-
Notifications
You must be signed in to change notification settings - Fork 40
/
simctgt5.py
82 lines (74 loc) · 2.77 KB
/
simctgt5.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
import sys
import ipdb
import os
import operator
from operator import itemgetter
import torch
from torch import nn
import random
import argparse
import numpy as np
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss
import seaborn as sns
import matplotlib.pyplot as plt
import pandas as pd
train_fct = CrossEntropyLoss()
val_fct = CrossEntropyLoss(reduction='none')
class SimCTGT5(nn.Module):
def __init__(self, model_name):
super(SimCTGT5, self).__init__()
from transformers import AutoTokenizer, T5ForConditionalGeneration
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.model = T5ForConditionalGeneration.from_pretrained(model_name)
self.vocab_size = len(self.tokenizer)
self.embed_dim = self.model.config.hidden_size
self.pad_token_id = self.tokenizer.pad_token_id
# decoding functions
# ------------------------------------------------------- #
@torch.no_grad()
def fast_contrastive_search(self, input_ids, decoder_ids, beam_width, alpha, decoding_len):
'''
input_ids: prefix input; 1 x prefix_len
decoding_len: how many tokens to generate
beam_width: size of candidate pool during decoding
alpha: regulates importance of model confidence and degeneration penalty
'''
self.model.eval()
from utlis import EncDecContrastiveDecodingOneStepFast
# sanity check
assert alpha >= 0. and alpha <= 1.0
batch_size, seqlen = input_ids.size()
generated = []
past_key_values = None
last_hidden_states = None
logits = None
input_embeds = None
for step in range(decoding_len):
decoder_ids, past_key_values, last_hidden_states, logits, input_embeds = EncDecContrastiveDecodingOneStepFast(
self.model,
input_ids,
decoder_ids,
beam_width,
alpha,
past_key_values,
last_hidden_states,
self.tokenizer,
logits,
first_step=step == 0,
input_embeds=input_embeds,
)
token = decoder_ids.squeeze(dim=-1).item()
generated.append(token)
return generated
def greedy_search(self, input_ids, decoding_len):
output = self.model.generate(
input_ids=input_ids,
max_length=decoding_len)
return output[0]
def beam_search(self, input_ids, beam_width, decoding_len):
output = self.model.generate(
input_ids=input_ids,
max_length=decoding_len,
num_beams=beam_width)
return output[0]