def train_pet_ensemble(model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[int],
output_dir: str,
ipet_data_dir: str = None,
repetitions: int = 3,
train_data: List[InputExample] = None,
unlabeled_data: List[InputExample] = None,
eval_data: List[InputExample] = None,
do_train: bool = True,
do_eval: bool = True,
save_unlabeled_logits: bool = False,
seed: int = 42):
"""
Train and evaluate an ensemble of PET models without knowledge distillation.
:param model_config: the model configuration to use
:param train_config: the training configuration to use
:param eval_config: the evaluation configuration to use
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param ipet_data_dir: optional directory containing additional training data for iPET
:param repetitions: the number of training repetitions
:param train_data: the training examples to use
:param unlabeled_data: the unlabeled examples to use
:param eval_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param save_unlabeled_logits: whether logits for unlabeled examples should be saved in a file ``logits.txt``. This
is required for both iPET and knowledge distillation.
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(list))
set_seed(seed)
for pattern_id in pattern_ids:
for iteration in range(repetitions):
model_config.pattern_id = pattern_id
results_dict = {}
pattern_iter_output_dir = "{}/p{}-i{}".format(
output_dir, pattern_id, iteration)
if os.path.exists(pattern_iter_output_dir):
logger.warning(
f"Path {pattern_iter_output_dir} already exists, skipping it..."
)
continue
if not os.path.exists(pattern_iter_output_dir):
os.makedirs(pattern_iter_output_dir)
wrapper = init_model(model_config)
# Training
if do_train:
if ipet_data_dir:
p = os.path.join(
ipet_data_dir,
'p{}-i{}-train.bin'.format(pattern_id, iteration))
ipet_train_data = InputExample.load_examples(p)
for example in ipet_train_data:
example.logits = None
else:
ipet_train_data = None
results_dict.update(
train_single_model(wrapper,
train_data,
train_config,
eval_config,
ipet_train_data=ipet_train_data,
unlabeled_data=unlabeled_data))
with open(os.path.join(pattern_iter_output_dir, 'results.txt'),
'w') as fh:
fh.write(str(results_dict))
logger.info("Saving trained model at {}...".format(
pattern_iter_output_dir))
wrapper.save(pattern_iter_output_dir)
train_config.save(
os.path.join(pattern_iter_output_dir, 'train_config.json'))
eval_config.save(
os.path.join(pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
if save_unlabeled_logits:
logits = evaluate(wrapper, unlabeled_data,
eval_config)['logits']
save_logits(
os.path.join(pattern_iter_output_dir, 'logits.txt'),
logits)
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
# Evaluation
if do_eval:
logger.info("Starting evaluation...")
if not wrapper:
wrapper = TransformerModelWrapper.from_pretrained(
pattern_iter_output_dir)
eval_result = evaluate(wrapper,
eval_data,
eval_config,
priming_data=train_data)
save_predictions(
os.path.join(pattern_iter_output_dir, 'predictions.jsonl'),
wrapper, eval_result)
save_logits(
os.path.join(pattern_iter_output_dir, 'eval_logits.txt'),
eval_result['logits'])
scores = eval_result['scores']
logger.info(
"--- RESULT (pattern_id={}, iteration={}) ---".format(
pattern_id, iteration))
logger.info(scores)
results_dict['test_set_after_training'] = scores
with open(
os.path.join(pattern_iter_output_dir, 'results.json'),
'w') as fh:
json.dump(results_dict, fh)
for metric, value in scores.items():
results[metric][pattern_id].append(value)
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
if do_eval:
logger.info("=== OVERALL RESULTS ===")
_write_results(os.path.join(output_dir, 'result_test.txt'), results)
else:
logger.info("=== ENSEMBLE TRAINING COMPLETE ===")
def train_pet_ensemble(
model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[Union[str, int]],
output_dir: str,
ipet_data_dir: str = None,
repetitions: int = 3,
train_data: List[InputExample] = None,
unlabeled_data: List[InputExample] = None,
dev_data: List[InputExample] = None,
test_data: List[InputExample] = None,
do_train: bool = True,
do_eval: bool = True,
save_unlabeled_logits: bool = False,
seed: int = 42,
overwrite_dir: bool = False,
save_model=False,
local_rank=-1,
):
"""
Train and evaluate an ensemble of PET models without knowledge distillation.
:param model_config: the model configuration to use
:param train_config: the training configuration to use
:param eval_config: the evaluation configuration to use
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param ipet_data_dir: optional directory containing additional training data for iPET
:param repetitions: the number of training repetitions
:param train_data: the training examples to use
:param unlabeled_data: the unlabeled examples to use
:param dev_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param save_unlabeled_logits: whether logits for unlabeled examples should be saved in a file ``logits.txt``. This
is required for both iPET and knowledge distillation.
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(lambda: defaultdict(list)))
set_seed(seed)
for pattern_id in pattern_ids:
for iteration in range(repetitions):
model_config.pattern_id = pattern_id
results_dict = {}
shots = 0 if train_data is None else len(train_data)
pattern_iter_output_dir = "{}/{}shots-{}-i{}-seed{}".format(
output_dir, shots, pattern_name(pattern_id), iteration, seed)
if os.path.exists(pattern_iter_output_dir) and not overwrite_dir:
logger.warning(
f"Path {pattern_iter_output_dir} already exists, skipping it..."
)
continue
if not os.path.exists(pattern_iter_output_dir) and local_rank in [
-1, 0
]:
os.makedirs(pattern_iter_output_dir)
wrapper = init_model(model_config)
# Training
if do_train:
if ipet_data_dir:
p = os.path.join(
ipet_data_dir,
"{}-i{}-train.bin".format(pattern_name(pattern_id),
iteration))
ipet_train_data = InputExample.load_examples(p)
for example in ipet_train_data:
example.logits = None
else:
ipet_train_data = None
results_dict.update(
train_single_model(
wrapper,
train_data,
train_config,
pattern_iter_output_dir,
dev_data,
eval_config,
ipet_train_data=ipet_train_data,
unlabeled_data=unlabeled_data,
return_train_set_results=False,
local_rank=local_rank,
))
with open(os.path.join(pattern_iter_output_dir, "results.txt"),
"w") as fh:
fh.write(str(results_dict))
if local_rank in [-1, 0]:
logger.info("Saving trained model at {}...".format(
pattern_iter_output_dir))
train_config.save(
os.path.join(pattern_iter_output_dir,
"train_config.json"))
eval_config.save(
os.path.join(pattern_iter_output_dir,
"eval_config.json"))
logger.info("Saving complete")
if save_unlabeled_logits:
logits = evaluate(wrapper,
unlabeled_data,
eval_config,
local_rank=local_rank)["logits"]
save_logits(
os.path.join(pattern_iter_output_dir,
"logits.txt"), logits)
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
# Evaluation
if do_eval:
logger.info("Starting evaluation...")
try:
wrapper = TransformerModelWrapper.from_pretrained(
pattern_iter_output_dir)
except OSError:
warnings.warn(
"No model found saved, proceeding with current model instead of best"
)
pass
for split, eval_data in {
"dev": dev_data,
"test": test_data
}.items():
if eval_data is None:
continue
eval_result = evaluate(wrapper,
eval_data,
eval_config,
priming_data=train_data,
local_rank=local_rank)
if local_rank in [-1, 0]:
save_predictions(
os.path.join(pattern_iter_output_dir,
"predictions.jsonl"), wrapper,
eval_result)
save_logits(
os.path.join(pattern_iter_output_dir,
"eval_logits.txt"),
eval_result["logits"])
scores = eval_result["scores"]
logger.info(
"--- {} result (pattern_id={}, iteration={}) ---".
format(split, pattern_id, iteration))
logger.info(scores)
results_dict[f"{split}_set_after_training"] = scores
with open(
os.path.join(pattern_iter_output_dir,
"results.json"), "w") as fh:
json.dump(results_dict, fh)
for metric, value in scores.items():
results[split][metric][pattern_id].append(value)
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
if do_eval:
logger.info("=== OVERALL RESULTS ===")
results_to_log = _write_results(
os.path.join(output_dir, "result_test.txt"), results)
else:
logger.info("=== ENSEMBLE TRAINING COMPLETE ===")
results_to_log = None
if do_train and not save_model:
outputs = os.listdir(pattern_iter_output_dir)
for item in outputs:
if item.endswith(".bin"):
os.remove(os.path.join(pattern_iter_output_dir, item))
return results_to_log
def train_pet(train_data: List[InputExample],
eval_data: List[InputExample],
dev32_data: List[InputExample],
model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[int],
output_dir: str,
repetitions: int = 3,
do_train: bool = True,
do_eval: bool = True,
seed: int = 42
):
"""
Train and evaluate a new PET model for a given task.
:param model_config: the model configuration for each model corresponding to an individual PVP
:param train_config: the training configuration for each model corresponding to an individual PVP
:param eval_config: the evaluation configuration for each model corresponding to an individual PVP
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param repetitions: the number of training repetitions for each model corresponding to an individual PVP
:param train_data: the training examples to use
:param dev32_data: the dev32 examples to use
:param eval_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(list))
dev32_results = defaultdict(lambda: defaultdict(list))
# set_seed(seed)
assert model_config.task_type == "single_task"
for pattern_id in pattern_ids: # pattern只有1个
model_config.pattern_id = pattern_id
results_dict = {}
pattern_iter_output_dir = "{}/p{}-i{}".format(output_dir, pattern_id, 1)
# if os.path.exists(pattern_iter_output_dir):
# logger.warning(f"Path {pattern_iter_output_dir} already exists, skipping it...")
# continue
if not os.path.exists(pattern_iter_output_dir):
os.makedirs(pattern_iter_output_dir)
wrapper = init_model(model_config) # 初始化一个模型
# Training
if do_train:
# 开始多轮epoch训练，并将训练的结果保存到results_dict中
results_dict.update(train_single_model(train_data, eval_data, dev32_data, pattern_iter_output_dir, \
wrapper, train_config, eval_config))
with open(os.path.join(pattern_iter_output_dir, 'results.txt'), 'w') as fh:
fh.write(str(results_dict))
train_config.save(os.path.join(pattern_iter_output_dir, 'train_config.json'))
eval_config.save(os.path.join(pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
# Evaluation
if do_eval:
logger.info("Starting evaluation...")
logger.info("Single: Task {} 's Test examples number: {}".format(model_config.task_name, len(eval_data)))
logger.info("************Test Example:**************")
logger.info("text_a={}".format(eval_data[0].text_a))
logger.info("text_b={}".format(eval_data[0].text_b))
logger.info("task={}".format(eval_data[0].task))
logger.info("label={}".format(eval_data[0].label))
logger.info("**********************************")
# if not wrapper:
wrapper = TransformerModelWrapper.from_pretrained(pattern_iter_output_dir)
eval_result = evaluate(wrapper, eval_data, eval_config)
# dev32_result = evaluate(wrapper, dev32_data, eval_config)
save_predictions(os.path.join(pattern_iter_output_dir, 'eval_predictions.jsonl'), wrapper, eval_result)
save_logits(os.path.join(pattern_iter_output_dir, 'eval_logits.txt'), eval_result['logits'])
# save_predictions(os.path.join(pattern_iter_output_dir, 'dev32_predictions.jsonl'), wrapper, dev32_result)
# save_logits(os.path.join(pattern_iter_output_dir, 'dev32_logits.txt'), dev32_result['logits'])
logger.info("--- RESULT (pattern_id={}, Task={}) ---".format(pattern_id, model_config.task_name))
logger.info("eval_results:")
logger.info(eval_result['scores'])
# logger.info("dev32_results:")
# logger.info(dev32_result['scores'])
# results_dict['eval_set_after_training'] = eval_result['scores']
# results_dict['dev32_set_after_training'] = dev32_result['scores']
# with open(os.path.join(pattern_iter_output_dir, 'results.json'), 'w') as fh:
# json.dump(results_dict, fh)
#
# for metric, value in eval_result['scores'].items():
# results[metric][pattern_id].append(value)
#
# for metric, value in dev32_result['scores'].items():
# dev32_results[metric][pattern_id].append(value)
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
def train_generalization_cross(unseen_task_train_data: List[InputExample],
unseen_task_dev_data: List[InputExample],
seen_task_train_data: List[InputExample],
seen_task_dev_data: List[InputExample],
# dev32_data: List[InputExample],
unseen_task: str,
model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[int],
output_dir: str,
repetitions: int = 3,
do_train: bool = True,
do_eval: bool = True,
seed: int = 42
):
"""
Train and evaluate a new PET model for a given task.
:param model_config: the model configuration for each model corresponding to an individual PVP
:param train_config: the training configuration for each model corresponding to an individual PVP
:param eval_config: the evaluation configuration for each model corresponding to an individual PVP
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param repetitions: the number of training repetitions for each model corresponding to an individual PVP
:param train_data: the training examples to use
:param dev32_data: the dev32 examples to use
:param eval_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(list))
dev32_results = defaultdict(lambda: defaultdict(list))
set_seed(seed)
assert model_config.task_type == "cross_task"
for pattern_id in pattern_ids: # 只选择1个模式
model_config.pattern_id = pattern_id
results_dict = {}
pattern_iter_output_dir = "{}/{}/generalization/{}".format(output_dir, model_config.scene, model_config.task_name)
if not os.path.exists(pattern_iter_output_dir):
os.makedirs(pattern_iter_output_dir)
# wrapper = TransPromptModelWrapper(model_config) # 初始化一个TransPrompt模型
wrapper = TransPromptModelWrapper2(model_config) # 初始化一个TransPrompt模型
# wrapper = TransformerModelWrapper(model_config)
# Multi-Task Meta-Learning Training
if do_train:
logger.info("========= Stage1: Starting Fine-tuning Multi-Task Meta-Learner ... =========")
# 开始多轮epoch训练，并将训练的结果保存到results_dict中
# edit by wjn : eval_data -> None
results_dict.update(train_single_model(seen_task_train_data, None, seen_task_dev_data, pattern_iter_output_dir, \
wrapper, train_config, eval_config, use_debias=True))
train_config.save(os.path.join(pattern_iter_output_dir, 'train_config.json'))
eval_config.save(os.path.join(pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
logger.info("========= Stage1: Finish Fine-tuning Multi-Task Meta-Learner =========")
# Task Adaptation Fine-tune
if do_eval:
logger.info("========= Stage2: Starting Task Generalization (Unseen Task-Specific Fine-tuning) ... =========")
# 重新加载训练好的meta-learner
# wrapper = TransPromptModelWrapper.from_pretrained(pattern_iter_output_dir)
wrapper = TransPromptModelWrapper2.from_pretrained(pattern_iter_output_dir)
cross_data_dir = "data/k-shot-cross/"
# add by wjn
## 当前是task generalization，对每个group的每个task:
# 在训练好的meta learner基础上，在当前task对应的训练集上再次task specific fine-tune；并在验证集上选择模型
# 最后在对应的测试集上测试；
# group内的每个task在此阶段独立地进行
### task-specific fine-tune
logger.info("========= Stage2: Specific fine-tuning on Unseen Task {} =========".format(unseen_task))
# wrapper.config.task_name = task_name # 在task-specific微调时，更改当前微调的task名称
train_config.max_steps = eval_config.max_steps # 在task-specific微调时，更改max_steps
train_config.per_gpu_train_batch_size = eval_config.per_gpu_eval_batch_size # 更改batch_size
# 在meta-learner基础上对unseen task继续做task-specific微调，并保存
train_single_model(unseen_task_train_data, None,
unseen_task_dev_data, pattern_iter_output_dir + '/' + unseen_task, \
wrapper, train_config, eval_config, use_debias=False)
# 将task-specific微调后保存的模型加载进来
# task_specific_wrapper = TransPromptModelWrapper.from_pretrained(pattern_iter_output_dir + '/' + unseen_task)
task_specific_wrapper = TransPromptModelWrapper2.from_pretrained(pattern_iter_output_dir + '/' + unseen_task)
logger.info("========= Stage2: Evaluating test set on Task {}".format(unseen_task))
### evaluate on test dataset
eval_data = load_examples(
unseen_task, cross_data_dir + data_to_name[unseen_task] + "/" + str(model_config.k) + "-" + str(seed),
TEST_SET, num_examples=-1, num_examples_per_label=None)
logger.info("Group {}: Task {} 's Test examples number: {}".format(model_config.task_name, unseen_task, len(eval_data)))
# logger.info("************Test Example:**************")
# logger.info("text_a={}".format(eval_data[0].text_a))
# logger.info("text_b={}".format(eval_data[0].text_b))
# logger.info("task={}".format(eval_data[0].task))
# logger.info("label={}".format(eval_data[0].label))
# logger.info("**********************************")
# 更新当前group task的metrics：
eval_config.metrics = METRICS.get(unseen_task, DEFAULT_METRICS) # cross-task group 的 metrics
eval_result = evaluate(task_specific_wrapper, eval_data, eval_config)
save_predictions(os.path.join(pattern_iter_output_dir + '/' + unseen_task, 'eval_predictions.jsonl'), task_specific_wrapper, eval_result)
save_logits(os.path.join(pattern_iter_output_dir + '/' + unseen_task, 'eval_logits.txt'), eval_result['logits'])
# save_predictions(os.path.join(pattern_iter_output_dir, 'dev32_predictions.jsonl'), wrapper, dev32_result)
# save_logits(os.path.join(pattern_iter_output_dir, 'dev32_logits.txt'), dev32_result['logits'])
logger.info("--- Unseen Task Generalization Result (pattern_id={}, Group={}, Task={}) ---".format(pattern_id, model_config.task_name, unseen_task))
logger.info("eval_results: {}".format(eval_result['scores']))
task_specific_wrapper.model = None
task_specific_wrapper = None
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
def train_adaptation_cross(train_data: List[InputExample],
# eval_data: List[InputExample],
dev32_data: List[InputExample],
model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[int],
output_dir: str,
repetitions: int = 3,
do_train: bool = True,
do_eval: bool = True,
seed: int = 42
):
"""
Train and evaluate a new PET model for a given task.
:param model_config: the model configuration for each model corresponding to an individual PVP
:param train_config: the training configuration for each model corresponding to an individual PVP
:param eval_config: the evaluation configuration for each model corresponding to an individual PVP
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param repetitions: the number of training repetitions for each model corresponding to an individual PVP
:param train_data: the training examples to use
:param dev32_data: the dev32 examples to use
:param eval_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(list))
dev32_results = defaultdict(lambda: defaultdict(list))
set_seed(seed)
assert model_config.task_type == "cross_task"
# 当前是cross-task，则task_name是group的名称，需要获得group内的所有task
tasks = groups[model_config.task_name]
for pattern_id in pattern_ids: # 只选择1个模式
model_config.pattern_id = pattern_id
results_dict = {}
pattern_iter_output_dir = "{}/{}/adaptation/{}".format(output_dir, model_config.scene, model_config.task_name)
if not os.path.exists(pattern_iter_output_dir):
os.makedirs(pattern_iter_output_dir)
# wrapper = TransPromptModelWrapper(model_config) # 初始化一个TransPrompt模型
wrapper = TransPromptModelWrapper2(model_config) # 初始化一个TransPrompt模型
# wrapper = TransformerModelWrapper(model_config)
# Multi-Task Meta-Learning Training
if do_train:
logger.info("========= Stage1: Starting Fine-tuning Multi-Task Meta-Learner ... =========")
# 开始多轮epoch训练，并将训练的结果保存到results_dict中
# edit by wjn : eval_data -> None
results_dict.update(train_single_model(train_data, None, dev32_data, pattern_iter_output_dir, \
wrapper, train_config, eval_config, use_debias=True))
train_config.save(os.path.join(pattern_iter_output_dir, 'train_config.json'))
eval_config.save(os.path.join(pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
logger.info("========= Stage1: Finish Fine-tuning Multi-Task Meta-Learner =========")
# Task Adaptation Fine-tune
if do_eval:
logger.info("========= Stage2: Starting Task Adaptation (Task-Specific Fine-tuning) ... =========")
# 用于保存每次试验跑出的结果
# 加载先前的结果
t = time.time()
ada_res_acc = dict()
if os.path.exists('ada_res_acc.npy'):
ada_res_acc = np.load('ada_res_acc.npy', allow_pickle=True)[()] # dict {time: {task:acc, ...}}
accs = dict()
# 重新加载训练好的meta-learner
# wrapper = TransPromptModelWrapper.from_pretrained(pattern_iter_output_dir)
wrapper = TransPromptModelWrapper2.from_pretrained(pattern_iter_output_dir)
cross_data_dir = "data/k-shot-cross/"
# add by wjn
## 当前是task adaptation，对每个group的每个task:
# 在训练好的meta learner基础上，在当前task对应的训练集上再次task specific fine-tune；并在验证集上选择模型
# 最后在对应的测试集上测试；
# group内的每个task在此阶段独立地进行
# 获得cross-task上每个task对应的训练集和测试集
task_to_train_example, task_to_dev_example = dict(), dict() # {task_name: [.., ..], ..}
task_to_train_example = load_examples(
model_config.task_name, None, SPE_TRAIN_SET, num_examples=-1, num_examples_per_label=None, examples=train_data)
task_to_dev_example = load_examples(
model_config.task_name, None, SPE_DEV_SET, num_examples=-1, num_examples_per_label=None, examples=dev32_data)
for ei, task_name in enumerate(tasks):
### task-specific fine-tune
logger.info("========= Stage2.{}: Specific fine-tuning on Task {} =========".format(ei + 1, task_name))
# wrapper.config.task_name = task_name # 在task-specific微调时，更改当前微调的task名称
train_config.max_steps = eval_config.max_steps # 在task-specific微调时，更改max_steps
train_config.per_gpu_train_batch_size = eval_config.per_gpu_eval_batch_size # 更改batch_size
if task_name == 'mrpc': # group3内的两个task（MRPC和QQP）训练集样本数量悬殊过大，直接指定MRPC只有1200steps
train_config.max_steps = 4800
train_config.per_gpu_train_batch_size = 16
eval_config.per_gpu_eval_batch_size = 8
# 在meta-learner基础上继续做task-specific微调，并保存
train_single_model(task_to_train_example[data_to_name[task_name]], None,
task_to_dev_example[data_to_name[task_name]], pattern_iter_output_dir + '/' + task_name, \
wrapper, train_config, eval_config, use_debias=False)
# 将task-specific微调后保存的模型加载进来
# task_specific_wrapper = TransPromptModelWrapper.from_pretrained(pattern_iter_output_dir + '/' + task_name)
task_specific_wrapper = TransPromptModelWrapper2.from_pretrained(pattern_iter_output_dir + '/' + task_name)
logger.info("========= Stage2.{}: Evaluating test set on Task {}".format(ei + 1, task_name))
### evaluate on test dataset
eval_data = load_examples(
task_name, cross_data_dir + data_to_name[task_name] + "/" + str(model_config.k) + "-" + str(seed),
TEST_SET, num_examples=-1, num_examples_per_label=None)
logger.info("Group {}: Task {} 's Test examples number: {}".format(model_config.task_name, task_name, len(eval_data)))
# logger.info("************Test Example:**************")
# logger.info("text_a={}".format(eval_data[0].text_a))
# logger.info("text_b={}".format(eval_data[0].text_b))
# logger.info("task={}".format(eval_data[0].task))
# logger.info("label={}".format(eval_data[0].label))
# logger.info("**********************************")
# 更新当前group task的metrics：
eval_config.metrics = METRICS.get(task_name, DEFAULT_METRICS) # cross-task group 的 metrics
eval_result = evaluate(task_specific_wrapper, eval_data, eval_config)
save_predictions(os.path.join(pattern_iter_output_dir + '/' + task_name, 'eval_predictions.jsonl'), task_specific_wrapper, eval_result)
save_logits(os.path.join(pattern_iter_output_dir + '/' + task_name, 'eval_logits.txt'), eval_result['logits'])
# save_predictions(os.path.join(pattern_iter_output_dir, 'dev32_predictions.jsonl'), wrapper, dev32_result)
# save_logits(os.path.join(pattern_iter_output_dir, 'dev32_logits.txt'), dev32_result['logits'])
logger.info("--- Task Adaptation Result (pattern_id={}, Group={}, Task={}) ---".format(pattern_id, model_config.task_name, task_name))
logger.info("eval_results: {}".format(eval_result['scores']))
accs[task_name] = eval_result['scores']
task_specific_wrapper.model = None
task_specific_wrapper = None
ada_res_acc[t] = accs
np.save('ada_res_acc.npy', ada_res_acc)
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
def train_pet_cross(train_data: List[InputExample],
# eval_data: List[InputExample],
dev32_data: List[InputExample],
model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[int],
output_dir: str,
repetitions: int = 3,
do_train: bool = True,
do_eval: bool = True,
seed: int = 42
):
"""
Train and evaluate a new PET model for a given task.
:param model_config: the model configuration for each model corresponding to an individual PVP
:param train_config: the training configuration for each model corresponding to an individual PVP
:param eval_config: the evaluation configuration for each model corresponding to an individual PVP
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param repetitions: the number of training repetitions for each model corresponding to an individual PVP
:param train_data: the training examples to use
:param dev32_data: the dev32 examples to use
:param eval_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(list))
dev32_results = defaultdict(lambda: defaultdict(list))
# set_seed(seed)
assert model_config.task_type == "cross_task"
# 当前是cross-task，则task_name是group的名称，需要获得group内的所有task
tasks = groups[model_config.task_name]
for pattern_id in pattern_ids: # 只选择1个模式
model_config.pattern_id = pattern_id
results_dict = {}
pattern_iter_output_dir = "{}/p{}-i{}".format(output_dir, pattern_id, 1)
# if os.path.exists(pattern_iter_output_dir):
# logger.warning(f"Path {pattern_iter_output_dir} already exists, skipping it...")
# continue
if not os.path.exists(pattern_iter_output_dir):
os.makedirs(pattern_iter_output_dir)
wrapper = init_model(model_config) # 初始化一个模型
# Training
if do_train:
# 开始多轮epoch训练，并将训练的结果保存到results_dict中
# edit by wjn : eval_data -> None
results_dict.update(train_single_model(train_data, None, dev32_data, pattern_iter_output_dir, \
wrapper, train_config, eval_config))
with open(os.path.join(pattern_iter_output_dir, 'results.txt'), 'w') as fh:
fh.write(str(results_dict))
train_config.save(os.path.join(pattern_iter_output_dir, 'train_config.json'))
eval_config.save(os.path.join(pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
# Evaluation
if do_eval:
logger.info("Starting evaluation...")
# if not wrapper:
wrapper = TransformerModelWrapper.from_pretrained(pattern_iter_output_dir)
cross_data_dir = "data/k-shot-cross/"
# add by wjn
## 当前是cross-task，对当前group内的所有task，分别进行测试
for task_name in tasks:
eval_data = load_examples(
task_name, cross_data_dir + data_to_name[task_name] + "/" + str(model_config.k) + "-" + str(seed),
TEST_SET, num_examples=-1, num_examples_per_label=None)
logger.info("Group {}: Task {} 's Test examples number: {}".format(model_config.task_name, task_name, len(eval_data)))
logger.info("************Test Example:**************")
logger.info("text_a={}".format(eval_data[0].text_a))
logger.info("text_b={}".format(eval_data[0].text_b))
logger.info("task={}".format(eval_data[0].task))
logger.info("label={}".format(eval_data[0].label))
logger.info("**********************************")
# 更新当前group task的metrics：
eval_config.metrics = METRICS.get(task_name, DEFAULT_METRICS) # cross-task group 的 metrics
eval_result = evaluate(wrapper, eval_data, eval_config)
# dev32_result = evaluate(wrapper, dev32_data, eval_config)
save_predictions(os.path.join(pattern_iter_output_dir, 'eval_predictions.jsonl'), wrapper, eval_result)
save_logits(os.path.join(pattern_iter_output_dir, 'eval_logits.txt'), eval_result['logits'])
# save_predictions(os.path.join(pattern_iter_output_dir, 'dev32_predictions.jsonl'), wrapper, dev32_result)
# save_logits(os.path.join(pattern_iter_output_dir, 'dev32_logits.txt'), dev32_result['logits'])
logger.info("--- RESULT (pattern_id={}, Group={}, Task={}) ---".format(pattern_id, model_config.task_name, task_name))
logger.info("eval_results:")
logger.info(eval_result['scores'])
# logger.info("dev32_results:")
# logger.info(dev32_result['scores'])
# results_dict['eval_set_after_training'] = eval_result['scores']
# # results_dict['dev32_set_after_training'] = dev32_result['scores']
# with open(os.path.join(pattern_iter_output_dir, 'results.json'), 'w') as fh:
# json.dump(results_dict, fh)
#
# for metric, value in eval_result['scores'].items():
# results[metric][pattern_id].append(value)
#
# for metric, value in dev32_result['scores'].items():
# dev32_results[metric][pattern_id].append(value)
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
def train_pet(train_data: List[InputExample],
eval_data: List[InputExample],
dev32_data: List[InputExample],
model_config: WrapperConfig,
train_config: TrainConfig,
eval_config: EvalConfig,
pattern_ids: List[int],
output_dir: str,
repetitions: int = 3,
do_train: bool = True,
do_eval: bool = True,
seed: int = 42
):
"""
Train and evaluate a new PET model for a given task.
:param model_config: the model configuration for each model corresponding to an individual PVP
:param train_config: the training configuration for each model corresponding to an individual PVP
:param eval_config: the evaluation configuration for each model corresponding to an individual PVP
:param pattern_ids: the ids of all PVPs to use
:param output_dir: the output directory
:param repetitions: the number of training repetitions for each model corresponding to an individual PVP
:param train_data: the training examples to use
:param dev32_data: the dev32 examples to use
:param eval_data: the evaluation examples to use
:param do_train: whether to perform training
:param do_eval: whether to perform evaluation
:param seed: the random seed to use
"""
results = defaultdict(lambda: defaultdict(list))
dev32_results = defaultdict(lambda: defaultdict(list))
set_seed(seed)
for pattern_id in pattern_ids:
for iteration in range(repetitions):
model_config.pattern_id = pattern_id
results_dict = {}
pattern_iter_output_dir = "{}/p{}-i{}".format(output_dir, pattern_id, iteration)
if os.path.exists(pattern_iter_output_dir):
logger.warning(f"Path {pattern_iter_output_dir} already exists, skipping it...")
continue
if not os.path.exists(pattern_iter_output_dir):
os.makedirs(pattern_iter_output_dir)
wrapper = init_model(model_config)
# Training
if do_train:
results_dict.update(train_single_model(train_data, eval_data, dev32_data, pattern_iter_output_dir, \
wrapper, train_config, eval_config))
with open(os.path.join(pattern_iter_output_dir, 'results.txt'), 'w') as fh:
fh.write(str(results_dict))
train_config.save(os.path.join(pattern_iter_output_dir, 'train_config.json'))
eval_config.save(os.path.join(pattern_iter_output_dir, 'eval_config.json'))
logger.info("Saving complete")
if not do_eval:
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
# Evaluation
if do_eval:
logger.info("Starting evaluation...")
# if not wrapper:
wrapper = TransformerModelWrapper.from_pretrained(pattern_iter_output_dir)
eval_result = evaluate(wrapper, eval_data, eval_config)
dev32_result = evaluate(wrapper, dev32_data, eval_config)
save_predictions(os.path.join(pattern_iter_output_dir, 'eval_predictions.jsonl'), wrapper, eval_result)
save_logits(os.path.join(pattern_iter_output_dir, 'eval_logits.txt'), eval_result['logits'])
save_predictions(os.path.join(pattern_iter_output_dir, 'dev32_predictions.jsonl'), wrapper,
dev32_result)
save_logits(os.path.join(pattern_iter_output_dir, 'dev32_logits.txt'), dev32_result['logits'])
logger.info("--- RESULT (pattern_id={}, iteration={}) ---".format(pattern_id, iteration))
logger.info("eval_results:")
logger.info(eval_result['scores'])
logger.info("dev32_results:")
logger.info(dev32_result['scores'])
results_dict['eval_set_after_training'] = eval_result['scores']
results_dict['dev32_set_after_training'] = dev32_result['scores']
with open(os.path.join(pattern_iter_output_dir, 'results.json'), 'w') as fh:
json.dump(results_dict, fh)
for metric, value in eval_result['scores'].items():
results[metric][pattern_id].append(value)
for metric, value in dev32_result['scores'].items():
dev32_results[metric][pattern_id].append(value)
wrapper.model = None
wrapper = None
torch.cuda.empty_cache()
if do_eval:
logger.info("=== OVERALL RESULTS ===")
_write_results(os.path.join(output_dir, 'result_test.txt'), results, dev32_results)
else:
logger.info("=== ENSEMBLE TRAINING COMPLETE ===")
