app.py

import os
os.system("pip uninstall -y gradio")
os.system("pip install --upgrade gradio")
os.system("pip install datamapplot==0.3.0")
os.system("pip install numba==0.59.1")
os.system("pip install umap-learn==0.5.6")
os.system("pip install pynndescent==0.5.12")



import spaces


from pathlib import Path
from fastapi import FastAPI
from fastapi.staticfiles import StaticFiles
import uvicorn
import gradio as gr
from datetime import datetime
import sys



gr.set_static_paths(paths=["static/"])



# create a FastAPI app
app = FastAPI()

# create a static directory to store the static files
static_dir = Path('./static')
static_dir.mkdir(parents=True, exist_ok=True)

# mount FastAPI StaticFiles server
app.mount("/static", StaticFiles(directory=static_dir), name="static")

# Gradio stuff



import datamapplot
import numpy as np
import requests
import io
import pandas as pd
from pyalex import Works, Authors, Sources, Institutions, Concepts, Publishers, Funders
from itertools import chain
from compress_pickle import load, dump



from transformers import AutoTokenizer
from adapters import AutoAdapterModel
import torch
from tqdm import tqdm
from numba.typed import List
import pickle
import pynndescent
import umap














def query_records(search_term):
    def invert_abstract(inv_index):
        if inv_index is not None:
            l_inv = [(w, p) for w, pos in inv_index.items() for p in pos]
            return " ".join(map(lambda x: x[0], sorted(l_inv, key=lambda x: x[1])))
        else:
            return ' '

    def get_pub(x):
        try: 
            source = x['source']['display_name']
            if source not in ['parsed_publication','Deleted Journal']:
                return source
            else: 
                return ' '
        except:
            return ' '

    # Fetch records based on the search term in the abstract!
    query = Works().search([search_term])
    query_length = Works().search([search_term]).count()
    
    records = []
    #total_pages = (query_length + 199) // 200  # Calculate total number of pages
    progress=gr.Progress()
    
    for i, record in progress.tqdm(enumerate(chain(*query.paginate(per_page=200)))):
        records.append(record)
        
        # Calculate progress from 0 to 0.1
        #achieved_progress = min(0.1, (i + 1) / query_length * 0.1)
        
        # Update progress bar
        #progress(achieved_progress, desc="Getting queried data...")


    
    records_df = pd.DataFrame(records)
    records_df['abstract'] = [invert_abstract(t) for t in records_df['abstract_inverted_index']]

    records_df['parsed_publication'] = [get_pub(x) for x in records_df['primary_location']]

    
    records_df['parsed_publication'] = records_df['parsed_publication'].fillna(' ')
    records_df['abstract'] = records_df['abstract'].fillna(' ')
    records_df['title'] = records_df['title'].fillna(' ')

    
    return records_df




################# Setting up the model for specter2 embeddings ###################



#device = torch.device("mps" if torch.backends.mps.is_available() else "cuda")
#print(f"Using device: {device}")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

tokenizer = AutoTokenizer.from_pretrained('allenai/specter2_aug2023refresh_base')
model = AutoAdapterModel.from_pretrained('allenai/specter2_aug2023refresh_base')


@spaces.GPU(duration=60)
def create_embeddings(texts_to_embedd):
    # Set up the device


    print(len(texts_to_embedd))

    # Load the proximity adapter and activate it
    model.load_adapter("allenai/specter2_aug2023refresh", source="hf", load_as="proximity", set_active=True)
    model.set_active_adapters("proximity")
 
    model.to(device)

    def batch_generator(data, batch_size):
        """Yield consecutive batches of data."""
        for i in range(0, len(data), batch_size):
            yield data[i:i + batch_size]

    
    def encode_texts(texts, device, batch_size=16):
        """Process texts in batches and return their embeddings."""
        model.eval()
        with torch.no_grad():
            all_embeddings = []
            count = 0
            for batch in tqdm(batch_generator(texts, batch_size)):
                inputs = tokenizer(batch, padding=True, truncation=True, return_tensors="pt", max_length=512).to(device)
                outputs = model(**inputs)
                embeddings = outputs.last_hidden_state[:, 0, :]  # Taking the [CLS] token representation
                
                all_embeddings.append(embeddings.cpu())  # Move to CPU to free GPU memory
                #torch.mps.empty_cache()  # Clear cache to free up memory
                if count == 100:
                    #torch.mps.empty_cache()
                    torch.cuda.empty_cache()
                    count = 0
                
                count +=1
            
            all_embeddings = torch.cat(all_embeddings, dim=0)
        return all_embeddings

    # Concatenate title and abstract
    embeddings = encode_texts(texts_to_embedd, device, batch_size=32).cpu().numpy()  # Process texts in batches of 10

    return embeddings





def predict(text_input, sample_size_slider, reduce_sample_checkbox, progress=gr.Progress()):

    
    # get data.
    records_df = query_records(text_input,progress=progress)
    if reduce_sample_checkbox:
        records_df = records_df.sample(sample_size_slider)
    print(records_df)
    
    
    progress(0.3, desc="Embedding Data...")
    texts_to_embedd = [title + tokenizer.sep_token + publication + tokenizer.sep_token  + abstract for title, publication, abstract in zip(records_df['title'],records_df['parsed_publication'], records_df['abstract'])]
    
    embeddings = create_embeddings(texts_to_embedd)
    print(embeddings)

    progress(0.5, desc="Project into UMAP-embedding...")
    umap_embeddings = mapper.transform(embeddings)
    records_df[['x','y']] = umap_embeddings

    basedata_df['color'] = '#ced4d211'
    records_df['color'] = '#f98e31'

    progress(0.6, desc="Set up data...")

    stacked_df = pd.concat([basedata_df,records_df], axis=0, ignore_index=True)
    stacked_df = stacked_df.fillna("Unlabelled")
    stacked_df = stacked_df.reset_index(drop=True)
    print(stacked_df)

    extra_data = pd.DataFrame(stacked_df['doi'])
    
    
    file_name = f"{datetime.utcnow().strftime('%s')}.html"
    file_path = static_dir / file_name
    print(file_path)
 
   #
    
    progress(0.7, desc="Plotting...")

    custom_css = """
    
       
            #title-container {
                background: #edededaa;
                border-radius: 2px;
    
                box-shadow: 2px 3px 10px #aaaaaa00;
            }
            
    
            
        #search-container {
            position: fixed !important;
            top: 20px !important;
            right: 20px !important;
            left: auto !important;
            width: 200px !important;
            z-index: 9999 !important;
        }
        
        #search {
         //   padding: 8px 8px !important;
         //   border: none !important;
         //   border-radius: 20px !important;
            background-color: #ffffffaa !important;
            font-family: 'Roboto Condensed', sans-serif !important;
            font-size: 14px;
         //   box-shadow: 0 0px 0px #aaaaaa00 !important;
        }
    
            
    
    """

    
    plot = datamapplot.create_interactive_plot(
        stacked_df[['x','y']].values,
        np.array(stacked_df['cluster_1_labels']),np.array(stacked_df['cluster_2_labels']),np.array(stacked_df['cluster_3_labels']),
        hover_text=[str(row['title']) for ix, row in stacked_df.iterrows()],
        marker_color_array=stacked_df['color'],

        use_medoids=True,
        width=1000,
        height=1000,
    #    title='The Science of <span style="color:#ab0b00;"> Consciousness </span>',
    #    sub_title=f'<div style="margin-top:20px;"> Large sample, n={len(dataset_df_filtered)}, embeddings with specter 2 & UMAP, labels: Claude 3.5 Sonnet </div>',
        point_radius_min_pixels=1,
        text_outline_width=5,
        point_hover_color='#5e2784',
        point_radius_max_pixels=7,
        color_label_text=False,
        font_family="Roboto Condensed",
        font_weight=700,
        tooltip_font_weight=600,
        tooltip_font_family="Roboto Condensed",
        extra_point_data=extra_data,
        on_click="window.open(`{doi}`)",
        custom_css=custom_css,
        initial_zoom_fraction=.8,
        enable_search=True)




    

    progress(0.9, desc="Saving plot...")
    plot.save(file_path)

    progress(1.0, desc="Done!")
    iframe = f"""<iframe src="/static/{file_name}" width="100%" height="500px"></iframe>"""
    link = f'<a href="/static/{file_name}" target="_blank">{file_name}</a>'
    return link, iframe






################ MAIN BLOCK #####################



# with gr.Blocks() as block:
#     gr.Markdown("""
# ## Mapping OpenAlex-Queries

# This is a tool to further interdisciplinary research – you are a neuroscientist who has used ..., What have the ... been doing with them. 
# Your a philosopher of science who wonders where the concept of a fitnesslandscape has appeared...
# """)
#     with gr.Row():
#         with gr.Column():
#             text_input = gr.Textbox(label="Name")
#             markdown = gr.Markdown(label="Output Box")
#             new_btn = gr.Button("New")
#         with gr.Column():
#             html = gr.HTML(label="HTML preview", show_label=True)

#     new_btn.click(fn=predict, inputs=[text_input], outputs=[markdown, html])



with gr.Blocks() as block:
    gr.Markdown("""
    ## Mapping OpenAlex-Queries

    This is a tool to further interdisciplinary research – you are a neuroscientist who has used ..., What have the ... been doing with them. 
    You're a philosopher of science who wonders where the concept of a fitness landscape has appeared...
    """)
    
    with gr.Column():
        text_input = gr.Textbox(label="OpenAlex Fulltext-Search")
        sample_size_slider = gr.Slider(label="Sample Size", minimum=10, maximum=20000, step=10, value=1000)
        reduce_sample_checkbox = gr.Checkbox(label="Reduce Sample Size", value=True)
        new_btn = gr.Button("Run Query")
        markdown = gr.Markdown(label="")
        html = gr.HTML(label="HTML preview", show_label=True)
    
    new_btn.click(fn=predict, inputs=[text_input, sample_size_slider, reduce_sample_checkbox], outputs=[markdown, html])








def setup_basemap_data():
    # get data.
    print("getting basemap data...")
    basedata_df = load("100k_filtered_OA_sample_cluster_and_positions.bz")
    print(basedata_df)
    return basedata_df


    
def setup_mapper():
    print("getting mapper...")

    params_new  = pickle.load(open('umap_mapper_300k_random_OA_specter_2_params.pkl', 'rb'))
    print("setting up mapper...")
    mapper = umap.UMAP()
    
    # Filter out 'target_backend' from umap_params if it exists
    umap_params = {k: v for k, v in params_new.get('umap_params', {}).items() if k != 'target_backend'}
    mapper.set_params(**umap_params)
    
    for attr, value in params_new.get('umap_attributes', {}).items():
        if attr != 'embedding_':
            setattr(mapper, attr, value)
    
    if 'embedding_' in params_new.get('umap_attributes', {}):
        mapper.embedding_ = List(params_new['umap_attributes']['embedding_'])
    
    return mapper


    

basedata_df = setup_basemap_data()
mapper = setup_mapper()


# mount Gradio app to FastAPI app
app = gr.mount_gradio_app(app, block, path="/")

# serve the app
if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)

# run the app with
# python app.py
# or
# uvicorn "app:app" --host "0.0.0.0" --port 7860 --reload