def prepare_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
list_passage = batch["input"]["passage"]
list_question = batch["input"]["question"]
list_true_entity = batch["input"]["true_entity"]
list_false_entities = batch["input"]["false_entities"]
list_lbl = batch["output"]["lbl"]
bs = len(batch["input"]["passage"])
prep_lbl = np.random.randint(self.num_lbl, size=bs)
tgt = torch.from_numpy(prep_lbl).long() == batch["output"]["lbl"]
list_orig_input_ids = []
list_masked_input_ids = []
for b_idx, (p, q, te, fe, lbl) in enumerate(
zip(list_passage, list_question, list_true_entity,
list_false_entities, list_lbl)):
txt_split_tuple = []
true_num_lbl_tok = self.get_lbl_num_lbl_tok(te)
max_num_lbl_tok = true_num_lbl_tok
for idx, wrong_enty in enumerate(fe):
num_lbl_tok = self.get_lbl_num_lbl_tok(wrong_enty)
if num_lbl_tok > max_num_lbl_tok:
max_num_lbl_tok = num_lbl_tok
txt_trim = -1
pattern = self.pet_patterns[self._pet_names.index(mode)]
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace("[PASSAGE]", p).replace(
"[QUESTION]", q + " [SEP]").replace("@highlight", "-")
txt_split_tuple.append(txt_split_inp)
# Trim the paragraph
if "[PASSAGE]" in txt_split:
txt_trim = idx
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(
self.tokenizer, self.config, txt_split_tuple[0],
txt_split_tuple[1], txt_split_tuple[2], txt_trim)
list_orig_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_orig_input_ids).to(device), torch.tensor(
list_masked_input_ids).to(device), prep_lbl, tgt.to(device)
def prepare_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
list_question = batch["input"]["question"]
list_passage = batch["input"]["passage"]
list_answer = batch["input"]["answer"]
bs = len(batch["input"]["answer"])
prep_lbl = np.random.randint(self.num_lbl, size=bs)
tgt = torch.from_numpy(prep_lbl).long() == batch["output"]["lbl"]
pattern, label = self.pet_pvps[self._pet_names.index(mode)]
list_orig_input_ids = []
list_masked_input_ids = []
for b_idx, (p, q, a, lbl) in enumerate(
zip(list_passage, list_question, list_answer, prep_lbl)):
txt_split_tuple = []
txt_trim = -1
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace("[PARAGRAPH]", p).replace(
"[QUESTION]", q).replace("[ANSWER]",
a).replace("[MASK]", label[lbl])
txt_split_tuple.append(txt_split_inp)
# Trim the paragraph
if "[PARAGRAPH]" in txt_split:
txt_trim = idx
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(
self.tokenizer, self.config, txt_split_tuple[0],
txt_split_tuple[1], txt_split_tuple[2], txt_trim)
list_orig_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_orig_input_ids).to(device), torch.tensor(
list_masked_input_ids).to(device), prep_lbl, tgt.to(device)
def prepare_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
list_hypothesis = batch["input"]["hypothesis"]
list_premise = batch["input"]["premise"]
bs = len(batch["input"]["hypothesis"])
prep_lbl = np.random.randint(self.num_lbl, size=bs)
tgt = torch.from_numpy(prep_lbl).long() == batch["output"]["lbl"]
pattern, label = self.pet_pvps[self._pet_names.index(mode)]
list_orig_input_ids = []
list_masked_input_ids = []
for b_idx, (h, p, lbl) in enumerate(
zip(list_hypothesis, list_premise, prep_lbl)):
txt_split_tuple = []
txt_trim = -1
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace("[HYPOTHESIS]", h).replace(
"[PREMISE]", p).replace("[MASK]", label[lbl])
txt_split_tuple.append(txt_split_inp)
# Trim the paragraph
if "[PREMISE]" in txt_split:
txt_trim = idx
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(
self.tokenizer, self.config, txt_split_tuple[0],
txt_split_tuple[1], txt_split_tuple[2], txt_trim)
list_orig_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_orig_input_ids).to(device), torch.tensor(
list_masked_input_ids).to(device), prep_lbl, tgt.to(device)
def prepare_eval_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
list_hypothesis = batch["input"]["hypothesis"]
list_premise = batch["input"]["premise"]
list_input_ids = []
list_masked_input_ids = []
pattern, label = self.pet_pvps[self._pet_names.index(mode)]
for b_idx, (h, p) in enumerate(zip(list_hypothesis, list_premise)):
mask_idx = None
for l_idx, lbl in enumerate(label):
txt_split_tuple = []
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace(
"[HYPOTHESIS]", h).replace("[PREMISE]",
p).replace("[MASK]", lbl)
txt_split_tuple.append(txt_split_inp)
# Trim the paragraph
if "[PREMISE]" in txt_split:
txt_trim = idx
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(
self.tokenizer,
self.config,
txt_split_tuple[0],
txt_split_tuple[1],
txt_split_tuple[2],
txt_trim,
mask_idx=mask_idx)
list_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_input_ids).to(device), torch.tensor(
list_masked_input_ids).to(device)
def prepare_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
list_sentence1 = batch["input"]["sentence1"]
list_sentence2 = batch["input"]["sentence2"]
list_word = batch["input"]["word"]
bs = len(batch["input"]["sentence1"])
prep_lbl = np.random.randint(self.num_lbl, size=bs)
tgt = torch.from_numpy(prep_lbl).long() == batch["output"]["lbl"]
pattern, label = self.pet_pvps[self._pet_names.index(mode)]
list_orig_input_ids = []
list_masked_input_ids = []
for b_idx, (s1, s2, w, lbl) in enumerate(zip(list_sentence1, list_sentence2, list_word, prep_lbl)):
txt_split_tuple = []
txt_trim = -1
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace("[SENTENCE1]", s1).replace("[SENTENCE2]", s2).replace("[WORD]", w).replace("[MASK]",
label[lbl])
txt_split_tuple.append(txt_split_inp)
# Trim the paragraph
if "[SENTENCE1]" in txt_split:
txt_trim = idx
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(self.tokenizer, self.config, txt_split_tuple[0], txt_split_tuple[1], txt_split_tuple[2], txt_trim)
list_orig_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_orig_input_ids).to(device), torch.tensor(list_masked_input_ids).to(device), prep_lbl, tgt.to(device)
def prepare_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
list_text = batch["input"]["text"]
list_pronoun = batch["input"]["pronoun"]
list_noun = batch["input"]["noun"]
list_lbl = batch["output"]["lbl"]
list_orig_input_ids = []
list_masked_input_ids = []
tgt = torch.tensor([1.]).long()
for b_idx, (t, p, n, lbl) in enumerate(
zip(list_text, list_pronoun, list_noun, list_lbl)):
txt_trim = -1
pattern = self.pet_patterns[self._pet_names.index(mode)]
txt_split_tuple = []
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace("[TEXT]", t).replace(
"[NNP]", p).replace("[MASK]", n)
txt_split_tuple.append(txt_split_inp)
if "[TEXT]" in txt_split:
txt_trim = idx
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(
self.tokenizer, self.config, txt_split_tuple[0],
txt_split_tuple[1], txt_split_tuple[2], txt_trim)
list_orig_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_orig_input_ids).to(device), torch.tensor(
list_masked_input_ids).to(device), None, tgt.to(device)
def prepare_pet_mlm_batch(self, batch, mode="PET1"):
'''
Prepare for train
:param batch:
:return:
'''
# Always use pattern 3 for COPA
mode = "PET3"
list_premise = batch["input"]["premise"]
list_choice1 = batch["input"]["choice1"]
list_choice2 = batch["input"]["choice2"]
list_question = batch["input"]["question"]
list_lbl = batch["output"]["lbl"]
bs = len(batch["input"]["question"])
prep_lbl = np.random.randint(self.num_lbl, size=bs)
tgt = torch.from_numpy(prep_lbl).long() == batch["output"]["lbl"]
list_orig_input_ids = []
list_masked_input_ids = []
for b_idx, (p, c1, c2, ques, lbl) in enumerate(
zip(list_premise, list_choice1, list_choice2, list_question,
list_lbl)):
txt_split_tuple = []
txt_trim = -1
if ques == "cause":
pet_pvps = self.pet_patterns_cause
elif ques == "effect":
pet_pvps = self.pet_patterns_effect
pattern = pet_pvps[self._pet_names.index(mode)]
if lbl.item() == 0:
lbl_choice = c1[:-1]
elif lbl.item() == 1:
lbl_choice = c2[:-1]
else:
raise ValueError("Invalid Lbl")
for idx, txt_split in enumerate(pattern):
txt_split_inp = txt_split.replace("[PREMISE]", p[:-1]).replace(
"[CHOICE1]",
c1[:-1]).replace("[CHOICE2]",
c2[:-1]).replace("[MASK]", lbl_choice)
txt_split_tuple.append(txt_split_inp)
if lbl.item() == 0:
# Trim the paragraph
if "[PREMISE]" in txt_split:
txt_trim = idx
elif lbl.item() == 1:
# Trim the paragraph
if "[PREMISE]" in txt_split:
txt_trim = idx
else:
raise ValueError("Invalid Lbl")
orig_input_ids, masked_input_ids, mask_idx = tokenize_pet_mlm_txt(
self.tokenizer, self.config, txt_split_tuple[0],
txt_split_tuple[1], txt_split_tuple[2], txt_trim)
list_orig_input_ids.append(orig_input_ids)
list_masked_input_ids.append(masked_input_ids)
return torch.tensor(list_orig_input_ids).to(device), torch.tensor(
list_masked_input_ids).to(device), prep_lbl, tgt.to(device)