-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathinference.py
165 lines (141 loc) · 7.01 KB
/
inference.py
1
2
3
4
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
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import argparse
import torch
import os
os.environ['HF_ENDPOINT']='https://hf-mirror.com'
import json
from tqdm import tqdm
import shortuuid
import whisper
from PIL import Image
from openomni.constants import SPEECH_TOKEN_INDEX, DEFAULT_SPEECH_TOKEN, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
from openomni.conversation import conv_templates, SeparatorStyle
from openomni.mm_utils import process_images
from openomni.model.builder import load_pretrained_qwen_model
from openomni.utils import disable_torch_init
from torch.utils.data import Dataset, DataLoader
import time
import math
def split_list(lst, n):
"""Split a list into n (roughly) equal-sized chunks"""
chunk_size = math.ceil(len(lst) / n) # integer division
return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
def get_chunk(lst, n, k):
chunks = split_list(lst, n)
return chunks[k]
def ctc_postprocess(tokens, blank):
_toks = tokens.squeeze(0).tolist()
print(_toks,len(_toks))
# deduplicated_toks = [v for i, v in enumerate(_toks) if i == 0 or v != _toks[i - 1]]
deduplicated_toks = [v for i, v in enumerate(_toks)]
hyp = [v for v in deduplicated_toks if v != blank]
hyp = " ".join(list(map(str, hyp)))
return hyp
def get_w(weights, keyword):
return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
def eval_model(args):
# Model
disable_torch_init()
# print(args.model_path)
model_path = os.path.expanduser(args.model_path)
# model_path=args.model_path
tokenizer, model, image_processor, context_len = load_pretrained_qwen_model(model_path, args.model_base, is_lora=args.is_lora)
# state_dict={k:v for k,v in model.named_parameters()}
# state_dict=get_w(state_dict, "model.mm_projector")
# torch.save(state_dict,"openomni/ckpt/pretrained/mm_projector/mm_projector.pt")
use_speech=True
tokenizer.add_tokens(["<image>"], special_tokens=True)
tokenizer.add_tokens(["<speech>"], special_tokens=True)
tokenizer.chat_template="{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
# tokenizer.chat_template="{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{''}}"
image_token_index = tokenizer.convert_tokens_to_ids("<image>")
speech_token_index = tokenizer.convert_tokens_to_ids("<speech>")
if args.square_eval:
model.config.image_grid_pinpoints = [
[
672,
672
]
]
speech_file = "./assets/question.wav"
image_file="./assets/example.png"
question="<image>\n Tell me something about the object in this image."
prompt=question.replace(DEFAULT_IMAGE_TOKEN, '').strip()
prompt = DEFAULT_IMAGE_TOKEN + '\n' + prompt
prompt = prompt.strip()
if use_speech:
prompt="<image>\n <speech>\n Please answer the questions in the user's input speech"
question_prompt=prompt
input_id=[]
system_message= "You are a helpful language, vision and speech assistant. You are able to understand the visual content that the user provides, and assist the user with a variety of tasks using natural language or speech."
input_id += tokenizer.apply_chat_template([{"role" : "system", "content" : system_message},
{"role" : "user", "content" : question_prompt}],
add_generation_prompt=True)
# input_id += encode_id
for idx, encode_id in enumerate(input_id):
if encode_id == image_token_index:
input_id[idx] = IMAGE_TOKEN_INDEX
if encode_id == speech_token_index:
input_id[idx] = SPEECH_TOKEN_INDEX
input_ids = torch.tensor([input_id], dtype=torch.long)
input_ids = input_ids.to(device='cuda', non_blocking=True)
image = Image.open(os.path.join(
'', image_file)).convert('RGB')
image_tensor = process_images(
[image], image_processor, model.config)[0]
speech = whisper.load_audio(os.path.join('',speech_file))
if args.input_type == "raw":
speech = torch.from_numpy(speech)
elif args.input_type == "mel":
speech = whisper.pad_or_trim(speech)
speech_tensor = whisper.log_mel_spectrogram(speech, n_mels=args.mel_size).permute(1, 0)
speech_length=torch.LongTensor([speech_tensor.shape[0]])
speech_tensor = speech_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True).unsqueeze(0)
speech_length = speech_length.to(device='cuda', non_blocking=True)
with torch.inference_mode():
time1=time.time()
outputs = model.generate(
input_ids,
images=image_tensor.unsqueeze(0).half().cuda(),
image_sizes=[image.size],
speech=speech_tensor,
speech_lengths=speech_length,
do_sample=True if args.temperature > 0 else False,
temperature=args.temperature,
top_p=args.top_p,
num_beams=args.num_beams,
max_new_tokens=args.max_new_tokens,
use_cache=True,
pad_token_id=tokenizer.pad_token_id,
streaming_unit_gen=False,
faster_infer=False if args.s2s else True
)
time2=time.time()
output_ids, output_units = outputs
outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
if args.s2s:
if args.speech_generator_type=="ar":
output_units=output_units
elif args.speech_generator_type=="ctc":
output_units = ctc_postprocess(output_units, blank=model.config.unit_vocab_size)
print(f"H-{time2-time1}-{idx}\t{outputs}")
if args.s2s:
print(f"U-{idx}\t{output_units}")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--model-path", type=str, default="./checkpoints/openomni_stage2_qwen_2/checkpoint-last")
parser.add_argument("--model-base", type=str, default=None)
parser.add_argument("--conv-mode", type=str, default="llava_qwen2")
parser.add_argument("--num-chunks", type=int, default=1)
parser.add_argument("--chunk-idx", type=int, default=0)
parser.add_argument("--temperature", type=float, default=0)
parser.add_argument("--top_p", type=float, default=None)
parser.add_argument("--num_beams", type=int, default=1)
parser.add_argument("--max_new_tokens", type=int, default=256)
parser.add_argument("--input_type", type=str, default="mel")
parser.add_argument("--mel_size", type=int, default=128)
parser.add_argument("--s2s", action="store_true", default=True)
parser.add_argument("--speech_generator_type", action="store_true", default="ar")
parser.add_argument("--is_lora", action="store_true", default=False)
parser.add_argument("--square_eval", type=bool, default=True)
args = parser.parse_args()
eval_model(args)