app.py

import spaces
import requests
import logging
import duckdb
from gradio_huggingfacehub_search import HuggingfaceHubSearch
from bertopic import BERTopic
from bertopic.representation import (
    KeyBERTInspired,
    TextGeneration,
)
from umap import UMAP
import numpy as np
from torch import cuda, bfloat16
from transformers import (
    BitsAndBytesConfig,
    AutoTokenizer,
    AutoModelForCausalLM,
    pipeline,
)
from prompts import REPRESENTATION_PROMPT
from hdbscan import HDBSCAN
from sklearn.feature_extraction.text import CountVectorizer

# from cuml.cluster import HDBSCAN
# from cuml.manifold import UMAP
from sentence_transformers import SentenceTransformer
import gradio as gr

logging.basicConfig(
    level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s"
)


session = requests.Session()
sentence_model = SentenceTransformer("all-MiniLM-L6-v2")
keybert = KeyBERTInspired()
vectorizer_model = CountVectorizer(stop_words="english")

model_id = "meta-llama/Llama-2-7b-chat-hf"
device = f"cuda:{cuda.current_device()}" if cuda.is_available() else "cpu"
logging.info(device)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,  # 4-bit quantization
    bnb_4bit_quant_type="nf4",  # Normalized float 4
    bnb_4bit_use_double_quant=True,  # Second quantization after the first
    bnb_4bit_compute_dtype=bfloat16,  # Computation type
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    quantization_config=bnb_config,
    device_map="auto",
    offload_folder="offload",  # Offloading part of the model to CPU to save GPU memory
)

# Enable gradient checkpointing for memory efficiency during backprop?
model.gradient_checkpointing_enable()

generator = pipeline(
    model=model,
    tokenizer=tokenizer,
    task="text-generation",
    temperature=0.1,
    max_new_tokens=200,  # Reduced max_new_tokens to limit memory consumption
    repetition_penalty=1.1,
)

llama2 = TextGeneration(generator, prompt=REPRESENTATION_PROMPT)
representation_model = {
    "KeyBERT": keybert,
    "Llama2": llama2,
}

# TODO: It should be proporcional to the number of rows
# For small datasets (1-200 rows) it worked fine with 2 neighbors
N_NEIGHBORS = 15

umap_model = UMAP(
    n_neighbors=N_NEIGHBORS,
    n_components=5,
    min_dist=0.0,
    metric="cosine",
    random_state=42,
)

hdbscan_model = HDBSCAN(
    min_cluster_size=N_NEIGHBORS,
    metric="euclidean",
    cluster_selection_method="eom",
    prediction_data=True,
)

reduce_umap_model = UMAP(
    n_neighbors=N_NEIGHBORS,
    n_components=2,
    min_dist=0.0,
    metric="cosine",
    random_state=42,
)


def get_parquet_urls(dataset, config, split):
    parquet_files = session.get(
        f"https://datasets-server.huggingface.co/parquet?dataset={dataset}&config={config}&split={split}",
        timeout=20,
    ).json()
    if "error" in parquet_files:
        raise Exception(f"Error fetching parquet files: {parquet_files['error']}")
    parquet_urls = [file["url"] for file in parquet_files["parquet_files"]]
    logging.debug(f"Parquet files: {parquet_urls}")
    return ",".join(f"'{url}'" for url in parquet_urls)


def get_docs_from_parquet(parquet_urls, column, offset, limit):
    SQL_QUERY = f"SELECT {column} FROM read_parquet([{parquet_urls}]) LIMIT {limit} OFFSET {offset};"
    df = duckdb.sql(SQL_QUERY).to_df()
    logging.debug(f"Dataframe: {df.head(5)}")
    return df[column].tolist()


@spaces.GPU
# TODO: Modify batch size to reduce memory consumption during embedding calculation, which value is better?
def calculate_embeddings(docs):
    return sentence_model.encode(docs, show_progress_bar=True, batch_size=32)


@spaces.GPU
def fit_model(base_model, docs, embeddings):
    new_model = BERTopic(
        "english",
        # Sub-models
        embedding_model=sentence_model,
        umap_model=umap_model,
        hdbscan_model=hdbscan_model,
        representation_model=representation_model,
        vectorizer_model=vectorizer_model,
        # Hyperparameters
        top_n_words=10,
        verbose=True,
        min_topic_size=15,  # TODO: Should this value be coherent with N_NEIGHBORS?
    )
    logging.debug("Fitting new model")
    new_model.fit(docs, embeddings)
    logging.debug("End fitting new model")

    if base_model is None:
        return new_model, new_model

    updated_model = BERTopic.merge_models([base_model, new_model])
    nr_new_topics = len(set(updated_model.topics_)) - len(set(base_model.topics_))
    new_topics = list(updated_model.topic_labels_.values())[-nr_new_topics:]
    logging.info(f"The following topics are newly found: {new_topics}")
    return updated_model, new_model


def generate_topics(dataset, config, split, column, nested_column):
    logging.info(
        f"Generating topics for {dataset} with config {config} {split} {column} {nested_column}"
    )

    parquet_urls = get_parquet_urls(dataset, config, split)
    limit = 1_000
    chunk_size = 300
    offset = 0
    base_model = None
    all_docs = []
    reduced_embeddings_list = []
    topics_info, topic_plot = None, None
    while offset < limit:
        docs = get_docs_from_parquet(parquet_urls, column, offset, chunk_size)
        if not docs:
            break

        logging.info(
            f"----> Processing chunk: {offset=} {chunk_size=} with {len(docs)} docs"
        )

        embeddings = calculate_embeddings(docs)
        base_model, _ = fit_model(base_model, docs, embeddings)

        repr_model_topics = {
            key: label[0][0].split("\n")[0]
            for key, label in base_model.get_topics(full=True)["Llama2"].items()
        }

        base_model.set_topic_labels(repr_model_topics)

        reduced_embeddings = reduce_umap_model.fit_transform(embeddings)
        reduced_embeddings_list.append(reduced_embeddings)

        all_docs.extend(docs)

        topics_info = base_model.get_topic_info()
        topic_plot = base_model.visualize_documents(
            all_docs,
            reduced_embeddings=np.vstack(reduced_embeddings_list),
            custom_labels=True,
        )

        logging.info(f"Topics: {repr_model_topics}")

        yield topics_info, topic_plot

        offset += chunk_size

    logging.info("Finished processing all data")
    cuda.empty_cache()  # Clear cache at the end of each chunk
    return topics_info, topic_plot


with gr.Blocks() as demo:
    gr.Markdown("# 💠 Dataset Topic Discovery 🔭")
    gr.Markdown("## Select dataset and text column")
    with gr.Accordion("Data details", open=True):
        with gr.Row():
            with gr.Column(scale=3):
                dataset_name = HuggingfaceHubSearch(
                    label="Hub Dataset ID",
                    placeholder="Search for dataset id on Huggingface",
                    search_type="dataset",
                )
            subset_dropdown = gr.Dropdown(label="Subset", visible=False)
            split_dropdown = gr.Dropdown(label="Split", visible=False)

        with gr.Accordion("Dataset preview", open=False):

            @gr.render(inputs=[dataset_name, subset_dropdown, split_dropdown])
            def embed(name, subset, split):
                html_code = f"""
                <iframe
                src="https://huggingface.co/datasets/{name}/embed/viewer/{subset}/{split}"
                frameborder="0"
                width="100%"
                height="600px"
                ></iframe>
                    """
                return gr.HTML(value=html_code)

        with gr.Row():
            text_column_dropdown = gr.Dropdown(label="Text column name")
            nested_text_column_dropdown = gr.Dropdown(
                label="Nested text column name", visible=False
            )

        generate_button = gr.Button("Generate Topics", variant="primary")

    gr.Markdown("## Datamap")
    topics_plot = gr.Plot()
    with gr.Accordion("Topics Info", open=False):
        topics_df = gr.DataFrame(interactive=False, visible=True)
    generate_button.click(
        generate_topics,
        inputs=[
            dataset_name,
            subset_dropdown,
            split_dropdown,
            text_column_dropdown,
            nested_text_column_dropdown,
        ],
        outputs=[topics_df, topics_plot],
    )

    def _resolve_dataset_selection(
        dataset: str, default_subset: str, default_split: str, text_feature
    ):
        if "/" not in dataset.strip().strip("/"):
            return {
                subset_dropdown: gr.Dropdown(visible=False),
                split_dropdown: gr.Dropdown(visible=False),
                text_column_dropdown: gr.Dropdown(label="Text column name"),
                nested_text_column_dropdown: gr.Dropdown(visible=False),
            }
        info_resp = session.get(
            f"https://datasets-server.huggingface.co/info?dataset={dataset}", timeout=20
        ).json()
        if "error" in info_resp:
            return {
                subset_dropdown: gr.Dropdown(visible=False),
                split_dropdown: gr.Dropdown(visible=False),
                text_column_dropdown: gr.Dropdown(label="Text column name"),
                nested_text_column_dropdown: gr.Dropdown(visible=False),
            }
        subsets: list[str] = list(info_resp["dataset_info"])
        subset = default_subset if default_subset in subsets else subsets[0]
        splits: list[str] = list(info_resp["dataset_info"][subset]["splits"])
        split = default_split if default_split in splits else splits[0]
        features = info_resp["dataset_info"][subset]["features"]

        def _is_string_feature(feature):
            return isinstance(feature, dict) and feature.get("dtype") == "string"

        text_features = [
            feature_name
            for feature_name, feature in features.items()
            if _is_string_feature(feature)
        ]
        nested_features = [
            feature_name
            for feature_name, feature in features.items()
            if isinstance(feature, dict)
            and isinstance(next(iter(feature.values())), dict)
        ]
        nested_text_features = [
            feature_name
            for feature_name in nested_features
            if any(
                _is_string_feature(nested_feature)
                for nested_feature in features[feature_name].values()
            )
        ]
        if not text_feature:
            return {
                subset_dropdown: gr.Dropdown(
                    value=subset, choices=subsets, visible=len(subsets) > 1
                ),
                split_dropdown: gr.Dropdown(
                    value=split, choices=splits, visible=len(splits) > 1
                ),
                text_column_dropdown: gr.Dropdown(
                    choices=text_features + nested_text_features,
                    label="Text column name",
                ),
                nested_text_column_dropdown: gr.Dropdown(visible=False),
            }
        if text_feature in nested_text_features:
            nested_keys = [
                feature_name
                for feature_name, feature in features[text_feature].items()
                if _is_string_feature(feature)
            ]
            return {
                subset_dropdown: gr.Dropdown(
                    value=subset, choices=subsets, visible=len(subsets) > 1
                ),
                split_dropdown: gr.Dropdown(
                    value=split, choices=splits, visible=len(splits) > 1
                ),
                text_column_dropdown: gr.Dropdown(
                    choices=text_features + nested_text_features,
                    label="Text column name",
                ),
                nested_text_column_dropdown: gr.Dropdown(
                    value=nested_keys[0],
                    choices=nested_keys,
                    label="Nested text column name",
                    visible=True,
                ),
            }
        return {
            subset_dropdown: gr.Dropdown(
                value=subset, choices=subsets, visible=len(subsets) > 1
            ),
            split_dropdown: gr.Dropdown(
                value=split, choices=splits, visible=len(splits) > 1
            ),
            text_column_dropdown: gr.Dropdown(
                choices=text_features + nested_text_features, label="Text column name"
            ),
            nested_text_column_dropdown: gr.Dropdown(visible=False),
        }

    @dataset_name.change(
        inputs=[dataset_name],
        outputs=[
            subset_dropdown,
            split_dropdown,
            text_column_dropdown,
            nested_text_column_dropdown,
        ],
    )
    def show_input_from_subset_dropdown(dataset: str) -> dict:
        return _resolve_dataset_selection(
            dataset, default_subset="default", default_split="train", text_feature=None
        )

    @subset_dropdown.change(
        inputs=[dataset_name, subset_dropdown],
        outputs=[
            subset_dropdown,
            split_dropdown,
            text_column_dropdown,
            nested_text_column_dropdown,
        ],
    )
    def show_input_from_subset_dropdown(dataset: str, subset: str) -> dict:
        return _resolve_dataset_selection(
            dataset, default_subset=subset, default_split="train", text_feature=None
        )

    @split_dropdown.change(
        inputs=[dataset_name, subset_dropdown, split_dropdown],
        outputs=[
            subset_dropdown,
            split_dropdown,
            text_column_dropdown,
            nested_text_column_dropdown,
        ],
    )
    def show_input_from_split_dropdown(dataset: str, subset: str, split: str) -> dict:
        return _resolve_dataset_selection(
            dataset, default_subset=subset, default_split=split, text_feature=None
        )

    @text_column_dropdown.change(
        inputs=[dataset_name, subset_dropdown, split_dropdown, text_column_dropdown],
        outputs=[
            subset_dropdown,
            split_dropdown,
            text_column_dropdown,
            nested_text_column_dropdown,
        ],
    )
    def show_input_from_text_column_dropdown(
        dataset: str, subset: str, split: str, text_column
    ) -> dict:
        return _resolve_dataset_selection(
            dataset,
            default_subset=subset,
            default_split=split,
            text_feature=text_column,
        )


demo.launch()