import streamlit as st import streamlit.components.v1 from transformers import CLIPProcessor, CLIPModel from PIL import Image import numpy as np import pandas as pd USE_DESCRIPTION_AS_LABEL = False NUM_OTHER_TYPES = 4 MAX_FILES = 3 def _get_probs(image, text): inputs = processor(text=text, images=image, return_tensors="pt", padding=True) outputs = model(**inputs) logits_per_image = outputs.logits_per_image probs = logits_per_image.softmax(dim=1).detach().numpy().flatten() return probs smaller_model = "openai/clip-vit-base-patch32" model = CLIPModel.from_pretrained(smaller_model) processor = CLIPProcessor.from_pretrained(smaller_model) themes = pd.read_csv("wedding_themes.csv") themes = themes.query("use=='Y'") st.title("What's your wedding persona?") st.subheader("") filenames = st.file_uploader(f"Modern chic? Fall foliage? Upload up to {MAX_FILES} photos that have aesthetics that you love (photos don't have to be taken at a wedding).", accept_multiple_files=True) if len(filenames) > MAX_FILES: st.warning(f"Only {MAX_FILES} images are allowed!") if 0 < len(filenames) <= MAX_FILES and st.button("Process"): labels = themes.theme.tolist() descriptions = themes.description.tolist() all_probs = [] images = [] for filename in filenames: image = Image.open(filename) if USE_DESCRIPTION_AS_LABEL: probs = _get_probs(image, descriptions) else: labels_with_wedding = [(x + " wedding").lower() for x in labels] probs = _get_probs(image, labels_with_wedding) all_probs.append(probs) images.append(image) # Calculate the total probability and write results. combined_probs = np.sum(all_probs, axis=0) sorted_args = np.argsort(combined_probs)[::-1][:(1+NUM_OTHER_TYPES)].tolist() top_labels = np.array(labels)[sorted_args].tolist() top_description = descriptions[sorted_args[0]] st.subheader(f"Your result: {top_labels[0]}") st.write(f"{top_description} \n\n Other personas you might like: {', '.join(top_labels[1:])}") # Display the uploaded images. cols = st.columns(len(filenames)) for col, image in zip(cols, images): col.image(image)