-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
122 lines (94 loc) · 4.07 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
from typing import Union, List
from torch import Tensor
import torch
import numpy as np
from datasets import load_dataset, load_metric
from transformers import T5Tokenizer
tokenizer = T5Tokenizer.from_pretrained('T5')
bleu_metric = load_metric("sacrebleu")
rouge_metric = load_metric("rouge")
meteor_metric = load_metric("meteor")
class InputExample:
"""
Structure for one input example with texts, the label and a unique id
"""
def __init__(self, guid: str = '', texts: List[str] = None, label: Union[int, float] = 0):
self.guid = guid
self.texts = texts
self.label = label
def __str__(self):
return "<InputExample> label: {}, texts: {}".format(str(self.label), "; ".join(self.texts))
def batch_to_device(batch, target_device):
"""
send a pytorch batch to a device (CPU/GPU)
"""
for key in batch:
if isinstance(batch[key], Tensor):
batch[key] = batch[key].to(target_device)
return batch
def sample_3d(probs, temperature=1):
'''probs.shape = (batch, seq_len, dim)'''
sample_idx = torch.zeros(probs.size(0), probs.size(1)).to(probs.device)
sample_probs = torch.zeros(probs.size(0), probs.size(1)).to(probs.device)
if temperature != 1:
temp = torch.exp(torch.div(torch.log(probs + 1e-20), temperature))
else:
temp = probs
for i, s in enumerate(temp):
temp_idx = torch.multinomial(s, 1) # shape = (seq_len, 1)
temp_probs = s.gather(1, temp_idx) # shape = (seq_len, 1)
sample_idx[i] = temp_idx.squeeze(1)
sample_probs[i] = temp_probs.squeeze(1)
return sample_probs, sample_idx.long()
def cal_reward_loss(sample_probs, reward, idxs=None):
sample_probs = sample_probs.contiguous()
sample_logprobs = torch.log(sample_probs)
reward = reward.unsqueeze(1).contiguous()
if idxs is not None:
batch_size, max_len = sample_probs.size()
mask = torch.zeros(batch_size, max_len).to(sample_probs.device)
for i, l in enumerate(idxs):
mask[i, :l] = 1
mask = mask.float().contiguous()
output = -sample_logprobs * reward * mask
output = (output.sum(-1) / mask.sum(-1)).mean()
else:
output = -sample_logprobs * reward
output = output.mean()
return output
def collate_fn(insts, pad_token_id=1):
''' Pad the instance to the max seq length in batch '''
max_len = max(len(inst) for inst in insts)
max_len = max_len if max_len > 4 else 5
batch_seq = np.array([
inst + [pad_token_id] * (max_len - len(inst))
for inst in insts])
batch_seq = torch.LongTensor(batch_seq)
return batch_seq
def postprocess_text(preds, labels):
preds = [pred.strip() for pred in preds]
labels = [[label.strip()] for label in labels]
return preds, labels
def compute_metrics(eval_preds):
preds, labels = eval_preds
if isinstance(preds, tuple):
preds = preds[0]
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
# Replace -100 in the labels as we can't decode them.
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
# Some simple post-processing
decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
bleu_result = bleu_metric.compute(predictions=decoded_preds, references=decoded_labels)
rouge_result = rouge_metric.compute(predictions=decoded_preds, references=decoded_labels)
meteor_result = meteor_metric.compute(predictions=decoded_preds, references=decoded_labels)
result = {key: value.mid.fmeasure * 100 for key, value in rouge_result.items()}
pair_list = []
for i in range(len(decoded_labels)):
pair_list.append([decoded_preds[i], decoded_labels[i][0]])
result['bleu'] = bleu_result['score']
result['meteor'] = meteor_result['meteor']
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
result["gen_len"] = np.mean(prediction_lens)
result = {k: round(v, 4) for k, v in result.items()}
return result