Update app.py
Browse files
app.py
CHANGED
@@ -7,101 +7,117 @@ import io
|
|
7 |
import boto3
|
8 |
import json
|
9 |
|
10 |
-
|
11 |
bedrock_runtime = boto3.client('bedrock-runtime',
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
)
|
16 |
|
|
|
17 |
def construct_bedrock_body(base64_string, text):
|
18 |
if text:
|
19 |
-
return json.dumps(
|
20 |
-
{
|
21 |
-
"inputImage": base64_string,
|
22 |
-
"embeddingConfig": {"outputEmbeddingLength": 1024},
|
23 |
-
"inputText": text
|
24 |
-
}
|
25 |
-
)
|
26 |
-
|
27 |
-
return json.dumps(
|
28 |
-
{
|
29 |
"inputImage": base64_string,
|
30 |
"embeddingConfig": {"outputEmbeddingLength": 1024},
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
|
|
|
|
|
|
|
|
35 |
def get_embedding_from_titan_multimodal(body):
|
36 |
-
|
37 |
-
|
38 |
response = bedrock_runtime.invoke_model(
|
39 |
body=body,
|
40 |
modelId="amazon.titan-embed-image-v1",
|
41 |
accept="application/json",
|
42 |
contentType="application/json",
|
43 |
)
|
44 |
-
|
45 |
response_body = json.loads(response.get("body").read())
|
46 |
return response_body["embedding"]
|
47 |
|
|
|
48 |
uri = os.environ.get('MONGODB_ATLAS_URI')
|
49 |
client = MongoClient(uri)
|
50 |
db_name = 'celebrity_1000_embeddings'
|
51 |
collection_name = 'celeb_images'
|
52 |
-
|
53 |
celeb_images = client[db_name][collection_name]
|
54 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
def start_image_search(image, text):
|
56 |
if not image:
|
57 |
-
## Alert the user to upload an image
|
58 |
raise gr.Error("Please upload an image first, make sure to press the 'Submit' button after selecting the image.")
|
59 |
buffered = io.BytesIO()
|
60 |
image = image.resize((800, 600))
|
61 |
-
image.save(buffered, format="JPEG", quality=85)
|
62 |
img_byte = buffered.getvalue()
|
63 |
-
# Encode this byte array to Base64
|
64 |
img_base64 = base64.b64encode(img_byte)
|
65 |
-
|
66 |
-
# Convert Base64 bytes to string for JSON serialization
|
67 |
img_base64_str = img_base64.decode('utf-8')
|
68 |
body = construct_bedrock_body(img_base64_str, text)
|
69 |
embedding = get_embedding_from_titan_multimodal(body)
|
70 |
|
71 |
-
doc =
|
72 |
-
|
73 |
-
"
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
|
|
|
|
|
|
79 |
|
80 |
-
|
81 |
-
for
|
82 |
-
|
83 |
-
|
84 |
-
|
|
|
|
|
|
|
85 |
|
|
|
|
|
|
|
86 |
with gr.Blocks() as demo:
|
87 |
-
gr.Markdown(
|
88 |
-
|
89 |
-
# MongoDB's Vector Celeb Image matcher
|
90 |
|
91 |
-
Upload an image and find the most similar celeb image from the database.
|
92 |
|
93 |
💪 Make a great pose to impact the search! 🤯
|
94 |
-
|
95 |
""")
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
inputs=[gr.Image(type="pil", label="Upload an image"),gr.Textbox(label="Enter an adjusment to the image")],
|
101 |
-
## outputs=gr.Image(type="pil")
|
102 |
-
outputs=gr.Gallery(
|
103 |
-
label="Located images", show_label=True, elem_id="gallery"
|
104 |
-
, columns=[3], rows=[1], object_fit="contain", height="auto")
|
105 |
)
|
106 |
|
107 |
demo.launch()
|
|
|
7 |
import boto3
|
8 |
import json
|
9 |
|
10 |
+
# AWS Bedrock client setup
|
11 |
bedrock_runtime = boto3.client('bedrock-runtime',
|
12 |
+
aws_access_key_id=os.environ.get('AWS_ACCESS_KEY'),
|
13 |
+
aws_secret_access_key=os.environ.get('AWS_SECRET_KEY'),
|
14 |
+
region_name="us-east-1")
|
|
|
15 |
|
16 |
+
# Function to construct the request body for Bedrock
|
17 |
def construct_bedrock_body(base64_string, text):
|
18 |
if text:
|
19 |
+
return json.dumps({
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
"inputImage": base64_string,
|
21 |
"embeddingConfig": {"outputEmbeddingLength": 1024},
|
22 |
+
"inputText": text
|
23 |
+
})
|
24 |
+
return json.dumps({
|
25 |
+
"inputImage": base64_string,
|
26 |
+
"embeddingConfig": {"outputEmbeddingLength": 1024},
|
27 |
+
})
|
28 |
+
|
29 |
+
# Function to get the embedding from Bedrock model
|
30 |
def get_embedding_from_titan_multimodal(body):
|
|
|
|
|
31 |
response = bedrock_runtime.invoke_model(
|
32 |
body=body,
|
33 |
modelId="amazon.titan-embed-image-v1",
|
34 |
accept="application/json",
|
35 |
contentType="application/json",
|
36 |
)
|
|
|
37 |
response_body = json.loads(response.get("body").read())
|
38 |
return response_body["embedding"]
|
39 |
|
40 |
+
# MongoDB setup
|
41 |
uri = os.environ.get('MONGODB_ATLAS_URI')
|
42 |
client = MongoClient(uri)
|
43 |
db_name = 'celebrity_1000_embeddings'
|
44 |
collection_name = 'celeb_images'
|
|
|
45 |
celeb_images = client[db_name][collection_name]
|
46 |
|
47 |
+
# Function to generate image description using Claude 3 Sonnet
|
48 |
+
def generate_image_description_with_claude(image_base64):
|
49 |
+
claude_body = json.dumps({
|
50 |
+
"anthropic_version": "bedrock-2023-05-31",
|
51 |
+
"max_tokens": 1000,
|
52 |
+
"system": "Please respond only in Spanish.",
|
53 |
+
"messages": [{
|
54 |
+
"role": "user",
|
55 |
+
"content": [
|
56 |
+
{"type": "image", "source": {"type": "base64", "media_type": "image/jpeg", "data": image_base64}},
|
57 |
+
{"type": "text", "text": "What's in this image?"}
|
58 |
+
]
|
59 |
+
}]
|
60 |
+
})
|
61 |
+
|
62 |
+
claude_response = bedrock_runtime.invoke_model(
|
63 |
+
body=claude_body,
|
64 |
+
modelId="anthropic.claude-3-sonnet-v1:0",
|
65 |
+
accept="application/json",
|
66 |
+
contentType="application/json",
|
67 |
+
)
|
68 |
+
response_body = json.loads(claude_response.get("body").read())
|
69 |
+
# Assuming the response contains a field 'content' with the description
|
70 |
+
return response_body["messages"][0]["content"][0].get("text", "No description available")
|
71 |
+
|
72 |
+
# Main function to start image search
|
73 |
def start_image_search(image, text):
|
74 |
if not image:
|
|
|
75 |
raise gr.Error("Please upload an image first, make sure to press the 'Submit' button after selecting the image.")
|
76 |
buffered = io.BytesIO()
|
77 |
image = image.resize((800, 600))
|
78 |
+
image.save(buffered, format="JPEG", quality=85)
|
79 |
img_byte = buffered.getvalue()
|
|
|
80 |
img_base64 = base64.b64encode(img_byte)
|
|
|
|
|
81 |
img_base64_str = img_base64.decode('utf-8')
|
82 |
body = construct_bedrock_body(img_base64_str, text)
|
83 |
embedding = get_embedding_from_titan_multimodal(body)
|
84 |
|
85 |
+
doc = list(celeb_images.aggregate([
|
86 |
+
{
|
87 |
+
"$vectorSearch": {
|
88 |
+
"index": "vector_index",
|
89 |
+
"path": "embeddings",
|
90 |
+
"queryVector": embedding,
|
91 |
+
"numCandidates": 15,
|
92 |
+
"limit": 3
|
93 |
+
}
|
94 |
+
}, {"$project": {"image": 1}}
|
95 |
+
]))
|
96 |
|
97 |
+
images_with_descriptions = []
|
98 |
+
for image_doc in doc:
|
99 |
+
pil_image = Image.open(io.BytesIO(base64.b64decode(image_doc['image'])))
|
100 |
+
img_byte = io.BytesIO()
|
101 |
+
pil_image.save(img_byte, format='JPEG')
|
102 |
+
img_base64 = base64.b64encode(img_byte.getvalue()).decode('utf-8')
|
103 |
+
description = generate_image_description_with_claude(img_base64)
|
104 |
+
images_with_descriptions.append((pil_image, description))
|
105 |
|
106 |
+
return images_with_descriptions
|
107 |
+
|
108 |
+
# Gradio Interface
|
109 |
with gr.Blocks() as demo:
|
110 |
+
gr.Markdown("""
|
111 |
+
# MongoDB's Vector Celeb Image Matcher
|
|
|
112 |
|
113 |
+
Upload an image and find the most similar celeb image from the database, along with an AI-generated description.
|
114 |
|
115 |
💪 Make a great pose to impact the search! 🤯
|
|
|
116 |
""")
|
117 |
+
gr.Interface(fn=start_image_search,
|
118 |
+
inputs=[gr.Image(type="pil", label="Upload an image"), gr.Textbox(label="Enter an adjustment to the image")],
|
119 |
+
outputs=gr.Gallery(label="Located images with AI-generated descriptions", show_label=True, elem_id="gallery",
|
120 |
+
columns=[3], rows=[1], object_fit="contain", height="auto")
|
|
|
|
|
|
|
|
|
|
|
121 |
)
|
122 |
|
123 |
demo.launch()
|