Update modeling_tinyllava_phi.py
Browse files- modeling_tinyllava_phi.py +165 -15
modeling_tinyllava_phi.py
CHANGED
@@ -1,15 +1,9 @@
|
|
1 |
-
|
2 |
-
from .configuration import TinyLlavaConfig, IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
|
3 |
-
|
4 |
-
#from .data_preprocess import load_image, process_images, tokenizer_image_token
|
5 |
-
from dataclasses import dataclass
|
6 |
from typing import List, Optional, Tuple, Union
|
7 |
-
import ast
|
8 |
import re
|
9 |
|
10 |
import torch
|
11 |
import torch.utils.checkpoint
|
12 |
-
from torch import nn
|
13 |
from torch.nn import functional as F
|
14 |
|
15 |
from transformers import PreTrainedModel
|
@@ -17,12 +11,11 @@ from transformers.modeling_outputs import CausalLMOutputWithPast
|
|
17 |
from transformers.generation.utils import GenerateOutput
|
18 |
from transformers import CLIPVisionModel, CLIPImageProcessor, SiglipVisionModel, SiglipImageProcessor
|
19 |
|
20 |
-
import
|
|
|
21 |
from transformers import AutoConfig, AutoModelForCausalLM, PhiForCausalLM
|
22 |
-
|
23 |
-
import
|
24 |
-
sys.path.append(os.path.dirname(sys.path[0]))
|
25 |
-
from . import test
|
26 |
|
27 |
# from tinyllava.utils.data_utils import get_value_from_kwargs
|
28 |
CONTROLLER_HEART_BEAT_EXPIRATION = 30
|
@@ -39,12 +32,170 @@ logger = logging.get_logger(__name__)
|
|
39 |
|
40 |
# this import has to be relative, otherwise, when setting trust_remote_code=True
|
41 |
# huggingface transformers won't be able to load the module correctly
|
42 |
-
from numbers import Number
|
43 |
from typing import List, Optional, Union
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
ACT_TYPE = {
|
49 |
'relu': nn.ReLU,
|
50 |
'gelu': nn.GELU
|
@@ -138,7 +289,6 @@ class TinyLlavaPreTrainedModel(PreTrainedModel):
|
|
138 |
return self.language_model._supports_sdpa
|
139 |
|
140 |
|
141 |
-
|
142 |
class TinyLlavaForConditionalGeneration(TinyLlavaPreTrainedModel):
|
143 |
def __init__(self, config: TinyLlavaConfig):
|
144 |
|
@@ -478,4 +628,4 @@ class TinyLlavaForConditionalGeneration(TinyLlavaPreTrainedModel):
|
|
478 |
|
479 |
|
480 |
AutoConfig.register("tinyllava", TinyLlavaConfig)
|
481 |
-
AutoModelForCausalLM.register(TinyLlavaConfig, TinyLlavaForConditionalGeneration)
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
from typing import List, Optional, Tuple, Union
|
|
|
2 |
import re
|
3 |
|
4 |
import torch
|
5 |
import torch.utils.checkpoint
|
6 |
+
from torch import nn
|
7 |
from torch.nn import functional as F
|
8 |
|
9 |
from transformers import PreTrainedModel
|
|
|
11 |
from transformers.generation.utils import GenerateOutput
|
12 |
from transformers import CLIPVisionModel, CLIPImageProcessor, SiglipVisionModel, SiglipImageProcessor
|
13 |
|
14 |
+
from .configuration import TinyLlavaConfig, IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
|
15 |
+
|
16 |
from transformers import AutoConfig, AutoModelForCausalLM, PhiForCausalLM
|
17 |
+
|
18 |
+
import time
|
|
|
|
|
19 |
|
20 |
# from tinyllava.utils.data_utils import get_value_from_kwargs
|
21 |
CONTROLLER_HEART_BEAT_EXPIRATION = 30
|
|
|
32 |
|
33 |
# this import has to be relative, otherwise, when setting trust_remote_code=True
|
34 |
# huggingface transformers won't be able to load the module correctly
|
|
|
35 |
from typing import List, Optional, Union
|
36 |
+
import requests
|
37 |
+
from PIL import Image
|
38 |
+
from io import BytesIO
|
39 |
+
import base64
|
40 |
+
|
41 |
+
# Model Constants
|
42 |
+
IGNORE_INDEX = -100
|
43 |
+
IMAGE_TOKEN_INDEX = -200
|
44 |
+
DEFAULT_IMAGE_TOKEN = "<image>"
|
45 |
+
DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
|
46 |
+
DEFAULT_IM_START_TOKEN = "<im_start>"
|
47 |
+
DEFAULT_IM_END_TOKEN = "<im_end>"
|
48 |
+
IMAGE_PLACEHOLDER = "<image-placeholder>"
|
49 |
+
import dataclasses
|
50 |
+
from enum import auto, Enum
|
51 |
+
from typing import List, Tuple
|
52 |
+
|
53 |
+
|
54 |
+
class SeparatorStyle(Enum):
|
55 |
+
"""Different separator style."""
|
56 |
+
SINGLE = auto()
|
57 |
+
TWO = auto()
|
58 |
+
MPT = auto()
|
59 |
+
PLAIN = auto()
|
60 |
+
LLAMA_2 = auto()
|
61 |
+
TINY_LLAMA = auto()
|
62 |
+
QWEN_2 = auto()
|
63 |
+
|
64 |
+
|
65 |
+
@dataclasses.dataclass
|
66 |
+
class Conversation:
|
67 |
+
"""A class that keeps all conversation history."""
|
68 |
+
system: str
|
69 |
+
roles: List[str]
|
70 |
+
messages: List[List[str]]
|
71 |
+
offset: int
|
72 |
+
sep_style: SeparatorStyle = SeparatorStyle.SINGLE
|
73 |
+
sep: str = "###"
|
74 |
+
sep2: str = None
|
75 |
+
version: str = "Unknown"
|
76 |
+
|
77 |
+
skip_next: bool = False
|
78 |
+
|
79 |
+
def get_prompt(self):
|
80 |
+
messages = self.messages
|
81 |
+
if len(messages) > 0 and type(messages[0][1]) is tuple:
|
82 |
+
messages = self.messages.copy()
|
83 |
+
init_role, init_msg = messages[0].copy()
|
84 |
+
init_msg = init_msg[0].replace("<image>", "").strip()
|
85 |
+
if 'mmtag' in self.version:
|
86 |
+
messages[0] = (init_role, init_msg)
|
87 |
+
messages.insert(0, (self.roles[0], "<Image><image></Image>"))
|
88 |
+
messages.insert(1, (self.roles[1], "Received."))
|
89 |
+
else:
|
90 |
+
messages[0] = (init_role, "<image>\n" + init_msg)
|
91 |
+
|
92 |
+
if self.sep_style == SeparatorStyle.TWO:
|
93 |
+
seps = [self.sep, self.sep2]
|
94 |
+
ret = self.system + seps[0]
|
95 |
+
for i, (role, message) in enumerate(messages):
|
96 |
+
if message:
|
97 |
+
if type(message) is tuple:
|
98 |
+
message, _, _ = message
|
99 |
+
ret += role + ": " + message + seps[i % 2]
|
100 |
+
else:
|
101 |
+
ret += role + ":"
|
102 |
+
else:
|
103 |
+
raise ValueError(f"Invalid style: {self.sep_style}")
|
104 |
+
|
105 |
+
return ret
|
106 |
+
|
107 |
+
def append_message(self, role, message):
|
108 |
+
self.messages.append([role, message])
|
109 |
+
|
110 |
+
def copy(self):
|
111 |
+
return Conversation(
|
112 |
+
system=self.system,
|
113 |
+
roles=self.roles,
|
114 |
+
messages=[[x, y] for x, y in self.messages],
|
115 |
+
offset=self.offset,
|
116 |
+
sep_style=self.sep_style,
|
117 |
+
sep=self.sep,
|
118 |
+
sep2=self.sep2,
|
119 |
+
version=self.version)
|
120 |
+
|
121 |
+
|
122 |
|
123 |
|
124 |
+
conv_phi_v0 = Conversation(
|
125 |
+
system="A chat between a curious user and an artificial intelligence assistant. "
|
126 |
+
"The assistant gives helpful, detailed, and polite answers to the user's questions.",
|
127 |
+
roles=("USER", "ASSISTANT"),
|
128 |
+
version="phi",
|
129 |
+
messages=(),
|
130 |
+
offset=0,
|
131 |
+
sep_style=SeparatorStyle.TWO,
|
132 |
+
sep=" ",
|
133 |
+
sep2="<|endoftext|>",
|
134 |
+
)
|
135 |
|
136 |
|
137 |
+
def load_image_from_base64(image):
|
138 |
+
return Image.open(BytesIO(base64.b64decode(image)))
|
139 |
+
|
140 |
+
|
141 |
+
def expand2square(pil_img, background_color):
|
142 |
+
width, height = pil_img.size
|
143 |
+
if width == height:
|
144 |
+
return pil_img
|
145 |
+
elif width > height:
|
146 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
147 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
148 |
+
return result
|
149 |
+
else:
|
150 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
151 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
152 |
+
return result
|
153 |
+
|
154 |
+
|
155 |
+
def process_images(images, image_processor, model_cfg):
|
156 |
+
image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
|
157 |
+
new_images = []
|
158 |
+
if image_aspect_ratio == 'pad':
|
159 |
+
for image in images:
|
160 |
+
image = expand2square(image, tuple(int(x*255) for x in image_processor.image_mean))
|
161 |
+
image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
|
162 |
+
new_images.append(image)
|
163 |
+
else:
|
164 |
+
return image_processor(images, return_tensors='pt')['pixel_values']
|
165 |
+
if all(x.shape == new_images[0].shape for x in new_images):
|
166 |
+
new_images = torch.stack(new_images, dim=0)
|
167 |
+
return new_images
|
168 |
+
|
169 |
+
|
170 |
+
def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
|
171 |
+
prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
|
172 |
+
|
173 |
+
def insert_separator(X, sep):
|
174 |
+
return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
|
175 |
+
|
176 |
+
input_ids = []
|
177 |
+
offset = 0
|
178 |
+
if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
|
179 |
+
offset = 1
|
180 |
+
input_ids.append(prompt_chunks[0][0])
|
181 |
+
|
182 |
+
for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
|
183 |
+
input_ids.extend(x[offset:])
|
184 |
+
|
185 |
+
if return_tensors is not None:
|
186 |
+
if return_tensors == 'pt':
|
187 |
+
return torch.tensor(input_ids, dtype=torch.long)
|
188 |
+
raise ValueError(f'Unsupported tensor type: {return_tensors}')
|
189 |
+
return input_ids
|
190 |
+
|
191 |
+
def load_image(image_file):
|
192 |
+
if image_file.startswith("http") or image_file.startswith("https"):
|
193 |
+
response = requests.get(image_file)
|
194 |
+
image = Image.open(BytesIO(response.content)).convert("RGB")
|
195 |
+
else:
|
196 |
+
image = Image.open(image_file).convert("RGB")
|
197 |
+
return image
|
198 |
+
|
199 |
ACT_TYPE = {
|
200 |
'relu': nn.ReLU,
|
201 |
'gelu': nn.GELU
|
|
|
289 |
return self.language_model._supports_sdpa
|
290 |
|
291 |
|
|
|
292 |
class TinyLlavaForConditionalGeneration(TinyLlavaPreTrainedModel):
|
293 |
def __init__(self, config: TinyLlavaConfig):
|
294 |
|
|
|
628 |
|
629 |
|
630 |
AutoConfig.register("tinyllava", TinyLlavaConfig)
|
631 |
+
AutoModelForCausalLM.register(TinyLlavaConfig, TinyLlavaForConditionalGeneration)
|