Spaces:
Sleeping
Sleeping
File size: 5,845 Bytes
2288165 cc6118a 2288165 cc6118a 2288165 cc6118a 2288165 cc6118a 2288165 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 |
import re
import streamlit as st
import pandas as pd
import numpy as np
from transformers import CLIPProcessor, CLIPModel
from st_clickable_images import clickable_images
def load():
model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
df = {0: pd.read_csv("data.csv"), 1: pd.read_csv("data2.csv")}
embeddings = {0: np.load("embeddings.npy"), 1: np.load("embeddings2.npy")}
for k in [0, 1]:
embeddings[k] = embeddings[k] / np.linalg.norm(
embeddings[k], axis=1, keepdims=True
)
return model, processor, df, embeddings
model, processor, df, embeddings = load()
source = {0: "\nSource: Unsplash", 1: "\nSource: The Movie Database (TMDB)"}
def compute_text_embeddings(list_of_strings):
inputs = processor(text=list_of_strings, return_tensors="pt", padding=True)
result = model.get_text_features(**inputs).detach().numpy()
return result / np.linalg.norm(result, axis=1, keepdims=True)
def image_search(query, corpus, n_results=24):
positive_embeddings = None
def concatenate_embeddings(e1, e2):
if e1 is None:
return e2
else:
return np.concatenate((e1, e2), axis=0)
splitted_query = query.split("EXCLUDING ")
dot_product = 0
k = 0 if corpus == "Unsplash" else 1
if len(splitted_query[0]) > 0:
positive_queries = splitted_query[0].split(";")
for positive_query in positive_queries:
match = re.match(r"\[(Movies|Unsplash):(\d{1,5})\](.*)", positive_query)
if match:
corpus2, idx, remainder = match.groups()
idx, remainder = int(idx), remainder.strip()
k2 = 0 if corpus2 == "Unsplash" else 1
positive_embeddings = concatenate_embeddings(
positive_embeddings, embeddings[k2][idx : idx + 1, :]
)
if len(remainder) > 0:
positive_embeddings = concatenate_embeddings(
positive_embeddings, compute_text_embeddings([remainder])
)
else:
positive_embeddings = concatenate_embeddings(
positive_embeddings, compute_text_embeddings([positive_query])
)
dot_product = embeddings[k] @ positive_embeddings.T
dot_product = dot_product - np.median(dot_product, axis=0)
dot_product = dot_product / np.max(dot_product, axis=0, keepdims=True)
dot_product = np.min(dot_product, axis=1)
if len(splitted_query) > 1:
negative_queries = (" ".join(splitted_query[1:])).split(";")
negative_embeddings = compute_text_embeddings(negative_queries)
dot_product2 = embeddings[k] @ negative_embeddings.T
dot_product2 = dot_product2 - np.median(dot_product2, axis=0)
dot_product2 = dot_product2 / np.max(dot_product2, axis=0, keepdims=True)
dot_product -= np.max(np.maximum(dot_product2, 0), axis=1)
results = np.argsort(dot_product)[-1 : -n_results - 1 : -1]
return [
(
df[k].iloc[i]["path"],
df[k].iloc[i]["tooltip"] + source[k],
i,
)
for i in results
]
def main():
st.markdown(
"""
<style>
.block-container{
max-width: 1200px;
}
div.row-widget.stRadio > div{
flex-direction:row;
display: flex;
justify-content: center;
}
div.row-widget.stRadio > div > label{
margin-left: 5px;
margin-right: 5px;
}
section.main>div:first-child {
padding-top: 0px;
}
section:not(.main)>div:first-child {
padding-top: 30px;
}
div.reportview-container > section:first-child{
max-width: 320px;
}
#MainMenu {
visibility: hidden;
}
footer {
visibility: hidden;
}
</style>""",
unsafe_allow_html=True,
)
st.sidebar.markdown("# Semantic Image Search")
st.sidebar.markdown("**Enter your query and hit enter**")
st.sidebar.markdown("- Click image to find similar images")
st.sidebar.markdown("- Use '**;**' to combine multiple queries")
st.sidebar.markdown("- Use '**EXCLUDING**' to exclude a query")
if "query" in st.session_state:
query = st.sidebar.text_input("Query", value=st.session_state["query"])
else:
query = st.sidebar.text_input("Query", value="lighthouse")
corpus = st.sidebar.radio("Corpus", ["Unsplash"])
if st.sidebar.button("Submit"):
if len(query) > 0:
results = image_search(query, corpus)
clicked = clickable_images(
[result[0] for result in results],
titles=[result[1] for result in results],
div_style={
"display": "flex",
"justify-content": "center",
"flex-wrap": "wrap",
},
img_style={"margin": "2px", "height": "200px"},
)
if clicked >= 0:
change_query = False
if "last_clicked" not in st.session_state:
change_query = True
else:
if clicked != st.session_state["last_clicked"]:
change_query = True
if change_query:
st.session_state["query"] = f"[{corpus}:{results[clicked][2]}]"
st.experimental_rerun()
if __name__ == "__main__":
main()
|