Spaces:
Running
Running
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
from PIL import Image
|
3 |
+
import requests
|
4 |
+
from io import BytesIO
|
5 |
+
from transformers import (
|
6 |
+
ViTFeatureExtractor,
|
7 |
+
ViTForImageClassification,
|
8 |
+
pipeline,
|
9 |
+
AutoFeatureExtractor,
|
10 |
+
AutoModelForObjectDetection,
|
11 |
+
CLIPTokenizerFast,
|
12 |
+
CLIPTextModel
|
13 |
+
)
|
14 |
+
import torch
|
15 |
+
from torchvision.transforms import functional as F
|
16 |
+
import emoji
|
17 |
+
|
18 |
+
# Load models
|
19 |
+
@st.cache_resource
|
20 |
+
def load_models():
|
21 |
+
age_model = ViTForImageClassification.from_pretrained('nateraw/vit-age-classifier')
|
22 |
+
age_transforms = ViTFeatureExtractor.from_pretrained('nateraw/vit-age-classifier')
|
23 |
+
|
24 |
+
gender_model = ViTForImageClassification.from_pretrained('rizvandwiki/gender-classification-2')
|
25 |
+
gender_transforms = ViTFeatureExtractor.from_pretrained('rizvandwiki/gender-classification-2')
|
26 |
+
|
27 |
+
emotion_model = ViTForImageClassification.from_pretrained('dima806/facial_emotions_image_detection')
|
28 |
+
emotion_transforms = ViTFeatureExtractor.from_pretrained('dima806/facial_emotions_image_detection')
|
29 |
+
|
30 |
+
object_detector = pipeline("object-detection", model="facebook/detr-resnet-50")
|
31 |
+
|
32 |
+
action_model = ViTForImageClassification.from_pretrained('rvv-karma/Human-Action-Recognition-VIT-Base-patch16-224')
|
33 |
+
action_transforms = ViTFeatureExtractor.from_pretrained('rvv-karma/Human-Action-Recognition-VIT-Base-patch16-224')
|
34 |
+
|
35 |
+
prompt_generator = pipeline("text2text-generation", model="succinctly/text2image-prompt-generator")
|
36 |
+
|
37 |
+
clip_tokenizer = CLIPTokenizerFast.from_pretrained("openai/clip-vit-base-patch32")
|
38 |
+
clip_model = CLIPTextModel.from_pretrained("openai/clip-vit-base-patch32")
|
39 |
+
|
40 |
+
return (age_model, age_transforms, gender_model, gender_transforms,
|
41 |
+
emotion_model, emotion_transforms, object_detector,
|
42 |
+
action_model, action_transforms, prompt_generator,
|
43 |
+
clip_tokenizer, clip_model)
|
44 |
+
|
45 |
+
models = load_models()
|
46 |
+
(age_model, age_transforms, gender_model, gender_transforms,
|
47 |
+
emotion_model, emotion_transforms, object_detector,
|
48 |
+
action_model, action_transforms, prompt_generator,
|
49 |
+
clip_tokenizer, clip_model) = models
|
50 |
+
|
51 |
+
def predict(image, model, transforms):
|
52 |
+
inputs = transforms(image, return_tensors='pt')
|
53 |
+
output = model(**inputs)
|
54 |
+
proba = output.logits.softmax(1)
|
55 |
+
return proba.argmax(1).item()
|
56 |
+
|
57 |
+
def detect_attributes(image):
|
58 |
+
age = predict(image, age_model, age_transforms)
|
59 |
+
gender = predict(image, gender_model, gender_transforms)
|
60 |
+
emotion = predict(image, emotion_model, emotion_transforms)
|
61 |
+
action = predict(image, action_model, action_transforms)
|
62 |
+
|
63 |
+
objects = object_detector(image)
|
64 |
+
|
65 |
+
return {
|
66 |
+
'age': age_model.config.id2label[age],
|
67 |
+
'gender': gender_model.config.id2label[gender],
|
68 |
+
'emotion': emotion_model.config.id2label[emotion],
|
69 |
+
'action': action_model.config.id2label[action],
|
70 |
+
'objects': [obj['label'] for obj in objects]
|
71 |
+
}
|
72 |
+
|
73 |
+
def generate_prompt(attributes):
|
74 |
+
prompt = f"A {attributes['age']} {attributes['gender']} person feeling {attributes['emotion']} "
|
75 |
+
prompt += f"while {attributes['action']}. "
|
76 |
+
if attributes['objects']:
|
77 |
+
prompt += f"Surrounded by {', '.join(attributes['objects'])}. "
|
78 |
+
return prompt
|
79 |
+
|
80 |
+
def generate_emoji(prompt):
|
81 |
+
inputs = clip_tokenizer(prompt, return_tensors="pt", padding=True, truncation=True)
|
82 |
+
outputs = clip_model(**inputs)
|
83 |
+
embedding = outputs.last_hidden_state.mean(dim=1)
|
84 |
+
|
85 |
+
# Simple emoji mapping based on embedding features
|
86 |
+
if embedding[0, 0] > 0:
|
87 |
+
return emoji.emojize(":grinning_face:")
|
88 |
+
elif embedding[0, 1] > 0:
|
89 |
+
return emoji.emojize(":smiling_face_with_heart-eyes:")
|
90 |
+
elif embedding[0, 2] > 0:
|
91 |
+
return emoji.emojize(":face_with_tears_of_joy:")
|
92 |
+
elif embedding[0, 3] > 0:
|
93 |
+
return emoji.emojize(":thinking_face:")
|
94 |
+
else:
|
95 |
+
return emoji.emojize(":neutral_face:")
|
96 |
+
|
97 |
+
st.title("Image Attribute Detection and Emoji Generation")
|
98 |
+
|
99 |
+
uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])
|
100 |
+
|
101 |
+
if uploaded_file is not None:
|
102 |
+
image = Image.open(uploaded_file)
|
103 |
+
st.image(image, caption='Uploaded Image', use_column_width=True)
|
104 |
+
|
105 |
+
if st.button('Analyze and Generate Emoji'):
|
106 |
+
with st.spinner('Detecting attributes...'):
|
107 |
+
attributes = detect_attributes(image)
|
108 |
+
|
109 |
+
st.write("Detected Attributes:")
|
110 |
+
for key, value in attributes.items():
|
111 |
+
st.write(f"{key.capitalize()}: {value}")
|
112 |
+
|
113 |
+
with st.spinner('Generating prompt...'):
|
114 |
+
prompt = generate_prompt(attributes)
|
115 |
+
st.write("Generated Prompt:")
|
116 |
+
st.write(prompt)
|
117 |
+
|
118 |
+
with st.spinner('Generating emoji...'):
|
119 |
+
emoji_result = generate_emoji(prompt)
|
120 |
+
st.write("Generated Emoji:")
|
121 |
+
st.write(emoji_result)
|