Spaces:
Runtime error
Runtime error
Create your Gradio app.py
Browse files
app.py
ADDED
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
3 |
+
import torch
|
4 |
+
from clip_interrogator import Config, Interrogator
|
5 |
+
import random
|
6 |
+
import re
|
7 |
+
import requests
|
8 |
+
import shutil
|
9 |
+
from PIL import Image
|
10 |
+
|
11 |
+
|
12 |
+
# Definir la funciΓ³n para generar prompt desde imagen
|
13 |
+
config = Config()
|
14 |
+
config.device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
15 |
+
config.blip_offload = False if torch.cuda.is_available() else True
|
16 |
+
config.chunk_size = 2048
|
17 |
+
config.flavor_intermediate_count = 512
|
18 |
+
config.blip_num_beams = 64
|
19 |
+
config.clip_model_name = "ViT-H-14/laion2b_s32b_b79k"
|
20 |
+
ci = Interrogator(config)
|
21 |
+
|
22 |
+
def get_prompt_from_image(image, mode):
|
23 |
+
image = image.convert('RGB')
|
24 |
+
if mode == 'best':
|
25 |
+
prompt = ci.interrogate(image)
|
26 |
+
elif mode == 'classic':
|
27 |
+
prompt = ci.interrogate_classic(image)
|
28 |
+
elif mode == 'fast':
|
29 |
+
prompt = ci.interrogate_fast(image)
|
30 |
+
elif mode == 'negative':
|
31 |
+
prompt = ci.interrogate_negative(image)
|
32 |
+
return prompt
|
33 |
+
|
34 |
+
|
35 |
+
# Definir la funciΓ³n para generar prompt desde texto
|
36 |
+
model = AutoModelForSeq2SeqLM.from_pretrained('Helsinki-NLP/opus-mt-zh-en').eval()
|
37 |
+
tokenizer = AutoTokenizer.from_pretrained('Helsinki-NLP/opus-mt-zh-en')
|
38 |
+
|
39 |
+
def translate(text):
|
40 |
+
with torch.no_grad():
|
41 |
+
encoded = tokenizer([text], return_tensors='pt')
|
42 |
+
sequences = model.generate(**encoded)
|
43 |
+
return tokenizer.batch_decode(sequences, skip_special_tokens=True)[0]
|
44 |
+
|
45 |
+
text_pipe = pipeline('text-generation', model='succinctly/text2image-prompt-generator')
|
46 |
+
|
47 |
+
def text_generate(input):
|
48 |
+
seed = random.randint(100, 1000000)
|
49 |
+
set_seed(seed)
|
50 |
+
text_in_english = translate(input)
|
51 |
+
for count in range(6):
|
52 |
+
sequences = text_pipe(text_in_english, max_length=random.randint(60, 90), num_return_sequences=8)
|
53 |
+
list = []
|
54 |
+
for sequence in sequences:
|
55 |
+
line = sequence['generated_text'].strip()
|
56 |
+
if line != text_in_english and len(line) > (len(text_in_english) + 4) and line.endswith((':', '-', 'β')) is False:
|
57 |
+
list.append(line)
|
58 |
+
|
59 |
+
result = "\n".join(list)
|
60 |
+
result = re.sub('[^ ]+\.[^ ]+','', result)
|
61 |
+
result = result.replace('<', '').replace('>', '')
|
62 |
+
if result != '':
|
63 |
+
return result
|
64 |
+
if count == 5:
|
65 |
+
return result
|
66 |
+
|
67 |
+
|
68 |
+
# Definir la funciΓ³n que permite al usuario cargar una imagen desde una URL
|
69 |
+
def load_image_from_url(url):
|
70 |
+
response = requests.get(url, stream=True)
|
71 |
+
if response.status_code == 200:
|
72 |
+
with open('./image.jpg', 'wb') as f:
|
73 |
+
response.raw.decode_content = True
|
74 |
+
shutil.copyfileobj(response.raw, f)
|
75 |
+
return Image.open('./image.jpg')
|
76 |
+
else:
|
77 |
+
raise ValueError("No se pudo cargar la imagen")
|
78 |
+
|
79 |
+
|
80 |
+
# Crear la interfaz de usuario de Gradio
|
81 |
+
with gr.Interface(
|
82 |
+
[get_prompt_from_image, text_generate],
|
83 |
+
[
|
84 |
+
gr.inputs.Image(type='pil', label='Imagen'),
|
85 |
+
gr.inputs.Radio(['best', 'fast', 'classic', 'negative'], value='best', label='Modo'),
|
86 |
+
gr.inputs.Textbox(lines=6, label='Texto de entrada'),
|
87 |
+
],
|
88 |
+
[
|