Justin Donaldson
commited on
Commit
•
e3d923c
1
Parent(s):
9dda517
fixing gallery
Browse files
.envrc
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
eval "$(conda shell.zsh hook)"
|
2 |
+
conda activate vibesearch
|
app.py
CHANGED
@@ -1,10 +1,11 @@
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
-
from transformers import CLIPProcessor, CLIPModel, CLIPTokenizer
|
3 |
import sentence_transformers
|
4 |
-
from sentence_transformers import SentenceTransformer, util
|
5 |
-
import pickle
|
6 |
from PIL import Image
|
7 |
-
import
|
|
|
8 |
|
9 |
# (Pdb) query_emb.shape
|
10 |
# torch.Size([1, 512])
|
@@ -12,25 +13,24 @@ import os
|
|
12 |
# (24996, 512)
|
13 |
|
14 |
|
15 |
-
|
16 |
## Define model
|
17 |
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
18 |
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
19 |
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
|
20 |
|
21 |
-
#Open the precomputed embeddings
|
22 |
-
emb_filename =
|
23 |
# emb_filename = 'unsplash-25k-photos-embeddings.pkl'
|
24 |
|
25 |
-
with open(emb_filename,
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
|
30 |
|
31 |
def search_text(query, top_k=4):
|
32 |
-
"""" Search an image based on the text query.
|
33 |
-
|
34 |
Args:
|
35 |
query ([string]): [query you want search for]
|
36 |
top_k (int, optional): [Amount of images o return]. Defaults to 1.
|
@@ -39,44 +39,49 @@ def search_text(query, top_k=4):
|
|
39 |
[list]: [list of images that are related to the query.]
|
40 |
"""
|
41 |
# First, we encode the query.
|
42 |
-
inputs = tokenizer([query],
|
43 |
query_emb = model.get_text_features(**inputs)
|
44 |
|
45 |
# import pdb; pdb.set_trace()
|
46 |
-
|
47 |
# Then, we use the util.semantic_search function, which computes the cosine-similarity
|
48 |
# between the query embedding and all image embeddings.
|
49 |
# It then returns the top_k highest ranked images, which we output
|
50 |
hits = util.semantic_search(query_emb, img_emb, top_k=top_k)[0]
|
51 |
-
|
52 |
-
image=[]
|
53 |
for hit in hits:
|
54 |
-
#print(img_names[hit['corpus_id']])
|
55 |
# object = Image.open(os.path.join("photos/", img_names[hit['corpus_id']]))
|
56 |
-
object = Image.open(os.path.join("lvphotos/", img_names[hit[
|
57 |
image.append(object)
|
58 |
-
#print(f'array length is: {len(image)}')
|
59 |
|
60 |
return image
|
61 |
-
|
62 |
|
63 |
iface = gr.Interface(
|
64 |
-
title
|
65 |
-
description
|
66 |
-
article
|
67 |
-
fn=search_text,
|
68 |
-
inputs=[
|
69 |
-
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
)
|
74 |
-
|
75 |
-
|
|
|
|
|
|
|
|
|
|
|
76 |
[("Vacation Star")],
|
77 |
[("Rock Star")],
|
78 |
[("Barbie")],
|
79 |
[("Small Purse")],
|
80 |
[("Big Bag")],
|
81 |
-
|
82 |
-
|
|
|
1 |
+
import os
|
2 |
+
import pickle
|
3 |
+
|
4 |
import gradio as gr
|
|
|
5 |
import sentence_transformers
|
|
|
|
|
6 |
from PIL import Image
|
7 |
+
from sentence_transformers import SentenceTransformer, util
|
8 |
+
from transformers import CLIPModel, CLIPProcessor, CLIPTokenizer
|
9 |
|
10 |
# (Pdb) query_emb.shape
|
11 |
# torch.Size([1, 512])
|
|
|
13 |
# (24996, 512)
|
14 |
|
15 |
|
|
|
16 |
## Define model
|
17 |
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
18 |
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
19 |
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")
|
20 |
|
21 |
+
# Open the precomputed embeddings
|
22 |
+
emb_filename = "lv-handbags.pkl"
|
23 |
# emb_filename = 'unsplash-25k-photos-embeddings.pkl'
|
24 |
|
25 |
+
with open(emb_filename, "rb") as fIn:
|
26 |
+
img_names, img_emb = pickle.load(fIn)
|
27 |
+
# print(f'img_emb: {print(img_emb)}')
|
28 |
+
# print(f'img_names: {print(img_names)}')
|
29 |
|
30 |
|
31 |
def search_text(query, top_k=4):
|
32 |
+
""" " Search an image based on the text query.
|
33 |
+
|
34 |
Args:
|
35 |
query ([string]): [query you want search for]
|
36 |
top_k (int, optional): [Amount of images o return]. Defaults to 1.
|
|
|
39 |
[list]: [list of images that are related to the query.]
|
40 |
"""
|
41 |
# First, we encode the query.
|
42 |
+
inputs = tokenizer([query], padding=True, return_tensors="pt")
|
43 |
query_emb = model.get_text_features(**inputs)
|
44 |
|
45 |
# import pdb; pdb.set_trace()
|
46 |
+
|
47 |
# Then, we use the util.semantic_search function, which computes the cosine-similarity
|
48 |
# between the query embedding and all image embeddings.
|
49 |
# It then returns the top_k highest ranked images, which we output
|
50 |
hits = util.semantic_search(query_emb, img_emb, top_k=top_k)[0]
|
51 |
+
|
52 |
+
image = []
|
53 |
for hit in hits:
|
54 |
+
# print(img_names[hit['corpus_id']])
|
55 |
# object = Image.open(os.path.join("photos/", img_names[hit['corpus_id']]))
|
56 |
+
object = Image.open(os.path.join("lvphotos/", img_names[hit["corpus_id"]]))
|
57 |
image.append(object)
|
58 |
+
# print(f'array length is: {len(image)}')
|
59 |
|
60 |
return image
|
61 |
+
|
62 |
|
63 |
iface = gr.Interface(
|
64 |
+
title="Hushh Vibe Search Model on Louis Vuitton API",
|
65 |
+
description="Quick demo of using text to perform vector search on an image collection",
|
66 |
+
article="TBD",
|
67 |
+
fn=search_text,
|
68 |
+
inputs=[
|
69 |
+
gr.Textbox(
|
70 |
+
lines=4,
|
71 |
+
label="Write what you are looking for in an image...",
|
72 |
+
placeholder="Text Here...",
|
73 |
+
)
|
74 |
+
],
|
75 |
+
outputs=[
|
76 |
+
gr.Gallery(
|
77 |
+
label="Generated images", show_label=False, elem_id="gallery", columns=2
|
78 |
+
)
|
79 |
+
],
|
80 |
+
examples=[
|
81 |
[("Vacation Star")],
|
82 |
[("Rock Star")],
|
83 |
[("Barbie")],
|
84 |
[("Small Purse")],
|
85 |
[("Big Bag")],
|
86 |
+
],
|
87 |
+
).launch(debug=True)
|