app.py

import gradio as gr
from datetime import datetime
import pandas as pd
import numpy as np
import time
#from sklearn.cluster import KMeans
from sklearn.feature_extraction.text import CountVectorizer
from transformers import AutoModel, AutoTokenizer
from transformers.pipelines import pipeline
from sklearn.pipeline import make_pipeline
from sklearn.decomposition import TruncatedSVD
from sklearn.feature_extraction.text import TfidfVectorizer
import funcs.anonymiser as anon
from umap import UMAP

from torch import cuda, backends, version

random_seed = 42

# Check for torch cuda
print("Is CUDA enabled? ", cuda.is_available())
print("Is a CUDA device available on this computer?", backends.cudnn.enabled)
if cuda.is_available():
    torch_device = "gpu"
    print("Cuda version installed is: ", version.cuda)
    low_resource_mode = "No"
    #os.system("nvidia-smi")
else: 
    torch_device =  "cpu"
    low_resource_mode = "Yes"

print("Device used is: ", torch_device)

#os.environ['CUDA_VISIBLE_DEVICES'] = '-1'

from bertopic import BERTopic
#from sentence_transformers import SentenceTransformer
#from bertopic.backend._hftransformers import HFTransformerBackend

#from cuml.manifold import UMAP

#umap_model = UMAP(n_components=5, n_neighbors=15, min_dist=0.0)

today = datetime.now().strftime("%d%m%Y")
today_rev = datetime.now().strftime("%Y%m%d")

from funcs.helper_functions import dummy_function, put_columns_in_df, read_file, get_file_path_end, zip_folder, delete_files_in_folder
#from funcs.representation_model import representation_model
from funcs.embeddings import make_or_load_embeddings


# Load embeddings
#embedding_model_name = "BAAI/bge-small-en-v1.5"
#embedding_model = SentenceTransformer(embedding_model_name)

# Pinning a Jina revision for security purposes: https://www.baseten.co/blog/pinning-ml-model-revisions-for-compatibility-and-security/
# Save Jina model locally as described here: https://huggingface.co/jinaai/jina-embeddings-v2-base-en/discussions/29
embeddings_name = "jinaai/jina-embeddings-v2-small-en"
local_embeddings_location = "model/jina/"
revision_choice = "b811f03af3d4d7ea72a7c25c802b21fc675a5d99"

if low_resource_mode == "No":
    try:
        embedding_model = AutoModel.from_pretrained(local_embeddings_location, revision = revision_choice, trust_remote_code=True,local_files_only=True, device_map="auto")
    except:
        embedding_model = AutoModel.from_pretrained(embeddings_name, revision = revision_choice, trust_remote_code=True, device_map="auto")

    tokenizer = AutoTokenizer.from_pretrained("jinaai/jina-embeddings-v2-small-en")

    embedding_model_pipe = pipeline("feature-extraction", model=embedding_model, tokenizer=tokenizer)

elif low_resource_mode == "Yes":
    embedding_model_pipe = make_pipeline(
                TfidfVectorizer(),
                TruncatedSVD(2) # 100 # set to 2 to be compatible with zero shot topics - can't be higher than number of topics
                )

# Model used for representing topics
hf_model_name =  'TheBloke/phi-2-orange-GGUF' #'NousResearch/Nous-Capybara-7B-V1.9-GGUF' # 'second-state/stablelm-2-zephyr-1.6b-GGUF'
hf_model_file =   'phi-2-orange.Q5_K_M.gguf' #'Capybara-7B-V1.9-Q5_K_M.gguf' # 'stablelm-2-zephyr-1_6b-Q5_K_M.gguf'


def extract_topics(in_files, in_file, min_docs_slider, in_colnames, max_topics_slider, candidate_topics, in_label, anonymise_drop, return_intermediate_files, embeddings_super_compress, low_resource_mode, create_llm_topic_labels, save_topic_model, visualise_topics, reduce_outliers, embeddings_out, progress=gr.Progress()):

    progress(0, desc= "Loading data")

    if not in_colnames or not in_label:
        error_message = "Please enter one column name for the topics and another for the labelling."
        print(error_message)
        return error_message, None, None, embeddings_out

    all_tic = time.perf_counter()

    output_list = []
    file_list = [string.name for string in in_file]

    data_file_names = [string.lower() for string in file_list if "tokenised" not in string and "npz" not in string.lower() and "gz" not in string.lower()]
    data_file_name = data_file_names[0]
    data_file_name_no_ext = get_file_path_end(data_file_name)

    in_colnames_list_first = in_colnames[0]

    if in_label:
        in_label_list_first = in_label[0]
    else:
        in_label_list_first = in_colnames_list_first

    # Make sure format of input series is good
    in_files[in_colnames_list_first] = in_files[in_colnames_list_first].fillna('').astype(str)
    in_files[in_label_list_first] = in_files[in_label_list_first].fillna('').astype(str)
    
    if anonymise_drop == "Yes":
        progress(0.1, desc= "Anonymising data")
        anon_tic = time.perf_counter()
        time_out = f"Creating visualisation took {all_toc - vis_tic:0.1f} seconds"
        in_files_anon_col, anonymisation_success = anon.anonymise_script(in_files, in_colnames_list_first, anon_strat="replace")
        in_files[in_colnames_list_first] = in_files_anon_col[in_colnames_list_first]
        anonymise_data_name = "anonymised_data.csv"
        in_files.to_csv(anonymise_data_name)
        output_list.append(anonymise_data_name)

        anon_toc = time.perf_counter()
        time_out = f"Anonymising text took {anon_toc - anon_tic:0.1f} seconds"

    docs = list(in_files[in_colnames_list_first].str.lower())
    label_list = list(in_files[in_label_list_first])

    # Check if embeddings are being loaded in
    ## Load in pre-embedded file if exists
    file_list = [string.name for string in in_file]

    print("Low resource mode: ", low_resource_mode)

    if low_resource_mode == "No":
        print("Using high resource Jina transformer model")
        try:
            embedding_model = AutoModel.from_pretrained(local_embeddings_location, revision = revision_choice, trust_remote_code=True,local_files_only=True, device_map="auto")
        except:
            embedding_model = AutoModel.from_pretrained(embeddings_name, revision = revision_choice, trust_remote_code=True, device_map="auto")

        tokenizer = AutoTokenizer.from_pretrained("jinaai/jina-embeddings-v2-small-en")

        embedding_model_pipe = pipeline("feature-extraction", model=embedding_model, tokenizer=tokenizer)

        # UMAP model uses Bertopic defaults
        umap_model = UMAP(n_neighbors=15, n_components=5, min_dist=0.0, metric='cosine', low_memory=False, random_state=random_seed)

    elif low_resource_mode == "Yes":
        print("Choosing low resource TF-IDF model.")

        embedding_model_pipe = make_pipeline(
                TfidfVectorizer(),
                TruncatedSVD(100) # 100 # To be compatible with zero shot, this needs to be lower than number of suggested topics
                )
        embedding_model = embedding_model_pipe

        umap_model = TruncatedSVD(n_components=5, random_state=random_seed)

    progress(0.2, desc= "Loading/creating embeddings")

    embeddings_out, reduced_embeddings = make_or_load_embeddings(docs, file_list, data_file_name_no_ext, embeddings_out, embedding_model, return_intermediate_files, embeddings_super_compress, low_resource_mode, create_llm_topic_labels)

    vectoriser_model = CountVectorizer(stop_words="english", ngram_range=(1, 2), min_df=0.1)
    
    from funcs.prompts import capybara_prompt, capybara_start, open_hermes_prompt, open_hermes_start, stablelm_prompt, stablelm_start
    from funcs.representation_model import create_representation_model, llm_config, chosen_start_tag

    
    progress(0.3, desc= "Embeddings loaded. Creating BERTopic model")

    if not candidate_topics:
        
        # Generate representation model here if topics won't be changed later
        # if reduce_outliers == "No":
        #     topic_model = BERTopic( embedding_model=embedding_model_pipe,
        #                             vectorizer_model=vectoriser_model,
        #                             umap_model=umap_model,
        #                             min_topic_size = min_docs_slider,
        #                             nr_topics = max_topics_slider,
        #                             representation_model=representation_model,
        #                             verbose = True)
        
        topic_model = BERTopic( embedding_model=embedding_model_pipe,
                                vectorizer_model=vectoriser_model,
                                umap_model=umap_model,
                                min_topic_size = min_docs_slider,
                                nr_topics = max_topics_slider,
                                verbose = True)

        topics_text, probs = topic_model.fit_transform(docs, embeddings_out)   


    # Do this if you have pre-defined topics
    else:
        if low_resource_mode == "Yes":
            error_message = "Zero shot topic modelling currently not compatible with low-resource embeddings. Please change this option to 'No' on the options tab and retry."
            print(error_message)

            return error_message, output_list, None

        zero_shot_topics = read_file(candidate_topics.name)
        zero_shot_topics_lower = list(zero_shot_topics.iloc[:, 0].str.lower())

        # Generate representation model here if topics won't be changed later
        # if reduce_outliers == "No":
        #     topic_model = BERTopic( embedding_model=embedding_model_pipe,
        #                             vectorizer_model=vectoriser_model,
        #                             umap_model=umap_model,
        #                             min_topic_size = min_docs_slider,
        #                             nr_topics = max_topics_slider,
        #                             zeroshot_topic_list = zero_shot_topics_lower,
        #                             zeroshot_min_similarity = 0.5,#0.7,
        #                             representation_model=representation_model,
        #                             verbose = True)
        # else:
        topic_model = BERTopic( embedding_model=embedding_model_pipe,
                                vectorizer_model=vectoriser_model,
                                umap_model=umap_model,
                                min_topic_size = min_docs_slider,
                                nr_topics = max_topics_slider,
                                zeroshot_topic_list = zero_shot_topics_lower,
                                zeroshot_min_similarity = 0.5,#0.7,
                                verbose = True)
        
        topics_text, probs = topic_model.fit_transform(docs, embeddings_out)

    if not topics_text:
        return "No topics found.", data_file_name, None
        
    else: 
        print("Topic model created.")

    progress(0.5, desc= "Loading in representation model")
    print("Create LLM topic labels:", create_llm_topic_labels)
    representation_model = create_representation_model(create_llm_topic_labels, llm_config, hf_model_name, hf_model_file, chosen_start_tag, low_resource_mode)  


    # Reduce outliers if required, then update representation
    if reduce_outliers == "Yes":
        progress(0.6, desc= "Reducing outliers then creating topic representations")
        print("Reducing outliers.")
        # Calculate the c-TF-IDF representation for each outlier document and find the best matching c-TF-IDF topic representation using cosine similarity.
        topics_text = topic_model.reduce_outliers(docs, topics_text, strategy="embeddings")
        # Then, update the topics to the ones that considered the new data
        print("Finished reducing outliers.")

    progress(0.6, desc= "Creating topic representations")
    topic_model.update_topics(docs, topics=topics_text, vectorizer_model=vectoriser_model, representation_model=representation_model)

    topic_dets = topic_model.get_topic_info()

    if topic_dets.shape[0] == 1:
        topic_det_output_name = "topic_details_" + data_file_name_no_ext + "_" + today_rev + ".csv"
        topic_dets.to_csv(topic_det_output_name)
        output_list.append(topic_det_output_name)

        return "No topics found, original file returned", output_list, None, embeddings_out

    # Replace original labels with LLM labels
    if "Phi" in topic_model.get_topic_info().columns:
        llm_labels = [label[0][0].split("\n")[0] for label in topic_model.get_topics(full=True)["Phi"].values()]
        topic_model.set_topic_labels(llm_labels)
    else:
        topic_model.set_topic_labels(list(topic_dets["Name"]))

    # Outputs
    progress(0.8, desc= "Saving output")
    
    topic_det_output_name = "topic_details_" + data_file_name_no_ext + "_" + today_rev + ".csv"
    topic_dets.to_csv(topic_det_output_name)
    output_list.append(topic_det_output_name)

    doc_det_output_name = "doc_details_" + data_file_name_no_ext + "_" + today_rev + ".csv"
    doc_dets = topic_model.get_document_info(docs)[["Document",	"Topic", "Name", "Representative_document"]] # "Probability",
    doc_dets.to_csv(doc_det_output_name)
    output_list.append(doc_det_output_name)

    topics_text_out_str = str(topic_dets["Name"])
    output_text = "Topics: " + topics_text_out_str
   
    # Save topic model to file
    if save_topic_model == "Yes":
        topic_model_save_name_folder = "output_model/" + data_file_name_no_ext + "_topics_" + today_rev# + ".safetensors"
        topic_model_save_name_zip = topic_model_save_name_folder + ".zip"

        # Clear folder before replacing files
        delete_files_in_folder(topic_model_save_name_folder)

        topic_model.save(topic_model_save_name_folder, serialization='pytorch', save_embedding_model=True, save_ctfidf=False)

        # Zip file example
        
        zip_folder(topic_model_save_name_folder, topic_model_save_name_zip)
        output_list.append(topic_model_save_name_zip)

     # If you want to save your embedding files
    if return_intermediate_files == "Yes":
        print("Saving embeddings to file")
        if low_resource_mode == "Yes":
            embeddings_file_name = data_file_name_no_ext + '_' + 'tfidf_embeddings.npz'
        else:
            if embeddings_super_compress == "No":
                embeddings_file_name = data_file_name_no_ext + '_' + 'ai_embeddings.npz'
            else:
                embeddings_file_name = data_file_name_no_ext + '_' + 'ai_embedding_compress.npz'

        np.savez_compressed(embeddings_file_name, embeddings_out)

        output_list.append(embeddings_file_name)

    if visualise_topics == "Yes":
        from funcs.bertopic_vis_documents import visualize_documents_custom
        progress(0.9, desc= "Creating visualisation (this can take a while)")
        # Visualise the topics:
        vis_tic = time.perf_counter()
        print("Creating visualisation")
        topics_vis = visualize_documents_custom(topic_model, docs, hover_labels = label_list, reduced_embeddings=reduced_embeddings, hide_annotations=True, hide_document_hover=False, custom_labels=True)
        topics_vis_name = data_file_name_no_ext + '_' + 'visualisation_' + today_rev + '.html'
        topics_vis.write_html(topics_vis_name)
        output_list.append(topics_vis_name)

        all_toc = time.perf_counter()
        time_out = f"Creating visualisation took {all_toc - vis_tic:0.1f} seconds"
        print(time_out)

        
        time_out = f"All processes took {all_toc - all_tic:0.1f} seconds"
        print(time_out)

        return output_text, output_list, topics_vis, embeddings_out

    all_toc = time.perf_counter()
    time_out = f"All processes took {all_toc - all_tic:0.1f} seconds."
    print(time_out)

    return output_text, output_list, None, embeddings_out

# ## Gradio app - extract topics

block = gr.Blocks(theme = gr.themes.Base())

with block:

    data_state = gr.State(pd.DataFrame())
    embeddings_state = gr.State(np.array([]))
 
    gr.Markdown(
    """
    # Topic modeller
    Generate topics from open text in tabular data. Upload a file (csv, xlsx, or parquet), then specify the open text column that you want to use to generate topics, and another for labels in the visualisation. If you have an embeddings .npz file of the text made using the 'jina-embeddings-v2-small-en' model, you can load this in at the same time to skip the first modelling step. If you have a pre-defined list of topics, you can upload this as a csv file under 'I have my own list of topics...'. Further configuration options are available under the 'Options' tab.

    Suggested test dataset: https://huggingface.co/datasets/rag-datasets/mini_wikipedia/tree/main/data (passages.parquet)
    """)    
          
    with gr.Tab("Load files and find topics"):
        with gr.Accordion("Load data file", open = True):
            in_files = gr.File(label="Input text from file", file_count="multiple")
            with gr.Row():
                in_colnames = gr.Dropdown(choices=["Choose a column"], multiselect = True, label="Select column to find topics (first will be chosen if multiple selected).")
                in_label = gr.Dropdown(choices=["Choose a column"], multiselect = True, label="Select column for labelling documents in the output visualisation.")

        with gr.Accordion("I have my own list of topics (zero shot topic modelling).", open = False):
            candidate_topics = gr.File(label="Input topics from file (csv). File should have at least one column with a header and topic keywords in cells below. Topics will be taken from the first column of the file. Currently not compatible with low-resource embeddings.")
            
        with gr.Row():
            min_docs_slider = gr.Slider(minimum = 2, maximum = 1000, value = 15, step = 1, label = "Minimum number of similar documents needed to make a topic.")
            max_topics_slider = gr.Slider(minimum = 2, maximum = 500, value = 10, step = 1, label = "Maximum number of topics")

        with gr.Row():
            topics_btn = gr.Button("Extract topics")
            
        with gr.Row():
            output_single_text = gr.Textbox(label="Output example (first example in dataset)")
            output_file = gr.File(label="Output file")

        plot = gr.Plot(label="Visualise your topics here. Go to the 'Options' tab to enable.")
    
    with gr.Tab("Options"):
        with gr.Accordion("Data load and processing options", open = True):
            with gr.Row():
                anonymise_drop = gr.Dropdown(value = "No", choices=["Yes", "No"], multiselect=False, label="Anonymise data on file load. Names and other details are replaced with tags e.g. '<person>'.")
                return_intermediate_files = gr.Dropdown(label = "Return intermediate processing files from file preparation. Files can be loaded in to save processing time in future.", value="Yes", choices=["Yes", "No"])
                embedding_super_compress = gr.Dropdown(label = "Round embeddings to three dp for smaller files with less accuracy.", value="No", choices=["Yes", "No"])
            with gr.Row():
                low_resource_mode_opt = gr.Dropdown(label = "Use low resource embeddings and processing.", value="No", choices=["Yes", "No"])
                create_llm_topic_labels = gr.Dropdown(label = "Create LLM-generated topic labels.", value="No", choices=["Yes", "No"])
                reduce_outliers = gr.Dropdown(label = "Reduce outliers by selecting closest topic.", value="No", choices=["Yes", "No"])
            with gr.Row():
                save_topic_model = gr.Dropdown(label = "Save topic model to file.", value="Yes", choices=["Yes", "No"])
                visualise_topics = gr.Dropdown(label = "Create a visualisation to map topics.", value="No", choices=["Yes", "No"])

    # Update column names dropdown when file uploaded
    in_files.upload(fn=put_columns_in_df, inputs=[in_files], outputs=[in_colnames, in_label, data_state, embeddings_state])    
    in_colnames.change(dummy_function, in_colnames, None)

    topics_btn.click(fn=extract_topics, inputs=[data_state, in_files, min_docs_slider, in_colnames, max_topics_slider, candidate_topics, in_label, anonymise_drop, return_intermediate_files, embedding_super_compress, low_resource_mode_opt, create_llm_topic_labels, save_topic_model, visualise_topics, reduce_outliers, embeddings_state], outputs=[output_single_text, output_file, plot, embeddings_state], api_name="topics")

block.queue().launch(debug=True)#, server_name="0.0.0.0", ssl_verify=False, server_port=7860)