import enum from turtle import onclick import streamlit as st import numpy as np import base64 from io import BytesIO import requests from multilingual_clip import pt_multilingual_clip from transformers import CLIPTokenizerFast, AutoTokenizer import torch import logging from os import environ environ['TOKENIZERS_PARALLELISM'] = 'true' from myscaledb import Client DB_NAME = "mqdb_demo.unsplash_25k_clip_indexer" MODEL_ID = 'M-CLIP/XLM-Roberta-Large-Vit-B-32' DIMS = 512 # Ignore some bad links (broken in the dataset already) BAD_IDS = {'9_9hzZVjV8s', 'RDs0THr4lGs', 'vigsqYux_-8', 'rsJtMXn3p_c', 'AcG-unN00gw', 'r1R_0ZNUcx0'} @st.experimental_singleton(show_spinner=False) def init_clip(): """ Initialize CLIP Model Returns: Tokenizer: CLIPTokenizerFast (which convert words into embeddings) """ clip = pt_multilingual_clip.MultilingualCLIP.from_pretrained(MODEL_ID) tokenizer = AutoTokenizer.from_pretrained(MODEL_ID) return tokenizer, clip @st.experimental_singleton(show_spinner=False) def init_db(): """ Initialize the Database Connection Returns: meta_field: Meta field that records if an image is viewed or not client: Database connection object """ client = Client(url=st.secrets["DB_URL"], user=st.secrets["USER"], password=st.secrets["PASSWD"]) resp = requests.get('https://www.google.com/') assert resp.status_code == 200 # We can check if the connection is alive assert client.is_alive() meta_field = {} return meta_field, client @st.experimental_singleton(show_spinner=False) def init_query_num(): print("init query_num") return 0 def query(xq, top_k=10): """ Query TopK matched w.r.t a given vector Args: xq (numpy.ndarray or list of floats): Query vector top_k (int, optional): Number of matched vectors. Defaults to 10. Returns: matches: list of Records object. Keys referrring to selected columns """ attempt = 0 xq = xq / np.linalg.norm(xq) while attempt < 3: try: xq_s = f"[{', '.join([str(float(fnum)) for fnum in list(xq)])}]" print('Excluded pre:', st.session_state.meta) if len(st.session_state.meta) > 0: exclude_list = ','.join([f'\'{i}\'' for i, v in st.session_state.meta.items() if v >= 1]) print("Excluded:", exclude_list) # Using PREWHERE allows you to do column filter before vector search xc = st.session_state.index.fetch(f"SELECT id, url, vector,\ distance('topK={top_k}')(vector, {xq_s}) AS dist\ FROM {DB_NAME} PREWHERE id NOT IN ({exclude_list})") else: xc = st.session_state.index.fetch(f"SELECT id, url, vector,\ distance('topK={top_k}')(vector, {xq_s}) AS dist\ FROM {DB_NAME}") # real_xc = st.session_state.index.fetch(f"SELECT id, url, vector,\ # 1 - arraySum(arrayMap((x, y) -> x * y, {xq_s}, vector)) AS dist\ # FROM {DB_NAME} ORDER BY dist LIMIT {top_k}") # FIXME: This is causing freezing on DB real_xc = st.session_state.index.fetch(f"SELECT id, url, vector,\ distance('topK={top_k}')(vector, {xq_s}) AS dist\ FROM {DB_NAME}") top_k = real_xc xc = [xi for xi in xc if xi['id'] not in st.session_state.meta or \ st.session_state.meta[xi['id']] < 1] logging.info(f'{len(xc)} records returned, {[_i["id"] for _i in xc]}') matches = xc break except Exception as e: # force reload if we have trouble on connections or something else logging.warning(str(e)) _, st.session_state.index = init_db() attempt += 1 matches = [] if len(matches) == 0: logging.error(f"No matches found for '{DB_NAME}'") return matches, top_k @st.experimental_singleton(show_spinner=False) def init_random_query(): xq = np.random.rand(DIMS).tolist() return xq, xq.copy() class Classifier: """ Zero-shot Classifier This Classifier provides proxy regarding to the user's reaction to the probed images. The proxy will replace the original query vector generated by prompted vector and finally give the user a satisfying retrieval result. This can be commonly seen in a recommendation system. The classifier will recommend more precise result as it accumulating user's activity. """ def __init__(self, xq: list): # initialize model with DIMS input size and 1 output # note that the bias is ignored, as we only focus on the inner product result self.model = torch.nn.Linear(DIMS, 1, bias=False) # convert initial query `xq` to tensor parameter to init weights init_weight = torch.Tensor(xq).reshape(1, -1) self.model.weight = torch.nn.Parameter(init_weight) # init loss and optimizer self.loss = torch.nn.BCEWithLogitsLoss() self.optimizer = torch.optim.SGD(self.model.parameters(), lr=0.1) def fit(self, X: list, y: list, iters: int = 5): # convert X and y to tensor X = torch.Tensor(X) y = torch.Tensor(y).reshape(-1, 1) for i in range(iters): # zero gradients self.optimizer.zero_grad() # Normalize the weight before inference # This will constrain the gradient or you will have an explosion on query vector self.model.weight.data = self.model.weight.data / torch.norm(self.model.weight.data, p=2, dim=-1) # forward pass out = self.model(X) # compute loss loss = self.loss(out, y) # backward pass loss.backward() # update weights self.optimizer.step() def get_weights(self): xq = self.model.weight.detach().numpy()[0].tolist() return xq def prompt2vec(prompt: str): """ Convert prompt into a computational vector Args: prompt (str): Text to be tokenized Returns: xq: vector from the tokenizer, representing the original prompt """ # inputs = tokenizer(prompt, return_tensors='pt') # out = clip.get_text_features(**inputs) out = clip.forward(prompt, tokenizer) xq = out.squeeze(0).cpu().detach().numpy().tolist() return xq def pil_to_bytes(img): """ Convert a Pillow image into base64 Args: img (PIL.Image): Pillow (PIL) Image Returns: img_bin: image in base64 format """ with BytesIO() as buf: img.save(buf, format='jpeg') img_bin = buf.getvalue() img_bin = base64.b64encode(img_bin).decode('utf-8') return img_bin def card(i, url): return f'' def card_with_conf(i, conf, url): conf = "%.4f"%(conf) return f'Relevance: {conf}' def get_top_k(xq, top_k=9): """ wrapper function for query Args: xq (numpy.ndarray or list of floats): Query vector top_k (int, optional): Number of returned vectors. Defaults to 9. Returns: matches: See `query()` """ matches = query( xq, top_k=top_k ) return matches def tune(X, y, iters=2): """ Train the Zero-shot Classifier Args: X (numpy.ndarray): Input vectors (retreived vectors) y (list of floats or numpy.ndarray): Scores given by user iters (int, optional): iterations of updates to be run """ # train the classifier st.session_state.clf.fit(X, y, iters=iters) # extract new vector st.session_state.xq = st.session_state.clf.get_weights() def refresh_index(): """ Clean the session """ del st.session_state["meta"] st.session_state.meta = {} st.session_state.query_num = 0 logging.info(f"Refresh for '{st.session_state.meta}'") init_db.clear() # refresh session states st.session_state.meta, st.session_state.index = init_db() del st.session_state.clf, st.session_state.xq def calc_dist(): xq = np.array(st.session_state.xq) orig_xq = np.array(st.session_state.orig_xq) return np.linalg.norm(xq - orig_xq) def submit(): """ Tune the model w.r.t given score from user. """ st.session_state.query_num += 1 matches = st.session_state.matches velocity = 1 #st.session_state.velocity scores = {} states = [ st.session_state[f"input{i}"] for i in range(len(matches)) ] for i, match in enumerate(matches): scores[match['id']] = float(states[i]) # reset states to 1.0 for i in range(len(matches)): st.session_state[f"input{i}"] = 1.0 # get training data and labels X = list([match['vector'] for match in matches]) y = [v for v in list(scores.values())] tune(X, y, iters=int(st.session_state.iters)) # update record metadata after training for match in matches: st.session_state.meta[match['id']] = 1 logging.info(f"Exclude List: {st.session_state.meta}") def delete_element(element): del element st.markdown(""" """, unsafe_allow_html=True) messages = [ f""" Find most relevant examples from a large visual dataset by combining text query and few-shot learning. """, f""" Then then you can adjust the weight on each image. Those weights should **represent how much it can meet your preference**. You can either choose the images that match your prompt or change your mind. You might notice that there is a iteration slide bar on the top of all retrieved images. This will control the speed of changes on vectors. More **iterations** will change the vector faster while lower values on **iterations** will make the retrieval smoother. """, f""" This example will manage to train a classifier to distinguish between samples you want and samples you don't want. By initializing the weight from prompt, you can get a good enough classifier to cluster images you want to search. If you think the result is not as perfect as you expected, you can also supervise the classifer with **Relevance** annotation. If you cannot see any difference in Top-K Retrieved results, try to enlarge **Number of Iteration** """, # TODO @ fangruil: fill the link with our tech blog f""" The app uses the [MyScale](http://mqdb.page.moqi.ai/mqdb-docs/) to store and query images using vector search. All images are sourced from the [Unsplash Lite dataset](https://unsplash-datasets.s3.amazonaws.com/lite/latest/unsplash-research-dataset-lite-latest.zip) and encoded using [OpenAI's CLIP](https://huggingface.co/openai/clip-vit-base-patch32). We explain how it all works [here](). """ ] with st.spinner("Connecting DB..."): st.session_state.meta, st.session_state.index = init_db() with st.spinner("Loading Models..."): # Initialize CLIP model if 'xq' not in st.session_state: tokenizer, clip = init_clip() st.session_state.query_num = 0 if 'xq' not in st.session_state: # If it's a fresh start if st.session_state.query_num < len(messages): msg = messages[st.session_state.query_num] else: msg = messages[-1] # Basic Layout with st.container(): st.title("Visual Dataset Explorer") start = [st.empty(), st.empty(), st.empty(), st.empty(), st.empty()] start[0].info(msg) prompt = start[1].text_input("Prompt:", value="", placeholder="Examples: a photo of white dogs, cats in the snow, a house by the lake") start[2].markdown( '

Don\'t know what to search? Try Random!

', unsafe_allow_html=True) with start[3]: col = st.columns(8) prompt_xq = col[6].button("Prompt", disabled=len(prompt) == 0) random_xq = col[7].button("Random", disabled=len(prompt) != 0) if random_xq: # Randomly pick a vector to query xq, orig_xq = init_random_query() st.session_state.xq = xq st.session_state.orig_xq = orig_xq _ = [elem.empty() for elem in start] elif prompt_xq: print(f"Input prompt is {prompt}") # Tokenize the vectors xq = prompt2vec(prompt) st.session_state.xq = xq st.session_state.orig_xq = xq _ = [elem.empty() for elem in start] if 'xq' in st.session_state: # If it is not a fresh start if st.session_state.query_num+1 < len(messages): msg = messages[st.session_state.query_num+1] else: msg = messages[-1] # initialize classifier if 'clf' not in st.session_state: st.session_state.clf = Classifier(st.session_state.xq) # if we want to display images we end up here st.info(msg) # first retrieve images from pinecone st.session_state.matches, st.session_state.top_k = get_top_k(st.session_state.clf.get_weights(), top_k=9) with st.container(): with st.sidebar: with st.container(): st.header("Top K Nearest in Database") for i, k in enumerate(st.session_state.top_k): url = k["url"] url += "?q=75&fm=jpg&w=200&fit=max" if k["id"] not in BAD_IDS: disabled = False else: disable = True dist = np.matmul(st.session_state.clf.get_weights() / np.linalg.norm(st.session_state.clf.get_weights()), np.array(k["vector"]).T) st.markdown(card_with_conf(i, dist, url), unsafe_allow_html=True) # once retrieved, display them alongside checkboxes in a form with st.form("batch", clear_on_submit=False): st.session_state.iters = st.slider("Number of Iterations to Update", min_value=0, max_value=10, step=1, value=2) col = st.columns([1,9]) col[0].form_submit_button("Train!", on_click=submit) col[1].form_submit_button("Choose a new prompt", on_click=refresh_index) # we have three columns in the form cols = st.columns(3) for i, match in enumerate(st.session_state.matches): # find good url url = match["url"] url += "?q=75&fm=jpg&w=200&fit=max" if match["id"] not in BAD_IDS: disabled = False else: disable = True # the card shows an image and a checkbox cols[i%3].markdown(card(i, url), unsafe_allow_html=True) # we access the values of the checkbox via st.session_state[f"input{i}"] cols[i%3].slider( "Relevance", min_value=0.0, max_value=1.0, value=1.0, step=0.05, key=f"input{i}", disabled=disabled )