def __init__(self,
opt,
datatype: str = 'train',
seed: Optional[int] = None):
"""
Initalize the context generator.
opt: only a 'datapath' key is required, to specify the ParlAI data folder
"""
if seed is not None:
self.rng = random.Random(seed)
else:
self.rng = random.Random()
convai2_opt = Opt({'datapath': opt['datapath'], 'datatype': datatype})
self.convai2_teacher = BothTeacher(convai2_opt)
ed_opt = Opt({
'datapath': opt['datapath'],
'datatype': datatype,
'train_experiencer_only': True,
})
# Specify train_experiencer_only = True because we want to ensure that the text
# will correspond to a Speaker utterance and the label to a Listener response
self.ed_teacher = EmpatheticDialoguesTeacher(ed_opt)
wow_opt = Opt({'datapath': opt['datapath'], 'datatype': datatype})
self.wow_teacher = WizardDialogKnowledgeTeacher(wow_opt)
self.topic_to_persona_path = _topic_to_persona_path(opt)
self.wow_topics_to_episode_idxes = self._setup_topics_to_episodes()
self.persona_strings_to_wow_topics = self._setup_personas_to_topics()
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 test_beamsearch_contextblocking(self):
"""
Test beamsearch context blocking.
"""
agent = create_agent_from_model_file(
'zoo:unittest/context_blocking/model')
agent.observe({'text': '5 4 3 2', 'episode_done': True})
assert agent.act()['text'] == '5 4 3 2'
agent = create_agent_from_model_file(
'zoo:unittest/context_blocking/model',
Opt(beam_context_block_ngram=1))
agent.observe({'text': '5 4 3 2', 'episode_done': True})
text = agent.act()['text']
assert '5' not in text
assert '4' not in text
assert '3' not in text
assert '2' not in text
agent = create_agent_from_model_file(
'zoo:unittest/context_blocking/model',
Opt(beam_context_block_ngram=2))
agent.observe({'text': '5 4 3 2', 'episode_done': True})
text = agent.act()['text']
assert '5' in text
assert '5 4' not in text
assert '4 3' not in text
assert '3 2' not in text
def get_tl(tmpdir):
final_opt = Opt({
'task': 'integration_tests',
'datatype': 'valid',
'validation_max_exs': 30,
'short_final_eval': True,
})
final_opt.save(os.path.join(tmpdir, "final_opt.opt"))
opt = Opt({
'task':
'integration_tests',
'validation_max_exs':
10,
'model':
'repeat_label',
'model_file':
os.path.join(tmpdir, 'model'),
'short_final_eval':
True,
'num_epochs':
1.0,
'final_extra_opt':
str(os.path.join(tmpdir, "final_opt.opt")),
})
parser = tms.setup_args()
parser.set_params(**opt)
popt = parser.parse_args([])
for k, v in opt.items():
popt[k] = v
return tms.TrainLoop(popt)
def test_counts(self):
with testing_utils.tempdir() as tmpdir:
data_path = tmpdir
# Check EmpatheticDialoguesTeacher, with multiple examples per episode
opts_episodes_and_examples = [
(
{'datatype': 'train'},
EPISODE_COUNTS['train_both_sides'],
EXAMPLE_COUNTS['train_both_sides'],
), # Test the default mode
(
{'datatype': 'train', 'train_experiencer_only': True},
EPISODE_COUNTS['train_experiencer_only'],
EXAMPLE_COUNTS['train_experiencer_only'],
),
(
{'datatype': 'train', 'train_experiencer_only': False},
EPISODE_COUNTS['train_both_sides'],
EXAMPLE_COUNTS['train_both_sides'],
),
(
{'datatype': 'valid'},
EPISODE_COUNTS['valid'],
EXAMPLE_COUNTS['valid'],
),
({'datatype': 'test'}, EPISODE_COUNTS['test'], EXAMPLE_COUNTS['test']),
]
for teacher_class in [EmpatheticDialoguesTeacher]:
for opt, num_episodes, num_examples in opts_episodes_and_examples:
full_opt = Opt({**opt, 'datapath': data_path})
teacher = teacher_class(full_opt)
self.assertEqual(teacher.num_episodes(), num_episodes)
self.assertEqual(teacher.num_examples(), num_examples)
# Check EmotionClassificationSituationTeacher, with one example per episode
train_episode_count = EPISODE_COUNTS['train_experiencer_only']
# For the situation classifier, we only want to have one episode per train
# conversation
opts_episodes = [
({'datatype': 'train'}, train_episode_count), # Test the default mode
(
{'datatype': 'train', 'train_experiencer_only': True},
train_episode_count,
),
(
{'datatype': 'train', 'train_experiencer_only': False},
train_episode_count,
),
({'datatype': 'valid'}, EPISODE_COUNTS['valid']),
({'datatype': 'test'}, EPISODE_COUNTS['test']),
]
for teacher_class in [EmotionClassificationSituationTeacher]:
for opt, num_episodes in opts_episodes:
full_opt = Opt({**opt, 'datapath': data_path})
teacher = teacher_class(full_opt)
self.assertEqual(teacher.num_episodes(), num_episodes)
self.assertEqual(teacher.num_examples(), num_episodes)
def _test_iterate(self, teacher_class):
for dt in [
'train:ordered',
'train:stream:ordered',
'valid',
'test',
'valid:stream',
'test:stream',
]:
opt = Opt({'datatype': dt, 'datapath': '/tmp', 'task': 'test'})
teacher = teacher_class(opt)
self._verify_act(teacher.act(), 1, 2, False)
self._verify_act(teacher.act(), 2, 4, False)
self._verify_act(teacher.act(), 3, 6, True)
self._verify_act(teacher.act(), 1, 2, False)
self._verify_act(teacher.act(), 2, 4, False)
self._verify_act(teacher.act(), 3, 6, True)
self._verify_act(teacher.act(), 1, 2, False)
self._verify_act(teacher.act(), 2, 4, False)
self._verify_act(teacher.act(), 3, 6, True)
assert teacher.epoch_done()
def test_save_load(self):
o = Opt({'a': 3, 'b': 'foo'})
with testing_utils.tempdir() as tmpdir:
fn = os.path.join(tmpdir, "opt")
o.save(fn)
o2 = Opt.load(fn)
assert o == o2
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 test_allow_missing_init_opts(self):
"""
Test --allow-missing-init-opts.
"""
with testing_utils.tempdir() as temp_dir:
init_opt_path = os.path.join(temp_dir, 'init_opt.opt')
# Save a test opt file with an argument that doesn't exist
init_opt = Opt({'made_up_arg': 'foo'})
init_opt.save(init_opt_path)
# Assert that the opt file normally can't be loaded in
with self.assertRaises(RuntimeError):
_ = ParlaiParser(True,
True).parse_kwargs(init_opt=init_opt_path)
# Assert that the opt file *can* be loaded in if we set
# --allow-missing-init-opts, and assert that the made-up arg does not exist
# in the opt
opt = ParlaiParser(True,
True).parse_kwargs(init_opt=init_opt_path,
allow_missing_init_opts=True)
self.assertFalse(hasattr(opt, 'made_up_arg'))
def add_extra_args(self, args=None):
super().add_extra_args(args)
parsed = vars(self.parse_known_args(args, nohelp=True)[0])
# Also load extra args options if a file is given.
if parsed.get("init_opt") is not None:
try:
self._load_known_opts(parsed.get("init_opt"), parsed)
except FileNotFoundError:
# don't die if -o isn't found here. See comment in second call
# later on.
pass
parsed = self._infer_datapath(parsed)
partial = Opt(parsed)
for model in [
"system_model",
"user_model",
"api_schema_grounding_model",
"goal_grounding_model",
"api_resp_model",
]:
if (model in partial and partial[model] is not None
and len(partial[model]) > 0):
self.add_model_subargs(partial[model], partial)
for model_file_prefix in ["system", "user"]:
key = model_file_prefix + "_model_file"
if key in partial and partial[key] and len(partial[key]) > 0:
model_name = self._get_model_name_from_model_file(key, partial)
self.add_model_subargs(model_name, partial)
def test_gpt2_bpe_tokenize(self):
with testing_utils.capture_output():
opt = Opt({'dict_tokenizer': 'gpt2', 'datapath': './data'})
agent = DictionaryAgent(opt)
self.assertEqual(
# grinning face emoji
agent.gpt2_tokenize(u'Hello, ParlAI! \U0001f600'),
[
'Hello',
',',
r'\xc4\xa0Par',
'l',
'AI',
'!',
r'\xc4\xa0\xc3\xb0\xc5\x81\xc4\xba',
r'\xc4\xa2',
],
)
self.assertEqual(
agent.vec2txt(agent.tok2ind[w] for w in [
'Hello',
',',
r'\xc4\xa0Par',
'l',
'AI',
'!',
r'\xc4\xa0\xc3\xb0\xc5\x81\xc4\xba',
r'\xc4\xa2',
]),
# grinning face emoji
u'Hello, ParlAI! \U0001f600',
)
def run(self):
"""
1) load model 2) generate embeddings 3) save embeddings.
"""
self.use_cuda = not self.opt.get('no_cuda') and torch.cuda.is_available()
overrides = {'interactive_mode': True, 'interactive_candidates': 'inline'}
if self.opt['dpr_model']:
overrides.update(
{
'model': 'dpr_agent',
'model_file': self.opt['model_file'],
'override': {
'model': 'dpr_agent',
'interactive_candidates': 'inline',
},
}
)
agent = create_agent(Opt(overrides))
else:
agent = create_agent_from_model_file(self.opt['model_file'], overrides)
model = agent.model.module if hasattr(agent.model, 'module') else agent.model
assert hasattr(model, 'encoder_cand') or hasattr(model, 'cand_encoder')
assert isinstance(agent, TorchRankerAgent)
passages = self.load_passages()
data = self.encode_passages(agent, passages)
self.save_data(data)
def _build_model(self, opt: Opt) -> Tuple[PolyEncoderModule, DictionaryAgent]:
"""
Build poly-encoder module.
:param opt:
options from base RAG Model
:return dropout poly-encoder:
return dropout poly agent.
"""
model_file = modelzoo_path(opt['datapath'], opt['poly_faiss_model_file'])
model_opt = Opt.load(f'{model_file}.opt')
create_model_opt = {
**{k: model_opt[k] for k in TRANSFORMER_RANKER_BASE_OPT},
**{k: model_opt[k] for k in POLYENCODER_OPT_KEYS},
'model': 'transformer/dropout_poly',
'init_model': model_file,
'dict_file': f'{model_file}.dict',
# necessary opt args
'multitask_weights': [1],
# dropout_poly args
'poly_dropout_reduction_type': model_opt['poly_dropout_reduction_type'],
'poly_dropout_use_codes': model_opt.get('poly_dropout_use_codes', True),
}
logging.disable()
agent = create_agent(Opt(create_model_opt))
logging.enable()
assert isinstance(agent, DropoutPolyAgent)
return agent.model, agent.dict
def add_extra_args(self, args=None):
"""
Add more args depending on how known args are set.
"""
parsed = vars(self.parse_known_args(args, nohelp=True)[0])
# Also load extra args options if a file is given.
if parsed.get('init_opt') is not None:
try:
self._load_known_opts(parsed.get('init_opt'), parsed)
except FileNotFoundError:
# don't die if -o isn't found here. See comment in second call
# later on.
pass
parsed = self._infer_datapath(parsed)
partial = Opt(parsed)
# find which image mode specified if any, and add additional arguments
image_mode = parsed.get('image_mode', None)
if image_mode is not None and image_mode != 'no_image_model':
self.add_image_args(image_mode)
# find which task specified if any, and add its specific arguments
task = parsed.get('task', None)
if task is not None:
self.add_task_args(task, partial)
evaltask = parsed.get('evaltask', None)
if evaltask is not None:
self.add_task_args(evaltask, partial)
# find which model specified if any, and add its specific arguments
model = get_model_name(parsed)
if model is not None:
self.add_model_subargs(model, partial)
# add world args, if we know a priori which world is being used
if task is not None:
self.add_world_args(
task,
parsed.get('interactive_task', False),
parsed.get('selfchat_task', False),
partial,
)
# reparse args now that we've inferred some things. specifically helps
# with a misparse of `-opt` as `-o pt`, which causes opt loading to
# try to load the file "pt" which doesn't exist.
# After adding model arguments, -opt becomes known (it's in TorchAgent),
# and we parse the `-opt` value correctly.
parsed = vars(self.parse_known_args(args, nohelp=True)[0])
if parsed.get('init_opt') is not None:
self._load_known_opts(parsed.get('init_opt'), parsed)
# reset parser-level defaults over any model-level defaults
try:
self.set_defaults(**self._defaults)
except AttributeError:
raise RuntimeError('Please file an issue on github that argparse '
'got an attribute error when parsing.')
def _process_args_to_opts(self,
args_that_override: Optional[List[str]] = None):
self.opt = Opt(vars(self.args))
# custom post-parsing
self.opt['parlai_home'] = self.parlai_home
self.opt = self._infer_datapath(self.opt)
# set all arguments specified in command line as overridable
option_strings_dict = {}
store_true = []
store_false = []
for group in self._action_groups:
for a in group._group_actions:
if hasattr(a, 'option_strings'):
for option in a.option_strings:
option_strings_dict[option] = a.dest
if '_StoreTrueAction' in str(type(a)):
store_true.append(option)
elif '_StoreFalseAction' in str(type(a)):
store_false.append(option)
if args_that_override is None:
args_that_override = _sys.argv[1:]
for i in range(len(args_that_override)):
if args_that_override[i] in option_strings_dict:
if args_that_override[i] in store_true:
self.overridable[option_strings_dict[
args_that_override[i]]] = True
elif args_that_override[i] in store_false:
self.overridable[option_strings_dict[
args_that_override[i]]] = False
elif (i < len(args_that_override) - 1 and
args_that_override[i + 1] not in option_strings_dict):
key = option_strings_dict[args_that_override[i]]
self.overridable[key] = self.opt[key]
self.opt['override'] = self.overridable
# load opts if a file is provided.
if self.opt.get('init_opt', None) is not None:
self._load_opts(self.opt)
# map filenames that start with 'zoo:' to point to the model zoo dir
options_to_change = {
'model_file', 'dict_file', 'bpe_vocab', 'bpe_merge'
}
for each_key in options_to_change:
if self.opt.get(each_key) is not None:
self.opt[each_key] = modelzoo_path(self.opt.get('datapath'),
self.opt[each_key])
if self.opt['override'].get(each_key) is not None:
# also check override
self.opt['override'][each_key] = modelzoo_path(
self.opt.get('datapath'), self.opt['override'][each_key])
# add start time of an experiment
self.opt['starttime'] = datetime.datetime.today().strftime(
'%b%d_%H-%M')
def test_save_withignore(self):
o = Opt({'a': 3, 'b': 'foo', 'override': {'a': 3}})
with testing_utils.tempdir() as tmpdir:
fn = os.path.join(tmpdir, "opt")
o.save(fn)
o2 = Opt.load(fn)
assert o != o2
assert 'override' not in o2
def train(epochs = 5):
opt = Opt({'num_epochs' : epochs,'datapath':datapath,'datatype':datatype})
# set up timers
train_time = Timer()
validate_time = Timer()
log_time = Timer()
save_time = Timer()
parleys = 0
def test_iter(self):
opt = Opt({'datatype': 'valid', 'datapath': '/tmp', 'task': 'test'})
teacher = TupleTeacher(opt)
# twice to ensure we reset iterators correctly
examples = list(teacher)
assert len(examples) == 9
examples = list(teacher)
assert len(examples) == 9
def test_no_truncate(self):
with self.assertRaises(ValueError):
testing_utils.train_model(
Opt({
**_DEFAULT_OPTIONS,
**{
'truncate': -1
}
}))
def get_dictionary(PATH: str) -> DictionaryAgent:
"""
读取字典
:param PATH: 字典工具目录
:return 读取的字典
"""
opt = Opt()
dictionary = DictionaryAgent(opt=opt)
dictionary.load(PATH)
return dictionary
def __init__(
self,
opt: Opt,
dpr_model: str = 'bert',
pretrained_path: str = DPR_ZOO_MODEL,
encoder_type: str = 'query',
):
# Override options
try:
config: BertConfig = BertConfig.from_pretrained(
'bert-base-uncased')
except OSError:
config_path = PathManager.get_local_path(
os.path.join(opt['datapath'], "bert_base_uncased",
self.CONFIG_PATH))
config: BertConfig = BertConfig.from_pretrained(config_path)
pretrained_path = modelzoo_path(opt['datapath'],
pretrained_path) # type: ignore
if not os.path.exists(pretrained_path):
# when initializing from parlai rag models, the pretrained path
# may not longer exist. This is fine if we've already trained
# the model.
assert dpr_model == 'bert_from_parlai_rag'
logging.error(f'Pretrained Path does not exist: {pretrained_path}')
pretrained_path = modelzoo_path(opt['datapath'],
DPR_ZOO_MODEL) # type: ignore
dpr_model = 'bert'
logging.error(f'Setting to zoo model: {pretrained_path}')
enc_opt = {
"n_heads": config.num_attention_heads,
"n_layers": config.num_hidden_layers,
"embedding_size": config.hidden_size,
"ffn_size": config.intermediate_size,
"dropout": config.hidden_dropout_prob,
"attention_dropout": config.attention_probs_dropout_prob,
"activation": config.hidden_act,
"variant": 'xlm',
"reduction_type": 'first',
"n_positions": config.max_position_embeddings,
"n_segments": config.type_vocab_size,
}
embedding = torch.nn.Embedding(config.vocab_size,
config.hidden_size,
padding_idx=config.pad_token_id)
super().__init__(
Opt(enc_opt),
vocabulary_size=config.vocab_size,
padding_idx=config.pad_token_id,
embedding=embedding,
reduction_type='first',
)
self._load_state(opt['datapath'], dpr_model, pretrained_path,
encoder_type)
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 _opt(self, **kwargs):
return Opt(
batchsize=4,
optimizer='adam',
n_layers=1,
n_heads=4,
ffn_size=16,
embedding_size=16,
skip_generation=True,
**kwargs,
)
def test_beamsearch_blocking(self):
"""
Test beamsearch blocking.
"""
with testing_utils.tempdir() as tmpdir:
agent = create_agent_from_model_file(
'zoo:unittest/beam_blocking/model')
agent.observe({'text': '5 5 5 5 5 5 5', 'episode_done': True})
assert agent.act()['text'] == '5 5 5 5 5 5 5'
agent = create_agent_from_model_file(
'zoo:unittest/beam_blocking/model', Opt(beam_block_ngram=1))
agent.observe({'text': '5 5 5 5 5 5 5', 'episode_done': True})
assert '5 5' not in agent.act()['text']
agent = create_agent_from_model_file(
'zoo:unittest/beam_blocking/model', Opt(beam_block_ngram=2))
agent.observe({'text': '5 5 5 5 5 5 5', 'episode_done': True})
assert '5 5 5' not in agent.act()['text']
with open(os.path.join(tmpdir, 'blocklist.txt'), 'w') as f:
f.write("38\n62\n34 34\n")
agent = create_agent_from_model_file(
'zoo:unittest/beam_blocking/model',
Opt(beam_block_list_filename=os.path.join(
tmpdir, 'blocklist.txt')),
)
agent.observe({'text': '4 4 4', 'episode_done': True})
assert agent.act()['text'] == '4 4 4'
agent.observe({'text': '38 38 38', 'episode_done': True})
assert '38' not in agent.act()['text']
agent.observe({'text': '62 62 62', 'episode_done': True})
assert '62' not in agent.act()['text']
agent.observe({'text': '34 34 34', 'episode_done': True})
text = agent.act()['text']
assert '34' in text
assert '34 34' not in text
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 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 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 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 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 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',
)