''
'\n\n--------------------------------------------------\n\n'
'WARNING: You have chosen to use Huggingface\'s tokenizer.\n'
'Please install HuggingFace tokenizer with: pip install tokenizers.\n'
'For now, defaulting to the GPT2Tokenizer.'
'\n\n--------------------------------------------------\n\n'
)
tokenizer = 'slow_bytelevel_bpe'
else:
raise ImportError(
'Please install HuggingFace tokenizer with: pip install tokenizers.\n'
)
if tokenizer == 'slow_bytelevel_bpe':
bpe_helper = SlowBytelevelBPE(opt, shared)
if tokenizer == 'gpt2':
bpe_helper = Gpt2BpeHelper(opt, shared)
if tokenizer == 'bpe':
bpe_helper = SubwordBPEHelper(opt, shared)
assert (
bpe_helper is not None
), f"bpe_factory called with invalid tokenizer: {tokenizer}"
return bpe_helper
def test_gpt2_bpe_tokenize(self):
opt = Opt({'dict_tokenizer': 'gpt2', 'datapath': './data'})
agent = DictionaryAgent(opt)
self.assertEqual(
# grinning face emoji
agent.gpt2_tokenize(u'Hello, ParlAI! \U0001f600'),
GPT2_BPE_RESULT,
)
self.assertEqual(
agent.vec2txt(agent.tok2ind[w] for w in GPT2_BPE_RESULT),
# grinning face emoji
u'Hello, ParlAI! \U0001f600',
)
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: Opt, shared=None):
"""
Setup reranker.
"""
super().__init__(opt, shared)
reranker_class = self.get_reranker_class()
self.inference_strategies = (opt['inference_strategies']
or opt['inference']).split(',')
self.debug_mode = opt.get('debug_mode', False)
if not shared:
self.reranker = reranker_class(opt, shared=None)
else:
self.reranker = reranker_class(opt, shared=shared['reranker'])
def test_safe_personas(self):
base_kwargs = Opt({'datatype': 'train', 'task': 'blended_skill_talk'})
safe_personas_only_to_count = {False: 4819, True: 3890}
for safe_personas_only, count in safe_personas_only_to_count.items():
full_kwargs = {
**base_kwargs, 'safe_personas_only': safe_personas_only
}
parser = setup_args()
parser.set_defaults(**full_kwargs)
opt = parser.parse_args([])
personas = _load_personas(opt)
self.assertEqual(len(personas), count)
def test_opt(self):
opt = {'x': 0}
opt = Opt(opt)
opt['x'] += 1
opt['x'] = 10
history = opt.history['x']
self.assertEqual(history[0][1], 1, 'History not set properly')
self.assertEqual(history[1][1], 10, 'History not set properly')
opt_copy = deepcopy(opt)
history = opt_copy.history['x']
self.assertEqual(history[0][1], 1, 'Deepcopy history not set properly')
self.assertEqual(history[1][1], 10, 'Deepcopy history not set properly')
def _all_split_datafiles(opt: Opt) -> List[str]:
datafiles = []
split_type = SplitType(opt.get("cmu_dog_split_type"))
if split_type in {SplitType.SEEN, SplitType.UNSEEN}:
# For seen/unseen split, the full set of dialogs is split
# across train, valid, test seen, and test unseen
for split in ['train', 'valid', 'test']:
datafiles.append(_datafile(split, SplitType.SEEN))
datafiles.append(_datafile('test', SplitType.UNSEEN))
else:
for split in ['train', 'valid', 'test']:
datafiles.append(_datafile(split, split_type))
return datafiles
def create_task(opt: Opt, user_agents, default_world=None):
"""
Create a world + task_agents (aka a task).
Assuming ``opt['task']="task_dir:teacher_class:options"`` e.g. ``"babi:Task1k:1"``
or ``"#babi-1k"`` or ``"#QA"``, see ``parlai/tasks/tasks.py`` and see
``parlai/tasks/task_list.py`` for list of tasks.
"""
task = opt.get('task')
if not task:
raise RuntimeError('No task specified. Please select a task with ' +
'--task {task_name}.')
if type(user_agents) != list:
user_agents = [user_agents]
# Convert any hashtag task labels to task directory path names.
# (e.g. "#QA" to the list of tasks that are QA tasks).
opt = copy.deepcopy(opt)
opt['task'] = ids_to_tasks(opt['task'])
logging.info(f"creating task(s): {opt['task']}")
if ',' not in opt['task']:
# Single task
world = create_task_world(opt,
user_agents,
default_world=default_world)
else:
# Multitask teacher/agent
# TODO: remove and replace with multiteachers only?
world = MultiWorld(opt, user_agents, default_world=default_world)
if opt.get('batchsize', 1) > 1 and opt.get('dynamic_batching'):
world = DynamicBatchWorld(opt, world)
elif opt.get('batchsize', 1) > 1:
# otherwise check if should use batchworld
world = BatchWorld(opt, world)
return world
def __init__(self, opt: Opt, dictionary: DictionaryAgent):
if opt.get('n_positions'):
# if the number of positions is explicitly provided, use that
n_positions = opt['n_positions']
else:
# else, use the worst case from truncate
n_positions = max(
opt.get('truncate') or 0,
opt.get('text_truncate') or 0,
opt.get('label_truncate') or 0,
)
if n_positions == 0:
# default to 1024
n_positions = 1024
super().__init__(opt, dictionary)
self.encoder = ContextWithImageEncoder(
n_heads=opt['n_heads'],
n_layers=opt['n_layers'],
embedding_size=opt['embedding_size'],
ffn_size=opt['ffn_size'],
vocabulary_size=len(dictionary),
embedding=self.embeddings,
dropout=opt['dropout'],
attention_dropout=opt['attention_dropout'],
relu_dropout=opt['relu_dropout'],
padding_idx=self.pad_idx,
learn_positional_embeddings=opt['learn_positional_embeddings'],
embeddings_scale=opt['embeddings_scale'],
n_positions=n_positions,
n_segments=opt.get('n_segments', 0),
activation=opt['activation'],
variant=opt['variant'],
output_scaling=opt['output_scaling'],
image_encoder_num_layers=opt['image_encoder_num_layers'],
image_features_dim=opt['image_features_dim'],
fusion=opt['image_fusion_type'],
)
def test_asymmetry(self):
opt = Opt({'model': 'transformer/generator', 'n_layers': 1})
agent = create_agent(opt)
self.assertEqual(agent.model.encoder.n_layers, 1)
self.assertEqual(agent.model.decoder.n_layers, 1)
opt = Opt({
'model': 'transformer/generator',
'n_layers': 1,
'n_encoder_layers': 2
})
agent = create_agent(opt)
self.assertEqual(agent.model.encoder.n_layers, 2)
self.assertEqual(agent.model.decoder.n_layers, 1)
opt = Opt({
'model': 'transformer/generator',
'n_layers': 1,
'n_encoder_layers': 2,
'n_decoder_layers': 4,
})
agent = create_agent(opt)
self.assertEqual(agent.model.encoder.n_layers, 2)
self.assertEqual(agent.model.decoder.n_layers, 4)
opt = Opt({
'model': 'transformer/generator',
'n_layers': 1,
'n_decoder_layers': 4
})
agent = create_agent(opt)
self.assertEqual(agent.model.encoder.n_layers, 1)
self.assertEqual(agent.model.decoder.n_layers, 4)
opt = Opt({'model': 'transformer/generator'})
agent = create_agent(opt)
self.assertEqual(agent.model.encoder.n_layers, 2)
self.assertEqual(agent.model.decoder.n_layers, 2)
def get_encoder(
self,
opt: Opt,
dict_: DictionaryAgent,
embeddings: torch.nn.Embedding,
module_klass: Type[TransformerEncoder],
null_idx: int,
reduction_type: Optional[str] = None,
):
"""
Return encoder, given options.
Ensures that multiobjective options are copied correctly.
:param opt:
opt dict
:param dict:
dictionary agent
:param null_idx:
null/pad index into dict
:param reduction_type:
reduction type for the encoder
:return:
a TransformerEncoder, initialized correctly
"""
opt = copy.deepcopy(opt)
opt['n_heads'] = opt.get('n_multiobjective_heads', 4)
opt['n_layers'] = opt.get('n_multiobjective_layers', 2)
opt['n_encoder_layers'] = opt.get('n_multiobjective_layers', 2)
opt['n_decoder_layers'] = opt.get('n_multiobjective_layers', 2)
return module_klass(
opt=opt,
vocabulary_size=len(dict_),
embedding=embeddings,
padding_idx=null_idx,
reduction_type=reduction_type,
n_segments=opt.get('n_segments', 2),
)
def _create_task_agents(opt: Opt):
"""
Create task agent(s) for the given task name.
It does this by calling the create_agent function in agents.py of the given task. If
create_agents function does not exist, it just looks for the teacher (agent) class
defined by the task name directly. (This saves the task creator bothering to define
the create_agents function when it is not needed.)
"""
if opt.get('interactive_task', False) or opt.get('selfchat_task', False):
# do not need task agents in interactive or self chat settings
return []
try:
# Tries to call the create_agent function in agents.py
my_module = load_task_module(opt['task'])
task_agents = my_module.create_agents(opt) # type: ignore
except (ModuleNotFoundError, AttributeError):
# Create_agent not found, so try to create the teacher directly.
return create_task_agent_from_taskname(opt)
if type(task_agents) != list:
task_agents = [task_agents]
return task_agents
def test_gpt2_bpe_tokenize(self):
datapath = ParlaiParser().parse_args([], print_args=False)['datapath']
opt = Opt({'dict_tokenizer': 'gpt2', 'datapath': datapath})
agent = DictionaryAgent(opt)
self.assertEqual(
# grinning face emoji
agent.gpt2_tokenize(u'Hello, ParlAI! \U0001f600'),
GPT2_BPE_RESULT,
)
self.assertEqual(
agent.vec2txt(agent.tok2ind[w] for w in GPT2_BPE_RESULT),
# grinning face emoji
u'Hello, ParlAI! \U0001f600',
)
def __init__(self, opt: Opt, dictionary: DictionaryAgent, retriever_shared=None):
super().__init__(opt, dictionary, retriever_shared)
if opt.get('filter_docs_with_label'):
assert (
RetrieverType(opt['rag_retriever_type']) == RetrieverType.SEARCH_ENGINE
)
self.retriever = FilterDocsForLabelSearchEngineRetrieverCombo(
opt, dictionary, shared=retriever_shared
) # type: ignore
else:
self.retriever = combo_fid_retriever_factory(
opt, dictionary, shared=retriever_shared
)
self.top_docs: List[List[Document]] = []
def __init__(self, opt: Opt, shared=None):
if not shared:
self.hf_tokenizer = self.get_tokenizer(opt)
self.tok2ind = self.hf_tokenizer.get_vocab()
self.ind2tok = {v: k for k, v in self.tok2ind.items()}
else:
self.hf_tokenizer = shared['hf_tokenizer']
self.tok2ind = shared['tok2ind']
self.ind2tok = shared['ind2tok']
self.freq = defaultdict(int)
for tok in self.tok2ind:
self.freq[tok] = 1
self.minfreq = opt.get('dict_minfreq', DictionaryAgent.default_minfreq)
self._unk_token_idx = self.hf_tokenizer.unk_token_id
self.override_special_tokens(opt)
self.lower = opt.get('dict_lower', DictionaryAgent.default_lower)
self.tokenizer = 'hf'
self.opt = opt
self.max_length = (self.opt.get('text_truncate')
or self.hf_tokenizer.model_max_length)
def __init__(self, opt: Opt, shared: TShared = None):
if not shared:
self.idx_to_ep = {}
else:
self.idx_to_ep = shared['idx_to_ep']
self.prepend_personality = opt.get('prepend_personality', True)
self.include_dialogue_history = opt.get('include_dialogue_history',
True)
self.category_frac = opt.get('category_frac', 0.0)
super().__init__(opt, shared)
self.num_eps = len(self.data) + len(
[d for d in self.data if len(d['dialog']) > 1])
# Replace personalities with polarity categories ("positive/neutral" or
# "negative"), with probability self.category_frac
if not shared:
category_map = get_category_map(self.personalities)
for i, d in enumerate(self.data):
use_category_rand = random.random()
if use_category_rand < self.category_frac:
self.data[i]['dialog'] = [[
category_map[personality], label
] for personality, label in d['dialog']]
def init_predictor(self, opt: Opt, shared=None):
"""
Initializes Predictor Module
"""
if not shared:
if not opt.get("predictor_model_file"):
logging.warn(
'Reranker MUST specify predictor_model_file unless subclass __init__() sets up the model in its own way (unusual). Skipping predictor setup!'
)
else:
self.predictor = create_agent_from_model_file(
self.predictor_model_file)
else:
self.predictor = shared['predictor']
def __init__(self, opt: Opt, shared: TShared = None):
assert opt['flatten_delimiter'] == opt.get(
'delimiter', '\n'
), '--flatten-delimiter and --delimiter are set differently, please inspect and set to the same to avoid unexpected results'
self.opt = opt
if shared and 'data' in shared:
self.data = shared['data']
else:
self.word_lists = self.build_wordlists(opt)
self.data = self._setup_data(opt)
super().__init__(opt, shared)
self.reset()
def combo_fid_retriever_factory(opt: Opt,
dictionary: DictionaryAgent,
shared=None) -> Optional[RagRetriever]:
"""
Bypass call to standard retriever factory to possibly build our own retriever.
"""
if opt.get('converting'):
return None
retriever = RetrieverType(opt['rag_retriever_type'])
if retriever is RetrieverType.SEARCH_ENGINE:
return ComboFidSearchQuerySearchEngineRetriever(
opt, dictionary, shared=shared) # type: ignore
else:
return retriever_factory(opt, dictionary, shared)
def init_predictor(self, opt: Opt, shared=None):
if not shared:
override = {
'return_cand_scores': True,
'datatype': 'valid',
'interactive_mode': opt.get('interactive_mode', True),
'ignore_bad_candidates': True,
'encode_candidate_vecs': True,
'interactive_candidates': 'inline',
} # to not init optim
self.predictor = create_agent_from_model_file(
self.predictor_model_file, opt_overrides=override)
else:
self.predictor = shared['predictor']
def update_state_dict(opt: Opt, state_dict: Dict[str, torch.Tensor],
model: torch.nn.Module):
"""
Update the given state dict to be RAG-ified.
:param opt:
options
:param state_dict:
weights to load
:param model:
underlying model that will load the state_dict
:return updated_state_dict:
return state_dict with appropriate keys/values
"""
# 1. Substitute all "encoder" and "decoder" keys with "seq2seq_encoder" and "seq2seq_decoder"
if not [k for k in state_dict if k.startswith('seq2seq')]:
for k in list(state_dict.keys()):
if k.startswith('encoder') or k.startswith('decoder'):
weights = state_dict.pop(k)
state_dict[f'seq2seq_{k}'] = weights
# 2. Retriever state
if not [k for k in state_dict if 'retriever' in k]:
retriever_state = {
f"retriever.{k}": v
for k, v in
model.retriever.state_dict().items() # type: ignore
}
state_dict.update(retriever_state)
# 3. Handle n_positional difference
if opt.get('n_extra_positions', 0) > 0:
key = 'seq2seq_encoder.position_embeddings.weight'
init_weight = (
model.seq2seq_encoder.position_embeddings.
weight # type: ignore
)
if state_dict[key].size(
0) < opt['n_positions'] + opt['n_extra_positions']:
# Make sure we're not adding more positions to a model trained
# with extra positions
state_dict[key] = torch.cat(
[
state_dict[key].to(init_weight), # type: ignore
init_weight[
-opt['n_extra_positions']:, :], # type: ignore
],
dim=0,
)
return state_dict
