webplip / text2image.py
vinid's picture
new features added (#7)
7ee7b26
raw history blame
No virus
8.31 kB
import streamlit as st
import pandas as pd
import numpy as np
from PIL import Image
import requests
import pickle
import tokenizers
from io import BytesIO
import torch
from transformers import (
VisionTextDualEncoderModel,
AutoFeatureExtractor,
AutoTokenizer,
CLIPModel,
AutoProcessor
)
import streamlit.components.v1 as components
@st.cache(
hash_funcs={
torch.nn.parameter.Parameter: lambda _: None,
tokenizers.Tokenizer: lambda _: None,
tokenizers.AddedToken: lambda _: None
}
)
def load_path_clip():
model = CLIPModel.from_pretrained("vinid/plip")
processor = AutoProcessor.from_pretrained("vinid/plip")
return model, processor
@st.cache
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 correlate texts directly with 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 โ€œBreast tumor surrounded by fatโ€.")
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:
fig1 = Image.open('resources/4x/image_retrieval.png')
st.image(fig1, caption='Image retrieval from text', width=400, output_format='png')
meta, image_embedding, text_embedding, validation_subset_index = init()
model, processor = load_path_clip()
st.markdown('#### Demo')
col1, col2 = st.columns(2)
with col1:
data_options = ["All twitter data (03/21/2006 โ€” 01/15/2023)",
"Twitter validation data (11/16/2022 โ€” 01/15/2023)"]
st.radio(
"Choose dataset for image retrieval ๐Ÿ‘‰",
key="datapool",
options=data_options,
)
with col2:
retrieval_options = ["Image only",
"Text and image (beta)",
]
st.radio(
"Similarity calcuation Mapping input with ๐Ÿ‘‰",
key="calculation_option",
options=retrieval_options,
)
col1, col2 = st.columns(2)
#query = st.text_input('Search Query', '')
col1_submit = False
show = False
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("Please select an example query", options=examples)
#st.info(f":white_check_mark: The written option is {query_1} ")
col1_submit = True
show = 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
show = True
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)
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
############################################################
st.markdown('Your input query: %s' % query)
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=800)
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('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.')