Update app.py
Browse files
app.py
CHANGED
@@ -9,14 +9,21 @@ tokenizer = BertTokenizer.from_pretrained("supermy/couplet-gpt2")
|
|
9 |
model = GPT2LMHeadModel.from_pretrained("supermy/couplet-gpt2")
|
10 |
model.eval()
|
11 |
|
|
|
|
|
12 |
def top_k_top_p_filtering( logits, top_k=0, top_p=0.0, filter_value=-float('Inf') ):
|
|
|
13 |
assert logits.dim() == 1
|
|
|
14 |
top_k = min( top_k, logits.size(-1) )
|
15 |
if top_k > 0:
|
|
|
|
|
|
|
16 |
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
|
17 |
-
logits[indices_to_remove] = filter_value
|
18 |
if top_p > 0.0:
|
19 |
-
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
|
20 |
cumulative_probs = torch.cumsum( F.softmax(sorted_logits, dim=-1), dim=-1 )
|
21 |
sorted_indices_to_remove = cumulative_probs > top_p
|
22 |
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
@@ -30,13 +37,17 @@ def generate(input_text):
|
|
30 |
input_ids.extend( tokenizer.encode(input_text + "-", add_special_tokens=False) )
|
31 |
input_ids = torch.tensor( [input_ids] )
|
32 |
|
|
|
33 |
generated = []
|
34 |
for _ in range(100):
|
35 |
output = model(input_ids)
|
36 |
-
|
37 |
next_token_logits = output.logits[0, -1, :]
|
|
|
38 |
next_token_logits[ tokenizer.convert_tokens_to_ids('[UNK]') ] = -float('Inf')
|
|
|
39 |
filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=8, top_p=1)
|
|
|
40 |
next_token = torch.multinomial( F.softmax(filtered_logits, dim=-1), num_samples=1 )
|
41 |
if next_token == tokenizer.sep_token_id:
|
42 |
break
|
|
|
9 |
model = GPT2LMHeadModel.from_pretrained("supermy/couplet-gpt2")
|
10 |
model.eval()
|
11 |
|
12 |
+
# top_k或top_p解码策略,仅保留top_k个或累积概率到达top_p的标记,其他标记设为filter_value,后续在选取标记的过程中会取不到值设为无穷小。
|
13 |
+
# 从模型输出的logit里面,划分出概率最高的几个
|
14 |
def top_k_top_p_filtering( logits, top_k=0, top_p=0.0, filter_value=-float('Inf') ):
|
15 |
+
#确保输出logit维度为1行若干列的矩阵,便于处理
|
16 |
assert logits.dim() == 1
|
17 |
+
#将top_k的值初始化为logit元素个数和设定的top_k之间的最小值
|
18 |
top_k = min( top_k, logits.size(-1) )
|
19 |
if top_k > 0:
|
20 |
+
# 去除掉概率值小于top_k里最后一个token概率的后续token
|
21 |
+
# torch.topk()返回最后一维最大的top_k个元素,返回值为二维(values,indices)
|
22 |
+
# ...表示其他维度由计算机自行推断
|
23 |
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
|
24 |
+
logits[indices_to_remove] = filter_value # 对于topk之外的其他元素的logits值设为负无穷
|
25 |
if top_p > 0.0:
|
26 |
+
sorted_logits, sorted_indices = torch.sort(logits, descending=True) # 对logits进行递减排序
|
27 |
cumulative_probs = torch.cumsum( F.softmax(sorted_logits, dim=-1), dim=-1 )
|
28 |
sorted_indices_to_remove = cumulative_probs > top_p
|
29 |
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
|
|
37 |
input_ids.extend( tokenizer.encode(input_text + "-", add_special_tokens=False) )
|
38 |
input_ids = torch.tensor( [input_ids] )
|
39 |
|
40 |
+
# 最多生成max_len个token
|
41 |
generated = []
|
42 |
for _ in range(100):
|
43 |
output = model(input_ids)
|
44 |
+
# 对于已生成的结果generated中的每个token添加一个重复惩罚项,降低其生成概率
|
45 |
next_token_logits = output.logits[0, -1, :]
|
46 |
+
# 对于[UNK]的概率设为无穷小,也就是说模型的预测结果不可能是[UNK]这个token
|
47 |
next_token_logits[ tokenizer.convert_tokens_to_ids('[UNK]') ] = -float('Inf')
|
48 |
+
# 使用top_k_top_p_filtering函数,按照top_k和top_p的值,对预测结果进行筛选
|
49 |
filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=8, top_p=1)
|
50 |
+
# torch.multinomial表示从候选集合中无放回地进行抽取num_samples个元素,权重越高,抽到的几率越高,返回元素的下标
|
51 |
next_token = torch.multinomial( F.softmax(filtered_logits, dim=-1), num_samples=1 )
|
52 |
if next_token == tokenizer.sep_token_id:
|
53 |
break
|