Upload phi_captioning.py
Browse files- phi_captioning.py +86 -0
phi_captioning.py
ADDED
@@ -0,0 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
|
3 |
+
from swift.tuners import Swift #chinese toolkit for finetunin and inference
|
4 |
+
|
5 |
+
|
6 |
+
from swift.llm import (
|
7 |
+
get_model_tokenizer, get_template, inference, ModelType,
|
8 |
+
get_default_template_type, inference_stream
|
9 |
+
)
|
10 |
+
from swift.utils import seed_everything
|
11 |
+
import torch
|
12 |
+
from tqdm import tqdm
|
13 |
+
import time
|
14 |
+
|
15 |
+
model_type = ModelType.phi3_vision_128k_instruct # model type
|
16 |
+
template_type = get_default_template_type(model_type)
|
17 |
+
print(f'template_type: {template_type}')
|
18 |
+
|
19 |
+
model_path = "./phi3-1476" # by default it is the lora path, not sure if it works the same way with merged checkpoint
|
20 |
+
model, tokenizer = get_model_tokenizer(model_type, torch.bfloat16, model_kwargs={'device_map': 'auto'})
|
21 |
+
model.generation_config.max_new_tokens = 1256 #generation params. As for me - defaults with do_sample=False works better than anything.
|
22 |
+
model.generation_config.do_sample = False
|
23 |
+
#model.generation_config.top_p = 0.7
|
24 |
+
#model.generation_config.temperature = 0.3
|
25 |
+
model = Swift.from_pretrained(model, model_path, "lora", inference_mode=True)
|
26 |
+
template = get_template(template_type, tokenizer)
|
27 |
+
#seed_everything(6321)
|
28 |
+
|
29 |
+
text = 'Make a caption that describe this image'
|
30 |
+
image_dir = './images/' # path to images
|
31 |
+
txt_dir = './tags/' # path to txt files with tags (from danbooru or from WD_Tagger)
|
32 |
+
maintxt_dir = './maintxt/' # path for result txt caprtions in natureal language
|
33 |
+
|
34 |
+
# image parsing
|
35 |
+
image_files = [f for f in os.listdir(image_dir) if f.endswith('.jpg')]
|
36 |
+
|
37 |
+
total_files = len(image_files)
|
38 |
+
start_time = time.time()
|
39 |
+
|
40 |
+
progress_bar = tqdm(total=total_files, unit='file', bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt} [{elapsed}<{remaining}, {rate_fmt}{postfix}]')
|
41 |
+
total_elapsed_time = 0
|
42 |
+
processed_files = 0
|
43 |
+
|
44 |
+
# Main captioning cycle
|
45 |
+
for image_file in image_files:
|
46 |
+
image_path = os.path.join(image_dir, image_file)
|
47 |
+
if os.path.exists(image_path):
|
48 |
+
txt_file = os.path.splitext(image_file)[0] + '.txt'
|
49 |
+
txt_path = os.path.join(txt_dir, txt_file)
|
50 |
+
|
51 |
+
if os.path.exists(txt_path):
|
52 |
+
with open(txt_path, 'r', encoding='utf-8') as f:
|
53 |
+
tags = f.read().strip()
|
54 |
+
|
55 |
+
text = f'<img>{image_path}</img> Make a caption that describe this image. Here is the tags describing image: {tags}\n Find the relevant character\'s names in the tags and use it.'
|
56 |
+
print(text)
|
57 |
+
step_start_time = time.time()
|
58 |
+
response, history = inference(model, template, text, do_sample=True, temperature=0, repetition_penalty=1.05)
|
59 |
+
step_end_time = time.time()
|
60 |
+
step_time = step_end_time - step_start_time
|
61 |
+
total_elapsed_time += step_time
|
62 |
+
remaining_time = (total_elapsed_time / (processed_files + 1)) * (total_files - processed_files)
|
63 |
+
|
64 |
+
remaining_hours = int(remaining_time // 3600)
|
65 |
+
remaining_minutes = int((remaining_time % 3600) // 60)
|
66 |
+
remaining_seconds = int(remaining_time % 60)
|
67 |
+
|
68 |
+
progress_bar.set_postfix(remaining=f'\n', refresh=False)
|
69 |
+
print(f"\n\n\nFile {image_file}\nConsumed time: {step_time:.2f} s\n{response}")
|
70 |
+
|
71 |
+
# Создаем имя файла для сохранения ответа
|
72 |
+
output_file = os.path.splitext(image_file)[0] + '.txt'
|
73 |
+
output_path = os.path.join(maintxt_dir, output_file)
|
74 |
+
|
75 |
+
# Записываем ответ в файл
|
76 |
+
with open(output_path, 'w', encoding='utf-8') as f:
|
77 |
+
f.write(response)
|
78 |
+
|
79 |
+
print(f"Caption saved in file: {output_file} \n")
|
80 |
+
processed_files += 1
|
81 |
+
progress_bar.update(1)
|
82 |
+
else:
|
83 |
+
print(f"File {txt_file} doesn't exist.")
|
84 |
+
else:
|
85 |
+
print(f"Image {image_file} not found.")
|
86 |
+
progress_bar.close()
|