from html import escape import re import streamlit as st import pandas as pd, numpy as np import torch from transformers import CLIPProcessor, CLIPModel from st_clickable_images import clickable_images MODEL_NAMES = [ # "base-patch32", # "base-patch16", # "large-patch14", "large-patch14-336" ] @st.cache(allow_output_mutation=True) def load(): df = {0: pd.read_csv("data.csv"), 1: pd.read_csv("data2.csv")} models = {} processors = {} embeddings = {} for name in MODEL_NAMES: models[name] = CLIPModel.from_pretrained(f"openai/clip-vit-{name}").eval() processors[name] = CLIPProcessor.from_pretrained(f"openai/clip-vit-{name}") embeddings[name] = { 0: np.load(f"embeddings-vit-{name}.npy"), 1: np.load(f"embeddings2-vit-{name}.npy"), } for k in [0, 1]: embeddings[name][k] = embeddings[name][k] / np.linalg.norm( embeddings[name][k], axis=1, keepdims=True ) return models, processors, df, embeddings models, processors, df, embeddings = load() source = {0: "\nSource: Unsplash", 1: "\nSource: The Movie Database (TMDB)"} def compute_text_embeddings(list_of_strings, name): inputs = processors[name](text=list_of_strings, return_tensors="pt", padding=True) with torch.no_grad(): result = models[name].get_text_features(**inputs).detach().numpy() return result / np.linalg.norm(result, axis=1, keepdims=True) def image_search(query, corpus, name, 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[name][k2][idx : idx + 1, :] ) if len(remainder) > 0: positive_embeddings = concatenate_embeddings( positive_embeddings, compute_text_embeddings([remainder], name) ) else: positive_embeddings = concatenate_embeddings( positive_embeddings, compute_text_embeddings([positive_query], name) ) dot_product = embeddings[name][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, name) dot_product2 = embeddings[name][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 ] description = """ # 意味による画像検索 **検索語を入力してから Enter キーを押してください** *OpenAI の [CLIP](https://openai.com/blog/clip/) model, 🤗 Hugging Face's [transformers library](https://huggingface.co/transformers/), [Streamlit](https://streamlit.io/), [Unsplash](https://unsplash.com/) の 25k images と [The Movie Database (TMDB)](https://www.themoviedb.org/) の 8k images を使用して構築しています。* *Vladimir Haltakov の [Unsplash Image Search](https://github.com/haltakov/natural-language-image-search) と Travis Hoppe の [Alph, The Sacred River](https://github.com/thoppe/alph-the-sacred-river)  に触発されました。* """ howto = """ - 画像をクリックすると、それをクエリとして使用し、類似画像を検索できます。 - 複数の検索語を組み合わせることができます(区切り文字として「**;**」を使用します)。 - 検索語に 「**EXCLUDING**」 が含まれている場合、その右側の部分が否定クエリとして使用されます。 """ div_style = { "display": "flex", "justify-content": "center", "flex-wrap": "wrap", } def main(): st.markdown( """ """, unsafe_allow_html=True, ) st.sidebar.markdown(description) with st.sidebar.expander("高度な使用方法"): st.markdown(howto) # mode = st.sidebar.selectbox( # "", ["Results for ViT-L/14@336px", "Comparison of 2 models"], index=0 # ) _, c, _ = st.columns((1, 3, 1)) if "query" in st.session_state: query = c.text_input("", value=st.session_state["query"]) else: query = c.text_input("", value="clouds at sunset") corpus = st.radio("", ["Unsplash", "Movies"]) models_dict = { "ViT-B/32 (quicker)": "base-patch32", "ViT-B/16 (average)": "base-patch16", # "ViT-L/14 (slow)": "large-patch14", "ViT-L/14@336px (slower)": "large-patch14-336", } if False: # "Comparison" in mode: c1, c2 = st.columns((1, 1)) selection1 = c1.selectbox("", models_dict.keys(), index=0) selection2 = c2.selectbox("", models_dict.keys(), index=2) name1 = models_dict[selection1] name2 = models_dict[selection2] else: name1 = MODEL_NAMES[-1] if len(query) > 0: results1 = image_search(query, corpus, name1) if False: # "Comparison" in mode: with c1: clicked1 = clickable_images( [result[0] for result in results1], titles=[result[1] for result in results1], div_style=div_style, img_style={"margin": "2px", "height": "150px"}, key=query + corpus + name1 + "1", ) results2 = image_search(query, corpus, name2) with c2: clicked2 = clickable_images( [result[0] for result in results2], titles=[result[1] for result in results2], div_style=div_style, img_style={"margin": "2px", "height": "150px"}, key=query + corpus + name2 + "2", ) else: clicked1 = clickable_images( [result[0] for result in results1], titles=[result[1] for result in results1], div_style=div_style, img_style={"margin": "2px", "height": "200px"}, key=query + corpus + name1 + "1", ) clicked2 = -1 if clicked2 >= 0 or clicked1 >= 0: change_query = False if "last_clicked" not in st.session_state: change_query = True else: if max(clicked2, clicked1) != st.session_state["last_clicked"]: change_query = True if change_query: if clicked1 >= 0: st.session_state["query"] = f"[{corpus}:{results1[clicked1][2]}]" # elif clicked2 >= 0: # st.session_state["query"] = f"[{corpus}:{results2[clicked2][2]}]" st.experimental_rerun() if __name__ == "__main__": main()