app.py

import gradio as gr
import pandas as pd
from huggingface_hub.hf_api import create_repo, upload_folder, upload_file, HfApi
from huggingface_hub.repository import Repository
import subprocess
import os
import tempfile
from uuid import uuid4
import pickle
import sweetviz as sv
import dabl
import re


def analyze_datasets(dataset, token, column=None, pairwise="off"):
    df = pd.read_csv(dataset.name)
    username = HfApi().whoami(token=token)["name"]
    if column is not None:
        analyze_report = sv.analyze(df, target_feat=column, pairwise_analysis=pairwise)
    else:
        analyze_report = sv.analyze(df, pairwise_analysis=pairwise)
    dataset_name = dataset.name.split("/")[-1].strip(".csv")
    analyze_report.show_html('./index.html', open_browser=False)
    
    repo_url = create_repo(f"{username}/{dataset_name}-report", repo_type = "space", token = token, space_sdk = "static", private=False)
    
    upload_file(path_or_fileobj ="./index.html", path_in_repo = "./index.html", repo_id =f"{username}/{dataset_name}-report", repo_type = "space", token=token)
    readme = f"---\ntitle: {dataset_name}\nemoji: ✨\ncolorFrom: green\ncolorTo: red\nsdk: static\npinned: false\ntags:\n- dataset-report\n---"    
    with open("README.md", "w+") as f:
        f.write(readme)
    upload_file(path_or_fileobj ="./README.md", path_in_repo = "README.md", repo_id =f"{username}/{dataset_name}-report", repo_type = "space", token=token)

    return f"Your dataset report will be ready at {repo_url}"


from sklearn.utils import estimator_html_repr


def extract_estimator_config(model):
    hyperparameter_dict = model.get_params(deep=True)
    table = "| Hyperparameters | Value |\n| :-- | :-- |\n"
    for hyperparameter, value in hyperparameter_dict.items():
        table += f"| {hyperparameter} | {value} |\n"
    return table

def detect_training(df, column):
    if dabl.detect_types(df)["continuous"][column] or dabl.detect_types(df)["dirty_float"][column]:
        trainer = dabl.SimpleRegressor()
        task = "regression"
    elif dabl.detect_types(df)["categorical"][column] or dabl.detect_types(df)["low_card_int"][column] or dabl.detect_types(df)["free_string"][column]:
        trainer = dabl.SimpleClassifier()
        task = "classification"
    return trainer, task

def edit_types(df):
    types = dabl.detect_types(df)
    low_cardinality = types[types["low_card_int"] == True].index.tolist()
    dirty_float = types[types["dirty_float"] == True].index.tolist()
    type_hints = {}
    for col in low_cardinality:
        type_hints[col] = "categorical"
    for col in dirty_float:
        type_hints[col] = "continuous"
    df_clean = dabl.clean(df, type_hints=type_hints)
    return df_clean

def train_baseline(dataset,  token, column):
    df = pd.read_csv(dataset.name)
    dataset_name = dataset.name.split("/")[-1].strip(".csv")
    df_clean = edit_types(df)
    fc, task = detect_training(df_clean, column)
    X = df_clean.drop(column, axis = 1)
    y = df_clean[column]
    
    with tempfile.TemporaryDirectory() as tmpdirname:
        from contextlib import redirect_stdout

        with open(f'{tmpdirname}/logs.txt', 'w') as f:
            with redirect_stdout(f):
                print('Logging training')
                fc.fit(X, y)
        username = HfApi().whoami(token=token)["name"]
        repo_url = create_repo(repo_id = f"{username}/{dataset_name}-{column}-{task}", token = token)
        if task == "regression":
            task_metadata = "tabular-regression"
        else:
            task_metadata = "tabular-classification"
        readme = f"---\nlicense: apache-2.0\nlibrary_name: sklearn\ntags:\n- {task_metadata}\n- baseline-trainer\n---\n\n"
        readme += f"## Baseline Model trained on {dataset_name} to apply {task} on {column}\n\n" 
        readme+="**Metrics of the best model:**\n\n"
        for elem in str(fc.current_best_).split("\n"):
            readme+= f"{elem}\n\n"
        readme+= "\n\n**See model plot below:**\n\n"
        readme+= re.sub(r"\n\s+", "", str(estimator_html_repr(fc.est_)))
        readme+= "\n\n**Disclaimer:** This model is trained with dabl library as a baseline, for better results, use [AutoTrain](https://huggingface.co/autotrain).\n\n"
        readme+= "**Logs of training** including the models tried in the process can be found in logs.txt" 
        with open(f"{tmpdirname}/README.md", "w+") as f:
            f.write(readme)
        with open(f"{tmpdirname}/clf.pkl", mode="bw") as f:
            pickle.dump(fc, file=f)
        upload_folder(repo_id =f"{username}/{dataset_name}-{column}-{task}", folder_path=tmpdirname, repo_type = "model", token=token, path_in_repo="./")

    return f"Your model will be ready at {repo_url}"



with gr.Blocks() as demo:
    main_title = gr.Markdown("""# Baseline Trainer 🪄🌟✨""")
    main_desc = gr.Markdown("""This app trains a baseline model for a given dataset and pushes it to your Hugging Face Hub Profile with a model card. For better results, use [AutoTrain](https://huggingface.co/autotrain).""")
    
    
    with gr.Tabs():
        with gr.TabItem("Baseline Trainer") as baseline_trainer:
            with gr.Row():
                with gr.Column():
                    title = gr.Markdown(""" ## Train a supervised baseline model""")
                    description = gr.Markdown("This app trains a model and pushes it to your Hugging Face Hub Profile.")
                    dataset = gr.File(label = "CSV Dataset")
                    column = gr.Text(label = "Enter target variable:")
                    pushing_desc = gr.Markdown("This app needs your Hugging Face Hub token and a unique name for your dataset report.")
                    token = gr.Textbox(label = "Your Hugging Face Token")
                    inference_run = gr.Button("Train")
                    inference_progress = gr.StatusTracker(cover_container=True)

                outcome = gr.outputs.Textbox(label = "Progress")
                inference_run.click(
                    train_baseline,
                    inputs=[dataset, token, column],
                    outputs=outcome,
                    status_tracker=inference_progress,
                )
        with gr.TabItem("Analyze") as analyze:
            with gr.Row():
                with gr.Column():
                    title = gr.Markdown(""" ## Analyze Dataset """)
                    description = gr.Markdown("Analyze a dataset or predictive variables against a target variable in a dataset (enter a column name to column section if you want to compare against target value). You can also do pairwise analysis, but it has quadratic complexity.")
                    dataset = gr.File(label = "CSV Dataset")
                    column = gr.Text(label = "Compare dataset against a target variable (Optional)")
                    pairwise = gr.Radio(["off", "on"], label = "Enable pairwise analysis")
                    token = gr.Textbox(label = "Your Hugging Face Token")
                    pushing_desc = gr.Markdown("This app needs your Hugging Face Hub token and a unique repository name for your dataset report.")
                    inference_run = gr.Button("Infer")
                    inference_progress = gr.StatusTracker(cover_container=True)
                outcome = gr.outputs.Textbox()
                inference_run.click(
                    analyze_datasets,
                    inputs=[dataset, token, column, pairwise],
                    outputs=outcome,
                    status_tracker=inference_progress,
                )

demo.launch(debug=True)