Spaces:
Running
Running
| import streamlit as st | |
| import pandas as pd | |
| import numpy as np | |
| from PIL import Image | |
| import pickle | |
| import tokenizers | |
| import torch | |
| from transformers import ( | |
| CLIPModel, | |
| AutoProcessor | |
| ) | |
| import streamlit.components.v1 as components | |
| import base64 | |
| def render_svg(svg_filename): | |
| with open(svg_filename,"r") as f: | |
| lines = f.readlines() | |
| svg=''.join(lines) | |
| """Renders the given svg string.""" | |
| b64 = base64.b64encode(svg.encode('utf-8')).decode("utf-8") | |
| html = r'<img src="data:image/svg+xml;base64,%s"/>' % b64 | |
| st.write(html, unsafe_allow_html=True) | |
| def load_path_clip(): | |
| model = CLIPModel.from_pretrained("vinid/plip") | |
| processor = AutoProcessor.from_pretrained("vinid/plip") | |
| return model, processor | |
| def init(): | |
| with open('data/twitter.asset', 'rb') as f: | |
| data = pickle.load(f) | |
| meta = data['meta'].reset_index(drop=True) | |
| image_embedding = data['image_embedding'] | |
| text_embedding = data['text_embedding'] | |
| print(meta.shape, image_embedding.shape) | |
| validation_subset_index = meta['source'].values == 'Val_Tweets' | |
| return meta, image_embedding, text_embedding, validation_subset_index | |
| def embed_images(model, images, processor): | |
| inputs = processor(images=images) | |
| pixel_values = torch.tensor(np.array(inputs["pixel_values"])) | |
| with torch.no_grad(): | |
| embeddings = model.get_image_features(pixel_values=pixel_values) | |
| return embeddings | |
| def embed_texts(model, texts, processor): | |
| inputs = processor(text=texts, padding="longest") | |
| input_ids = torch.tensor(inputs["input_ids"]) | |
| attention_mask = torch.tensor(inputs["attention_mask"]) | |
| with torch.no_grad(): | |
| embeddings = model.get_text_features( | |
| input_ids=input_ids, attention_mask=attention_mask | |
| ) | |
| return embeddings | |
| def app(): | |
| st.title('Text to Image Retrieval') | |
| st.markdown('#### A pathology image search engine that geos from texts to images.') | |
| col1, col2 = st.columns([1,1]) | |
| with col1: | |
| st.markdown("The text-to-image retrieval system can serve as an image search engine, enabling users to match images from multiple queries and retrieve the most relevant image based on a sentence description. This generic system can comprehend semantic and interrelated knowledge, such as “Breast tumor surrounded by fat”.") | |
| st.markdown("Unlike searching keywords and sentences from Google and indirectly matching the images from the target text, our proposed pathology image retrieval allows direct comparison between input sentences and images.") | |
| with col2: | |
| render_svg("resources/SVG/Asset 54.svg") | |
| meta, image_embedding, text_embedding, validation_subset_index = init() | |
| model, processor = load_path_clip() | |
| st.markdown('### Search') | |
| st.markdown('How to use this: first of all, select a dataset on which to do retrieval.\n' | |
| 'Then, either select a predefined search query or input one yourself.') | |
| col1, col2 = st.columns(2) | |
| with col1: | |
| data_options = ["All Twitter Data (03/21/2006 — 01/15/2023)", | |
| "Validation Twitter data (11/16/2022 — 01/15/2023)"] | |
| st.selectbox( | |
| "Dataset", | |
| key="datapool", | |
| options=data_options, | |
| ) | |
| with col2: | |
| retrieval_options = ["Image only", | |
| "Text and image (beta)", | |
| ] | |
| st.radio( | |
| "Similarity calcuation 👉", | |
| key="calculation_option", | |
| options=retrieval_options, | |
| ) | |
| col1, col2 = st.columns(2) | |
| with col1: | |
| # Create selectbox | |
| examples = ['Breast tumor surrounded by fat', | |
| 'HER2+ breast tumor', | |
| 'Colorectal cancer tumor on epithelium', | |
| 'An image of endometrium epithelium', | |
| 'Breast cancer DCIS', | |
| 'Papillary carcinoma in breast tissue', | |
| ] | |
| query_1 = st.selectbox("Select an example", options=examples) | |
| col1_submit = True | |
| with col2: | |
| form = st.form(key='my_form') | |
| query_2 = form.text_input(label='Or input your custom query:') | |
| submit_button = form.form_submit_button(label='Submit') | |
| if submit_button: | |
| col1_submit = False | |
| if col1_submit: | |
| query = query_1 | |
| else: | |
| query = query_2 | |
| input_text = embed_texts(model, [query], processor)[0].detach().cpu().numpy() | |
| input_text = input_text/np.linalg.norm(input_text) | |
| # Sort IDs by cosine-similarity from high to low | |
| if st.session_state.calculation_option == retrieval_options[0]: # Image only | |
| similarity_scores = input_text.dot(image_embedding.T) | |
| else: # Text and Image | |
| similarity_scores_i = input_text.dot(image_embedding.T) | |
| similarity_scores_t = input_text.dot(text_embedding.T) | |
| similarity_scores_i = similarity_scores_i / np.max(similarity_scores_i) | |
| similarity_scores_t = similarity_scores_t / np.max(similarity_scores_t) | |
| similarity_scores = (similarity_scores_i + similarity_scores_t) / 2 | |
| ############################################################ | |
| # Get top results | |
| ############################################################ | |
| topn = 5 | |
| df = pd.DataFrame(np.c_[np.arange(len(meta)), similarity_scores, meta['weblink'].values], columns = ['idx', 'score', 'twitterlink']) | |
| if st.session_state.datapool == data_options[1]: #Use val twitter data | |
| df = df.loc[validation_subset_index,:] | |
| df = df.sort_values('score', ascending=False) | |
| df = df.drop_duplicates(subset=['twitterlink']) | |
| best_id_topk = df['idx'].values[:topn] | |
| target_scores = df['score'].values[:topn] | |
| target_weblinks = df['twitterlink'].values[:topn] | |
| ############################################################ | |
| # Display results | |
| ############################################################ | |
| text = '<font size="4">Your input query: <span style="background-color: rgb(230,230,230);"><b>%s</b></span>' % query + \ | |
| ' (Try search it directly on [Twitter](https://twitter.com/search?q=%s&src=typed_query) or [Google](https://www.google.com/search?q=%s))</font>' % (query.replace(' ', '%20'), query.replace(' ', '+')) | |
| st.markdown(text, unsafe_allow_html=True) | |
| st.markdown('#### Top 5 results:') | |
| topk_options = ['1st', '2nd', '3rd', '4th', '5th'] | |
| tab = {} | |
| tab[0], tab[1], tab[2] = st.columns(3) | |
| for i in [0,1,2]: | |
| with tab[i]: | |
| topn_value = i | |
| topn_txt = topk_options[i] | |
| st.caption(f'The {topn_txt} relevant image (similarity = {target_scores[topn_value]:.4f})') | |
| components.html(''' | |
| <blockquote class="twitter-tweet"> | |
| <a href="%s"></a> | |
| </blockquote> | |
| <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"> | |
| </script> | |
| ''' % target_weblinks[topn_value], | |
| height=600) | |
| tab[3], tab[4], tab[5] = st.columns(3) | |
| for i in [3,4]: | |
| with tab[i]: | |
| topn_value = i | |
| topn_txt = topk_options[i] | |
| st.caption(f'The {topn_txt} relevant image (similarity = {target_scores[topn_value]:.4f})') | |
| components.html(''' | |
| <blockquote class="twitter-tweet"> | |
| <a href="%s"></a> | |
| </blockquote> | |
| <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"> | |
| </script> | |
| ''' % target_weblinks[topn_value], | |
| height=800) | |
| st.markdown("""---""") | |
| st.markdown('Disclaimer') | |
| st.caption('Please be advised that this function has been developed in compliance with the Twitter policy of data usage and sharing. It is important to note that the results obtained from this function are not intended to constitute medical advice or replace consultation with a qualified medical professional. The use of this function is solely at your own risk and should be consistent with applicable laws, regulations, and ethical considerations. We do not warrant or guarantee the accuracy, completeness, suitability, or usefulness of this function for any particular purpose, and we hereby disclaim any liability arising from any reliance placed on this function or any results obtained from its use. If you wish to review the original Twitter post, you should access the source page directly on Twitter.') | |
| st.markdown('Privacy statement') | |
| st.caption('In accordance with the privacy and control policy of Twitter, we hereby declared that the data redistributed by us shall only comprise of Tweet IDs. The Tweet IDs will be employed to establish a linkage with the original Twitter post, as long as the original post is still accessible. The hyperlink will cease to function if the user deletes the original post. It is important to note that all tweets displayed on our service have already been classified as non-sensitive by Twitter. It is strictly prohibited to redistribute any content apart from the Tweet IDs. Any distribution carried out must adhere to the laws and regulations applicable in your jurisdiction, including export control laws and embargoes.') | |