Update main.py
Browse files
main.py
CHANGED
@@ -1,67 +1,69 @@
|
|
1 |
-
import os
|
2 |
-
import torch
|
3 |
-
import torch.nn as nn
|
4 |
-
import numpy as np
|
5 |
-
import random
|
6 |
-
from transformers import (
|
7 |
-
BartForConditionalGeneration,
|
8 |
-
AutoModelForCausalLM,
|
9 |
-
BertModel,
|
10 |
-
Wav2Vec2Model,
|
11 |
-
CLIPModel,
|
12 |
-
AutoTokenizer
|
13 |
-
)
|
14 |
-
|
15 |
-
class MultiModalModel(nn.Module):
|
16 |
-
def __init__(self):
|
17 |
-
super(MultiModalModel, self).__init__()
|
18 |
-
# 初始化子模型
|
19 |
-
self.text_generator = BartForConditionalGeneration.from_pretrained('facebook/bart-base')
|
20 |
-
self.code_generator = AutoModelForCausalLM.from_pretrained('gpt2')
|
21 |
-
self.nlp_encoder = BertModel.from_pretrained('bert-base-uncased')
|
22 |
-
self.speech_encoder = Wav2Vec2Model.from_pretrained('facebook/wav2vec2-base-960h')
|
23 |
-
self.vision_encoder = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')
|
24 |
-
|
25 |
-
# 初始化分词器和处理器
|
26 |
-
self.text_tokenizer = AutoTokenizer.from_pretrained('facebook/bart-base')
|
27 |
-
self.code_tokenizer = AutoTokenizer.from_pretrained('gpt2')
|
28 |
-
self.nlp_tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
|
29 |
-
self.speech_processor = AutoTokenizer.from_pretrained('facebook/wav2vec2-base-960h')
|
30 |
-
self.vision_processor = AutoTokenizer.from_pretrained('openai/clip-vit-base-patch32')
|
31 |
-
|
32 |
-
def forward(self, task, inputs):
|
33 |
-
if task == 'text_generation':
|
34 |
-
# 确保 attention_mask 在 inputs 中
|
35 |
-
attention_mask = inputs.get('attention_mask')
|
36 |
-
print("输入数据:", inputs)
|
37 |
-
outputs = self.text_generator.generate(
|
38 |
-
inputs['input_ids'],
|
39 |
-
max_new_tokens=100, # 增加生成的最大新令牌数
|
40 |
-
pad_token_id=self.text_tokenizer.eos_token_id,
|
41 |
-
attention_mask=attention_mask,
|
42 |
-
top_p=0.9, # 调整 top_p 值
|
43 |
-
top_k=50, # 保持 top_k 值
|
44 |
-
temperature=0.8, # 调整 temperature 值
|
45 |
-
do_sample=True
|
46 |
-
)
|
47 |
-
print("生成的输出:", outputs)
|
48 |
-
return self.text_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
49 |
-
# 根据需要添加其他任务的逻辑...
|
50 |
-
|
51 |
-
# 主函数
|
52 |
-
if __name__ == "__main__":
|
53 |
-
# 初始化模型
|
54 |
-
model = MultiModalModel()
|
55 |
-
|
56 |
-
# 示例任务和输入数据
|
57 |
-
task = "text_generation"
|
58 |
-
input_text = "This is a sample input."
|
59 |
-
tokenizer = model.text_tokenizer
|
60 |
-
inputs = tokenizer(input_text, return_tensors='pt')
|
61 |
-
|
62 |
-
# 添加 attention_mask 键值对
|
63 |
-
inputs['attention_mask'] = torch.ones_like(inputs['input_ids'])
|
64 |
-
|
65 |
-
# 模型推理
|
66 |
-
result = model(task, inputs)
|
67 |
-
print("最终输出结果:", result)
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import numpy as np
|
5 |
+
import random
|
6 |
+
from transformers import (
|
7 |
+
BartForConditionalGeneration,
|
8 |
+
AutoModelForCausalLM,
|
9 |
+
BertModel,
|
10 |
+
Wav2Vec2Model,
|
11 |
+
CLIPModel,
|
12 |
+
AutoTokenizer
|
13 |
+
)
|
14 |
+
|
15 |
+
class MultiModalModel(nn.Module):
|
16 |
+
def __init__(self):
|
17 |
+
super(MultiModalModel, self).__init__()
|
18 |
+
# 初始化子模型
|
19 |
+
self.text_generator = BartForConditionalGeneration.from_pretrained('facebook/bart-base')
|
20 |
+
self.code_generator = AutoModelForCausalLM.from_pretrained('gpt2')
|
21 |
+
self.nlp_encoder = BertModel.from_pretrained('bert-base-uncased')
|
22 |
+
self.speech_encoder = Wav2Vec2Model.from_pretrained('facebook/wav2vec2-base-960h')
|
23 |
+
self.vision_encoder = CLIPModel.from_pretrained('openai/clip-vit-base-patch32')
|
24 |
+
|
25 |
+
# 初始化分词器和处理器
|
26 |
+
self.text_tokenizer = AutoTokenizer.from_pretrained('facebook/bart-base')
|
27 |
+
self.code_tokenizer = AutoTokenizer.from_pretrained('gpt2')
|
28 |
+
self.nlp_tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
|
29 |
+
self.speech_processor = AutoTokenizer.from_pretrained('facebook/wav2vec2-base-960h')
|
30 |
+
self.vision_processor = AutoTokenizer.from_pretrained('openai/clip-vit-base-patch32')
|
31 |
+
|
32 |
+
def forward(self, task, inputs):
|
33 |
+
if task == 'text_generation':
|
34 |
+
# 确保 attention_mask 在 inputs 中
|
35 |
+
attention_mask = inputs.get('attention_mask')
|
36 |
+
print("输入数据:", inputs)
|
37 |
+
outputs = self.text_generator.generate(
|
38 |
+
inputs['input_ids'],
|
39 |
+
max_new_tokens=100, # 增加生成的最大新令牌数
|
40 |
+
pad_token_id=self.text_tokenizer.eos_token_id,
|
41 |
+
attention_mask=attention_mask,
|
42 |
+
top_p=0.9, # 调整 top_p 值
|
43 |
+
top_k=50, # 保持 top_k 值
|
44 |
+
temperature=0.8, # 调整 temperature 值
|
45 |
+
do_sample=True
|
46 |
+
)
|
47 |
+
print("生成的输出:", outputs)
|
48 |
+
return self.text_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
49 |
+
# 根据需要添加其他任务的逻辑...
|
50 |
+
|
51 |
+
# 主函数
|
52 |
+
if __name__ == "__main__":
|
53 |
+
# 初始化模型
|
54 |
+
model = MultiModalModel()
|
55 |
+
|
56 |
+
# 示例任务和输入数据
|
57 |
+
task = "text_generation"
|
58 |
+
input_text = "This is a sample input."
|
59 |
+
tokenizer = model.text_tokenizer
|
60 |
+
inputs = tokenizer(input_text, return_tensors='pt')
|
61 |
+
|
62 |
+
# 添加 attention_mask 键值对
|
63 |
+
inputs['attention_mask'] = torch.ones_like(inputs['input_ids'])
|
64 |
+
|
65 |
+
# 模型推理
|
66 |
+
result = model(task, inputs)
|
67 |
+
print("最终输出结果:", result)
|
68 |
+
|
69 |
+
trust_remote_code=True
|