Upload 2 files
Browse files- config.json +51 -48
- vqapair.py +65 -0
config.json
CHANGED
@@ -1,48 +1,51 @@
|
|
1 |
-
{
|
2 |
-
"_name_or_path": "/content/drive/MyDrive/model/iter_39620_hf",
|
3 |
-
"architectures": [
|
4 |
-
"LlavaForConditionalGeneration"
|
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 |
-
}
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "/content/drive/MyDrive/model/iter_39620_hf",
|
3 |
+
"architectures": [
|
4 |
+
"LlavaForConditionalGeneration"
|
5 |
+
],
|
6 |
+
"auto_map": {
|
7 |
+
"AutoModelForCausalLM": "vqapair.VQApair"
|
8 |
+
},
|
9 |
+
"ignore_index": -100,
|
10 |
+
"image_token_index": 32011,
|
11 |
+
"model_type": "llava",
|
12 |
+
"pad_token_id": 32012,
|
13 |
+
"projector_hidden_act": "gelu",
|
14 |
+
"text_config": {
|
15 |
+
"_name_or_path": "../iter_32_xtuner_llama_llm",
|
16 |
+
"architectures": [
|
17 |
+
"LlamaForCausalLM"
|
18 |
+
],
|
19 |
+
"eos_token_id": 32000,
|
20 |
+
"hidden_size": 3072,
|
21 |
+
"intermediate_size": 8192,
|
22 |
+
"max_position_embeddings": 4096,
|
23 |
+
"model_type": "llama",
|
24 |
+
"original_max_position_embeddings": 4096,
|
25 |
+
"pad_token_id": 32000,
|
26 |
+
"rms_norm_eps": 1e-05,
|
27 |
+
"sliding_window": 2047,
|
28 |
+
"torch_dtype": "float16",
|
29 |
+
"vocab_size": 32064
|
30 |
+
},
|
31 |
+
"torch_dtype": "float16",
|
32 |
+
"transformers_version": "4.41.2",
|
33 |
+
"vision_config": {
|
34 |
+
"_name_or_path": "./visual_encoder",
|
35 |
+
"architectures": [
|
36 |
+
"CLIPVisionModel"
|
37 |
+
],
|
38 |
+
"dropout": 0.0,
|
39 |
+
"hidden_size": 1024,
|
40 |
+
"image_size": 336,
|
41 |
+
"intermediate_size": 4096,
|
42 |
+
"model_type": "clip_vision_model",
|
43 |
+
"num_attention_heads": 16,
|
44 |
+
"num_hidden_layers": 24,
|
45 |
+
"patch_size": 14,
|
46 |
+
"projection_dim": 768,
|
47 |
+
"torch_dtype": "float32"
|
48 |
+
},
|
49 |
+
"vision_feature_layer": -2,
|
50 |
+
"vision_feature_select_strategy": "default"
|
51 |
+
}
|
vqapair.py
ADDED
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn
|
3 |
+
from transformers import LlavaForConditionalGeneration, LlavaConfig
|
4 |
+
import re
|
5 |
+
|
6 |
+
from PIL import Image
|
7 |
+
from random import randint
|
8 |
+
|
9 |
+
class VQApair(LlavaForConditionalGeneration):
|
10 |
+
config_class = LlavaConfig
|
11 |
+
def __init__(self, config, **kwargs):
|
12 |
+
super().__init__(config)
|
13 |
+
self.processor = kwargs.pop("proc")
|
14 |
+
|
15 |
+
def genChoice(self, question, base_prompt, img_obj):
|
16 |
+
base_prompt += "{}<|end|>\n<|user|> Suggest 1 correct answer<|end|><|assistant|> ".format(question)
|
17 |
+
inputs = self.processor(base_prompt, img_obj, return_tensors='pt').to(0)
|
18 |
+
output = self.generate(**inputs, eos_token_id=32007, max_new_tokens=500)
|
19 |
+
index = torch.where(output[0]==32001)[0][-1].item()
|
20 |
+
answer = self.processor.decode(output[0][index:], skip_special_tokens=True)
|
21 |
+
base_prompt += "{}<|end|>\n<|user|> Suggest 3 incorrect answers<|end|><|assistant|> ".format(answer)
|
22 |
+
inputs = self.processor(base_prompt, img_obj, return_tensors='pt').to(0)
|
23 |
+
output = self.generate(**inputs, eos_token_id=32007, max_new_tokens=500)
|
24 |
+
index = torch.where(output[0]==32001)[0][-1].item()
|
25 |
+
choices = self.processor.decode(output[0][index:], skip_special_tokens=True)
|
26 |
+
a = choices.split("\n")
|
27 |
+
a = [x[3:].strip() for x in a]
|
28 |
+
a = [x for x in a if x]
|
29 |
+
correct_answer = randint(0,len(a))
|
30 |
+
a.insert(correct_answer, answer)
|
31 |
+
|
32 |
+
a = ["{}) {}".format(i+1, a[i]) for i in range(len(a))]
|
33 |
+
ans = "Correct Answer: {}".format(a[correct_answer])
|
34 |
+
return {"Choices": a, "Answers": ans}
|
35 |
+
|
36 |
+
def generateQn(self, img_path, n):
|
37 |
+
#commands = ["Generate a simple question",""]
|
38 |
+
prompt ='''
|
39 |
+
<|user|>\n<image>\nDescribe this image in a passage<|end|><|assistant|>
|
40 |
+
'''
|
41 |
+
artifacts = []
|
42 |
+
img_obj = Image.open(img_path)
|
43 |
+
|
44 |
+
|
45 |
+
inputs = self.processor(prompt, img_obj, return_tensors='pt').to(0)
|
46 |
+
#Generate Desc
|
47 |
+
output = self.generate(**inputs, eos_token_id=32007, max_new_tokens=500)
|
48 |
+
index = torch.where(output[0]==32001)[0][-1].item()
|
49 |
+
desc = self.processor.decode(output[0][index:], skip_special_tokens=True)
|
50 |
+
#Update Prompt to generate question
|
51 |
+
prompt += "{}<|end|>\n<|user|> {}<|end|><|assistant|> ".format(desc,"Generate a simple question")
|
52 |
+
inputs = self.processor(prompt, img_obj, return_tensors='pt').to(0)
|
53 |
+
#Generate k questions
|
54 |
+
output = self.generate(**inputs, eos_token_id=32007, max_new_tokens=500, do_sample=False, num_beams=3,num_beam_groups=3,diversity_penalty=10.0, num_return_sequences=n)
|
55 |
+
for out in output:
|
56 |
+
entry = {}
|
57 |
+
index = torch.where(out==32001)[0][-1].item()
|
58 |
+
text = self.processor.decode(out[index:], skip_special_tokens=True)
|
59 |
+
entry.update({"desc":desc})
|
60 |
+
entry.update({"question":text})
|
61 |
+
entry.update(self.genChoice(text, prompt, img_obj))
|
62 |
+
artifacts.append(entry)
|
63 |
+
|
64 |
+
return artifacts
|
65 |
+
|