upload generation_utils.py
Browse files- generation_utils.py +288 -0
generation_utils.py
ADDED
@@ -0,0 +1,288 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List
|
2 |
+
from queue import Queue
|
3 |
+
|
4 |
+
import torch
|
5 |
+
from PIL import Image
|
6 |
+
from copy import deepcopy
|
7 |
+
import requests, os
|
8 |
+
|
9 |
+
IMAGE_TOKEN_INDEX=-200
|
10 |
+
blacklist = ['<image>', '<s>', '</s>']
|
11 |
+
max_num_images = 3 # phi has a context length limit of 2048 and each image occupies 576 tokens.
|
12 |
+
|
13 |
+
def input_moderation(texts: list[list[str]]):
|
14 |
+
# perform input moderation on each message
|
15 |
+
for text_pair in texts:
|
16 |
+
# in-place operation
|
17 |
+
for b in blacklist:
|
18 |
+
text_pair[0] = text_pair[0].replace(b, '')
|
19 |
+
if text_pair[1] is not None:
|
20 |
+
text_pair[1] = text_pair[1].replace(b, '')
|
21 |
+
|
22 |
+
return texts
|
23 |
+
|
24 |
+
def insert_image_placeholder(t, num_images, placeholder='<image>', sep='\n'):
|
25 |
+
for _ in range(num_images):
|
26 |
+
t = f"{placeholder}{sep}" + t
|
27 |
+
return t
|
28 |
+
|
29 |
+
def get_conv(texts):
|
30 |
+
ret = []
|
31 |
+
|
32 |
+
for conv in texts:
|
33 |
+
ret.append({'from': 'human', 'value': conv[0]})
|
34 |
+
ret.append({'from': 'gpt', 'value': conv[1]}) # this is None for the last one
|
35 |
+
|
36 |
+
return ret
|
37 |
+
|
38 |
+
# copied from llava
|
39 |
+
def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
|
40 |
+
prompt_chunks = [tokenizer(chunk, add_special_tokens=False).input_ids for chunk in prompt.split('<image>')]
|
41 |
+
|
42 |
+
def insert_separator(X, sep):
|
43 |
+
return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
|
44 |
+
|
45 |
+
input_ids = []
|
46 |
+
offset = 0
|
47 |
+
if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
|
48 |
+
offset = 1
|
49 |
+
input_ids.append(prompt_chunks[0][0])
|
50 |
+
|
51 |
+
for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
|
52 |
+
input_ids.extend(x[offset:])
|
53 |
+
|
54 |
+
if return_tensors is not None:
|
55 |
+
if return_tensors == 'pt':
|
56 |
+
return torch.tensor(input_ids, dtype=torch.long)
|
57 |
+
raise ValueError(f'Unsupported tensor type: {return_tensors}')
|
58 |
+
return input_ids
|
59 |
+
|
60 |
+
def preprocess(tokenizer, data: list, return_tensors='pt'):
|
61 |
+
'''
|
62 |
+
[
|
63 |
+
{
|
64 |
+
'from': 'human',
|
65 |
+
'value': xxx,
|
66 |
+
},
|
67 |
+
{
|
68 |
+
'from': 'gpt',
|
69 |
+
'value': xxx
|
70 |
+
}
|
71 |
+
]
|
72 |
+
'''
|
73 |
+
# needs update
|
74 |
+
if not isinstance(data, list):
|
75 |
+
raise ValueError('must be a list')
|
76 |
+
|
77 |
+
# this is per model (tokenizer)
|
78 |
+
return preprocess_allava(tokenizer, data, return_tensors=return_tensors)
|
79 |
+
|
80 |
+
|
81 |
+
|
82 |
+
def preprocess_vicuna_v1(self, convs: list, return_tensors) -> list: # tokenize and concat the coversations
|
83 |
+
input_ids = None
|
84 |
+
for ind, conv in enumerate(convs):
|
85 |
+
if ind % 2 == 0: # human
|
86 |
+
h = conv['value'].strip()
|
87 |
+
h = f"USER: {h} "
|
88 |
+
cur_input_ids = self.tokenizer_image_token(prompt=h, return_tensors=return_tensors)
|
89 |
+
|
90 |
+
if input_ids is None:
|
91 |
+
input_ids = cur_input_ids
|
92 |
+
else:
|
93 |
+
input_ids = torch.cat([input_ids, cur_input_ids])
|
94 |
+
|
95 |
+
else: # gpt
|
96 |
+
g = conv['value']
|
97 |
+
if g is not None:
|
98 |
+
cur_input_ids = self.tokenizer(f"ASSISTANT: {g}</s>", add_special_tokens= False, max_length=self.maxlen, truncation=True, return_tensors='pt').input_ids[0]
|
99 |
+
input_ids = torch.cat([input_ids, cur_input_ids])
|
100 |
+
else:
|
101 |
+
cur_input_ids = self.tokenizer(f"ASSISTANT:", add_special_tokens= False, max_length=self.maxlen, truncation=True, return_tensors='pt').input_ids[0]
|
102 |
+
input_ids = torch.cat([input_ids, cur_input_ids])
|
103 |
+
|
104 |
+
|
105 |
+
return input_ids
|
106 |
+
|
107 |
+
def preprocess_allava(tokenizer, convs: list, return_tensors) -> list: # tokenize and concat the coversations
|
108 |
+
input_ids = None
|
109 |
+
|
110 |
+
|
111 |
+
for ind, conv in enumerate(convs):
|
112 |
+
if ind % 2 == 0: # human
|
113 |
+
h = conv['value'].strip()
|
114 |
+
h = f"[INST] {h} [/INST] "
|
115 |
+
cur_input_ids = tokenizer_image_token(prompt=h, tokenizer=tokenizer, return_tensors=return_tensors)
|
116 |
+
|
117 |
+
if input_ids is None:
|
118 |
+
input_ids = cur_input_ids
|
119 |
+
else:
|
120 |
+
input_ids = torch.cat([input_ids, cur_input_ids])
|
121 |
+
|
122 |
+
else: # gpt
|
123 |
+
g = conv['value']
|
124 |
+
if g is not None:
|
125 |
+
cur_input_ids = tokenizer(f"{g}{tokenizer.eos_token}", add_special_tokens= False, truncation=True, return_tensors='pt').input_ids[0]
|
126 |
+
input_ids = torch.cat([input_ids, cur_input_ids])
|
127 |
+
|
128 |
+
return input_ids
|
129 |
+
|
130 |
+
|
131 |
+
# copied from llava
|
132 |
+
def get_image_tensors(processor, images, device):
|
133 |
+
list_image_tensors = []
|
134 |
+
crop_size = processor.crop_size
|
135 |
+
for fp in images:
|
136 |
+
if fp is None: # None is used as a placeholder
|
137 |
+
list_image_tensors.append(torch.zeros(3, crop_size['height'], crop_size['width']).to(device))
|
138 |
+
continue
|
139 |
+
elif isinstance(fp, str):
|
140 |
+
image = Image.open(fp).convert('RGB')
|
141 |
+
elif isinstance(fp, Image.Image):
|
142 |
+
image = fp # already an image
|
143 |
+
else:
|
144 |
+
raise TypeError(f'Unsupported type {type(fp)}')
|
145 |
+
|
146 |
+
# this is the way of preprocessing images we used in training, so we impose it here
|
147 |
+
if True:
|
148 |
+
# self.data_args.image_aspect_ratio == 'pad'
|
149 |
+
def expand2square(pil_img, background_color):
|
150 |
+
width, height = pil_img.size
|
151 |
+
if pil_img.mode == 'L':
|
152 |
+
pil_img = pil_img.convert('RGB')
|
153 |
+
|
154 |
+
if width == height:
|
155 |
+
return pil_img
|
156 |
+
elif width > height:
|
157 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
158 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
159 |
+
return result
|
160 |
+
else:
|
161 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
162 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
163 |
+
return result
|
164 |
+
|
165 |
+
image = expand2square(image, tuple(int(x*255) for x in processor.image_mean))
|
166 |
+
image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
|
167 |
+
else:
|
168 |
+
image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0] # a tensor
|
169 |
+
list_image_tensors.append(image.to(device))
|
170 |
+
# list_image_tensors.append(image)
|
171 |
+
return list_image_tensors
|
172 |
+
|
173 |
+
|
174 |
+
|
175 |
+
|
176 |
+
def build_allava_input(tokenizer, processor, texts, images, history=None, return_history=False, device='cuda'):
|
177 |
+
'''
|
178 |
+
texts: [[]]
|
179 |
+
'''
|
180 |
+
|
181 |
+
############################
|
182 |
+
# 1. preprocess texts
|
183 |
+
############################
|
184 |
+
if isinstance(texts, str):
|
185 |
+
texts = [[texts, None]]
|
186 |
+
else:
|
187 |
+
assert isinstance(texts, list) and isinstance(texts[0], list) , 'texts must be a list of list'
|
188 |
+
|
189 |
+
if history is not None:
|
190 |
+
texts = history + texts # concat them together
|
191 |
+
|
192 |
+
texts = input_moderation(texts)
|
193 |
+
|
194 |
+
|
195 |
+
############################
|
196 |
+
# 2. preprocess images
|
197 |
+
############################
|
198 |
+
if isinstance(images, str) or isinstance(images, Image.Image):
|
199 |
+
images = [images]
|
200 |
+
|
201 |
+
valid_images = []
|
202 |
+
if images is None:
|
203 |
+
images = [None]
|
204 |
+
|
205 |
+
for img in images:
|
206 |
+
try:
|
207 |
+
if os.path.exists(img): # make sure that the path exists
|
208 |
+
img = Image.open(img).convert('RGB')
|
209 |
+
else: # else it must be a URL
|
210 |
+
img = Image.open(requests.get(img, stream=True).raw)
|
211 |
+
|
212 |
+
valid_images.append(img)
|
213 |
+
except:
|
214 |
+
continue
|
215 |
+
|
216 |
+
images = valid_images
|
217 |
+
|
218 |
+
if images == []:
|
219 |
+
images = [None]
|
220 |
+
|
221 |
+
|
222 |
+
assert len(images) < max_num_images, f'Currently at most {max_num_images} images are supported'
|
223 |
+
|
224 |
+
############################
|
225 |
+
# 3. collate conv
|
226 |
+
############################
|
227 |
+
|
228 |
+
history = deepcopy(texts) # history is the texts without <image> placeholders
|
229 |
+
|
230 |
+
# insert <image>
|
231 |
+
image_place_holder_inserted = insert_image_placeholder(texts[0][0], len(images) if None not in images else 0) # only insert the placeholders for user input at the 1st round
|
232 |
+
texts[0][0] = image_place_holder_inserted
|
233 |
+
|
234 |
+
# collate strings into conv
|
235 |
+
conv = get_conv(texts)
|
236 |
+
|
237 |
+
# make input ids
|
238 |
+
input_ids = preprocess(tokenizer, conv, return_tensors='pt').unsqueeze(0).to(device)
|
239 |
+
|
240 |
+
list_image_tensors = get_image_tensors(processor, images, device)
|
241 |
+
image_tensors = torch.stack(list_image_tensors)
|
242 |
+
|
243 |
+
try:
|
244 |
+
dtype = torch.bfloat16
|
245 |
+
# if your hardware does not support bf16, the following line raises an error
|
246 |
+
torch.tensor(1, dtype=dtype).cuda()
|
247 |
+
except:
|
248 |
+
# default using fp16
|
249 |
+
dtype = torch.float16
|
250 |
+
|
251 |
+
if return_history:
|
252 |
+
return input_ids, image_tensors, history
|
253 |
+
|
254 |
+
return input_ids, image_tensors, None
|
255 |
+
|
256 |
+
|
257 |
+
|
258 |
+
class TextIterStreamer:
|
259 |
+
def __init__(self, tokenizer, skip_prompt=False, skip_special_tokens=False):
|
260 |
+
self.tokenizer = tokenizer
|
261 |
+
self.skip_prompt = skip_prompt
|
262 |
+
self.skip_special_tokens = skip_special_tokens
|
263 |
+
self.tokens = []
|
264 |
+
self.text_queue = Queue()
|
265 |
+
self.next_tokens_are_prompt = True
|
266 |
+
|
267 |
+
def put(self, value):
|
268 |
+
if self.skip_prompt and self.next_tokens_are_prompt:
|
269 |
+
self.next_tokens_are_prompt = False
|
270 |
+
else:
|
271 |
+
if len(value.shape) > 1:
|
272 |
+
value = value[0]
|
273 |
+
self.tokens.extend(value.tolist())
|
274 |
+
self.text_queue.put(
|
275 |
+
self.tokenizer.decode(self.tokens, skip_special_tokens=self.skip_special_tokens))
|
276 |
+
|
277 |
+
def end(self):
|
278 |
+
self.text_queue.put(None)
|
279 |
+
|
280 |
+
def __iter__(self):
|
281 |
+
return self
|
282 |
+
|
283 |
+
def __next__(self):
|
284 |
+
value = self.text_queue.get()
|
285 |
+
if value is None:
|
286 |
+
raise StopIteration()
|
287 |
+
else:
|
288 |
+
return value
|