Vivien
commited on
Commit
•
5185219
1
Parent(s):
dab4ce7
Add possibility to compose queries and use images as queries
Browse files- app.py +75 -27
- requirements.txt +1 -0
app.py
CHANGED
@@ -1,8 +1,9 @@
|
|
|
|
|
|
1 |
import streamlit as st
|
2 |
import pandas as pd, numpy as np
|
3 |
-
from html import escape
|
4 |
-
import os
|
5 |
from transformers import CLIPProcessor, CLIPModel
|
|
|
6 |
|
7 |
|
8 |
@st.cache(
|
@@ -19,47 +20,72 @@ def load():
|
|
19 |
df = {0: pd.read_csv("data.csv"), 1: pd.read_csv("data2.csv")}
|
20 |
embeddings = {0: np.load("embeddings.npy"), 1: np.load("embeddings2.npy")}
|
21 |
for k in [0, 1]:
|
22 |
-
embeddings[k] = np.
|
23 |
-
|
|
|
24 |
)
|
25 |
return model, processor, df, embeddings
|
26 |
|
27 |
|
28 |
model, processor, df, embeddings = load()
|
29 |
-
|
30 |
source = {0: "\nSource: Unsplash", 1: "\nSource: The Movie Database (TMDB)"}
|
31 |
|
32 |
|
33 |
-
def get_html(url_list, height=200):
|
34 |
-
html = "<div style='margin-top: 20px; max-width: 1200px; display: flex; flex-wrap: wrap; justify-content: space-evenly'>"
|
35 |
-
for url, title, link in url_list:
|
36 |
-
html2 = f"<img title='{escape(title)}' style='height: {height}px; margin: 5px' src='{escape(url)}'>"
|
37 |
-
if len(link) > 0:
|
38 |
-
html2 = f"<a href='{escape(link)}' target='_blank'>" + html2 + "</a>"
|
39 |
-
html = html + html2
|
40 |
-
html += "</div>"
|
41 |
-
return html
|
42 |
-
|
43 |
-
|
44 |
def compute_text_embeddings(list_of_strings):
|
45 |
inputs = processor(text=list_of_strings, return_tensors="pt", padding=True)
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
st.cache(show_spinner=False)
|
50 |
|
51 |
|
52 |
def image_search(query, corpus, n_results=24):
|
53 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
k = 0 if corpus == "Unsplash" else 1
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
return [
|
59 |
(
|
60 |
df[k].iloc[i]["path"],
|
61 |
df[k].iloc[i]["tooltip"] + source[k],
|
62 |
-
|
63 |
)
|
64 |
for i in results
|
65 |
]
|
@@ -112,11 +138,33 @@ def main():
|
|
112 |
)
|
113 |
st.sidebar.markdown(description)
|
114 |
_, c, _ = st.columns((1, 3, 1))
|
115 |
-
|
|
|
|
|
|
|
116 |
corpus = st.radio("", ["Unsplash", "Movies"])
|
117 |
if len(query) > 0:
|
118 |
results = image_search(query, corpus)
|
119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
|
121 |
|
122 |
if __name__ == "__main__":
|
|
|
1 |
+
from html import escape
|
2 |
+
import re
|
3 |
import streamlit as st
|
4 |
import pandas as pd, numpy as np
|
|
|
|
|
5 |
from transformers import CLIPProcessor, CLIPModel
|
6 |
+
from st_clickable_images import clickable_images
|
7 |
|
8 |
|
9 |
@st.cache(
|
|
|
20 |
df = {0: pd.read_csv("data.csv"), 1: pd.read_csv("data2.csv")}
|
21 |
embeddings = {0: np.load("embeddings.npy"), 1: np.load("embeddings2.npy")}
|
22 |
for k in [0, 1]:
|
23 |
+
embeddings[k] = embeddings[k] - np.mean(embeddings[k], axis=0)
|
24 |
+
embeddings[k] = embeddings[k] / np.linalg.norm(
|
25 |
+
embeddings[k], axis=1, keepdims=True
|
26 |
)
|
27 |
return model, processor, df, embeddings
|
28 |
|
29 |
|
30 |
model, processor, df, embeddings = load()
|
|
|
31 |
source = {0: "\nSource: Unsplash", 1: "\nSource: The Movie Database (TMDB)"}
|
32 |
|
33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
def compute_text_embeddings(list_of_strings):
|
35 |
inputs = processor(text=list_of_strings, return_tensors="pt", padding=True)
|
36 |
+
result = model.get_text_features(**inputs).detach().numpy()
|
37 |
+
return result / np.linalg.norm(result, axis=1, keepdims=True)
|
|
|
|
|
38 |
|
39 |
|
40 |
def image_search(query, corpus, n_results=24):
|
41 |
+
positive_embeddings = None
|
42 |
+
|
43 |
+
def concatenate_embeddings(e1, e2):
|
44 |
+
if e1 is None:
|
45 |
+
return e2
|
46 |
+
else:
|
47 |
+
return np.concatenate((e1, e2), axis=0)
|
48 |
+
|
49 |
+
splitted_query = query.split("/")
|
50 |
+
|
51 |
+
positive_queries = splitted_query[0].split(";")
|
52 |
+
for positive_query in positive_queries:
|
53 |
+
match = re.match(r"\[(Movies|Unsplash):(\d{1,5})\](.*)", positive_query)
|
54 |
+
if match:
|
55 |
+
corpus2, idx, remainder = match.groups()
|
56 |
+
idx, remainder = int(idx), remainder.strip()
|
57 |
+
k = 0 if corpus2 == "Unsplash" else 1
|
58 |
+
positive_embeddings = concatenate_embeddings(
|
59 |
+
positive_embeddings, embeddings[k][idx : idx + 1, :]
|
60 |
+
)
|
61 |
+
if len(remainder) > 0:
|
62 |
+
positive_embeddings = concatenate_embeddings(
|
63 |
+
positive_embeddings, compute_text_embeddings([remainder])
|
64 |
+
)
|
65 |
+
else:
|
66 |
+
positive_embeddings = concatenate_embeddings(
|
67 |
+
positive_embeddings, compute_text_embeddings([positive_query])
|
68 |
+
)
|
69 |
k = 0 if corpus == "Unsplash" else 1
|
70 |
+
dot_product = embeddings[k] @ positive_embeddings.T
|
71 |
+
dot_product = dot_product - np.mean(dot_product, axis=0)
|
72 |
+
dot_product = dot_product / np.linalg.norm(dot_product, axis=0)
|
73 |
+
dot_product = np.min(dot_product, axis=1)
|
74 |
+
|
75 |
+
if len(splitted_query) > 1:
|
76 |
+
negative_queries = (" ".join(splitted_query[1:])).split(";")
|
77 |
+
negative_embeddings = compute_text_embeddings(negative_queries)
|
78 |
+
dot_product2 = embeddings[k] @ negative_embeddings.T
|
79 |
+
dot_product2 = dot_product2 - np.mean(dot_product2, axis=0)
|
80 |
+
dot_product2 = dot_product2 / np.linalg.norm(dot_product2, axis=0)
|
81 |
+
dot_product -= np.max(dot_product2, axis=1)
|
82 |
+
|
83 |
+
results = np.argsort(dot_product)[-1 : -n_results - 1 : -1]
|
84 |
return [
|
85 |
(
|
86 |
df[k].iloc[i]["path"],
|
87 |
df[k].iloc[i]["tooltip"] + source[k],
|
88 |
+
i,
|
89 |
)
|
90 |
for i in results
|
91 |
]
|
|
|
138 |
)
|
139 |
st.sidebar.markdown(description)
|
140 |
_, c, _ = st.columns((1, 3, 1))
|
141 |
+
if "query" in st.session_state:
|
142 |
+
query = c.text_input("", value=st.session_state["query"])
|
143 |
+
else:
|
144 |
+
query = c.text_input("", value="clouds at sunset")
|
145 |
corpus = st.radio("", ["Unsplash", "Movies"])
|
146 |
if len(query) > 0:
|
147 |
results = image_search(query, corpus)
|
148 |
+
clicked = clickable_images(
|
149 |
+
[result[0] for result in results],
|
150 |
+
titles=[result[1] for result in results],
|
151 |
+
div_style={
|
152 |
+
"display": "flex",
|
153 |
+
"justify-content": "center",
|
154 |
+
"flex-wrap": "wrap",
|
155 |
+
},
|
156 |
+
img_style={"margin": "2px", "height": "200px"},
|
157 |
+
)
|
158 |
+
if clicked >= 0:
|
159 |
+
change_query = False
|
160 |
+
if "last_clicked" not in st.session_state:
|
161 |
+
change_query = True
|
162 |
+
else:
|
163 |
+
if clicked != st.session_state["last_clicked"]:
|
164 |
+
change_query = True
|
165 |
+
if change_query:
|
166 |
+
st.session_state["query"] = f"[{corpus}:{results[clicked][2]}]"
|
167 |
+
st.experimental_rerun()
|
168 |
|
169 |
|
170 |
if __name__ == "__main__":
|
requirements.txt
CHANGED
@@ -2,3 +2,4 @@ torch
|
|
2 |
transformers
|
3 |
numpy
|
4 |
pandas
|
|
|
|
2 |
transformers
|
3 |
numpy
|
4 |
pandas
|
5 |
+
st-clickable-images
|