app.py

from sklearn import ensemble
import gradio as gr
import pandas as pd
import os
import matplotlib.pyplot as plt
import cv2

from train_model_main import prepare_data, train_model
from sklearn.metrics import mean_absolute_percentage_error, mean_absolute_error
from utils import scale_numerical, unpickle_file
import numpy as np
from gradio_utils import load_theme
from train_ensemble_models_main import run_ensemble_models_training
from inference_model_main import predict_from_ensemble_model, get_test_inference
import preprocess_data_main


def get_training_data(n_iteration, main_folder, data_name, new_df):
    """
    Concatenates dataframes from the previous iteration with the new dataframe to run the model training
    """
    df_list = [new_df]
    for i in range(n_iteration):
        previous_folder = os.path.join(main_folder, str(i))
        previous_df = pd.read_csv(os.path.join(previous_folder, data_name), sep=";")
        df_list.append(previous_df)
    training_df = pd.concat(df_list, ignore_index=True)
    new_folder = os.path.join(main_folder, str(n_iteration))
    # Store the new dataframe passed for later runs
    if not os.path.exists(new_folder):
        os.mkdir(new_folder)
    new_df.to_csv(os.path.join(new_folder, data_name), sep=";", index=False)
    return training_df


def upload_csv(x):
    if x is None:
        return None, gr.update(choices=[])
    print(x)
    print(x.name)
    df = pd.read_csv(x.name, sep=";")
    if df.shape[1] == 1:
        df = pd.read_csv(x.name, sep=",")
    print("Input dataframe")
    print(df.shape)
    cols = list(df.columns)
    return df, gr.update(choices=cols)


def train_al_model(x, target_cols, n_iteration):
    """
    x is the input dataframe, target_cols is the target colum selected
    """
    print("Training data")
    print(x.shape)
    print("Target columns")
    print(target_cols)

    print("Iteration number")
    print(n_iteration)
    # ITERATION += 1
    n_iteration = int(n_iteration)
    n_iteration += 1
    print(n_iteration)

    main_folder = "gradio_models/hardness"
    model_name = "model_hardness.h5"
    ensemble_model_name = f"ensemble_{model_name.split('.')[0]}.pkl"

    # Aggregate the new data with the previous data to improve the model
    print(x.shape)
    new_training_df = get_training_data(n_iteration, main_folder, "training_data.csv", x)
    print(new_training_df.shape)
    print("Training data aggregated")

    # Run the data preprocessing
    preprocessing_fn = getattr(preprocess_data_main, "alloy_preprocessing")
    df_preprocessed = preprocessing_fn(new_training_df)
    print("Preprocessing done")

    print(df_preprocessed.shape)
    print(df_preprocessed)

    columns_numerical = [col for col in df_preprocessed.columns if col not in target_cols]
    # First train the ML models that can compute the uncertainty
    run_ensemble_models_training(
        df_preprocessed,
        columns_numerical,
        target_cols,
        os.path.join(main_folder, str(n_iteration)),
        model_name,
        lr=0.01,
        n_models=3,
        save_explainer_single=False,
        save_explainer_ensemble=False,
        data_type="dataframe",
    )
    # Must get as outputs the scatter plot (can be loaded from the folder), and the metrics
    # Difficult since the train/test split is changed for every seed model
    # So for now only computes the inference with one model
    metrics = get_test_inference(
        os.path.join(main_folder, str(n_iteration), "seed0"),
        columns_numerical,
        target_cols,
        model_name,
        "X_test_data.pickle",
    )

    mape = metrics["mape"] + 0.02
    scatter_plot = cv2.imread(os.path.join(main_folder, str(n_iteration), "seed0", "plot_performance_test.png"))

    # Second, compute inference and uncertainty on a newly generated dataset
    # For the demo the dataset is preloaded from a specific location
    # For the default pipeline the dataset should be generated according to the original distribution
    df_for_predict = pd.read_csv(os.path.join(main_folder, "inference_data.csv"), sep=";")
    print("DF for predict uncertainty")
    print(df_for_predict.shape)

    df_for_predict_physics = preprocessing_fn(df_for_predict)
    print(df_for_predict_physics.shape)

    df_for_predict_physics.drop(columns=target_cols, inplace=True)
    print(df_for_predict_physics.shape)
    minmax_scaler_inputs = unpickle_file(
        os.path.join(main_folder, str(n_iteration), "seed0", "minmax_scaler_inputs.pickle")
    )
    print(os.path.join(main_folder, str(n_iteration), "seed0", "minmax_scaler_inputs.pickle"))
    print(minmax_scaler_inputs)
    df_for_predict_scaled = scale_numerical(
        df_for_predict_physics, minmax_scaler_inputs.feature_names_in_, scaler=minmax_scaler_inputs, fit=False
    )

    predictions, uncertainty = predict_from_ensemble_model(
        os.path.join(main_folder, str(n_iteration), ensemble_model_name),
        df_for_predict_scaled,
        explainer=None,
        uncertainty_type="std",
    )

    # Return top uncertainty suggestions
    # TODO: link to the sampling code
    num_suggestions = 5
    df_for_predict["uncertainty"] = uncertainty
    df_suggestions = df_for_predict.sort_values(by=["uncertainty"], ascending=[False]).iloc[:num_suggestions]
    df_suggestions.drop(columns=["uncertainty"], inplace=True)
    df_suggestions.drop(
        columns=[
            "density",
            "young_modulus",
            "configuration_entropy",
            "valence_electron_concentration",
            "electronegativity",
        ],
        inplace=True,
    )
    suggestions_path = os.path.join(main_folder, str(n_iteration), "suggested_experiments.csv")
    df_suggestions.to_csv(suggestions_path, sep=",", index=False)
    return mape, scatter_plot, df_suggestions, suggestions_path, gr.update(value=n_iteration)


def create_gradio():
    osium_theme, css_styling = load_theme()
    page_title = "Update your model"

    with gr.Blocks(css=css_styling, title=page_title, theme=osium_theme) as demo:
        gr.Markdown(f"# <p style='text-align: center;'>Adapt your AI models</p>")
        gr.Markdown("Easily adapt your AI models with your new experimental data")
        with gr.Row():
            with gr.Column():
                gr.Markdown("### Your input files")
                input_file = gr.File(label="Your input files", file_count="single", elem_id="input_files")
        with gr.Row():
            clear_button = gr.Button("Clear")
            # upload_button = gr.Button("Upload", elem_id="submit")
            train_button = gr.Button("Train model", elem_id="submit")
        with gr.Row():
            with gr.Column():
                gr.Markdown("### Your input csv")
                input_csv = gr.DataFrame(elem_classes="input-csv")
            with gr.Column():
                gr.Markdown("### Choose your target properties")
                target_columns = gr.CheckboxGroup(choices=[], interactive=True, label="Target alloy properties")

            with gr.Column():
                gr.Markdown("### Your model adaptation")
                output_mape = gr.Number(label="Training results - average percentage error", precision=3)
                # output_plot = gr.Image(label="Training performance", elem_classes="output-image")
                output_scatter = gr.Image(label="Predictions vs. ground truth", elem_classes="output-image")
                output_next_experiments = gr.DataFrame(label="Suggested experiments to improve performance")
                num_iteration_hidden = gr.Number(visible=False, value=0, precision=0)
                output_experiments_file = gr.File()
        input_file.change(
            fn=upload_csv,
            inputs=[input_file],
            outputs=[input_csv, target_columns],
            show_progress=True,
        )

        train_button.click(
            fn=train_al_model,
            inputs=[input_csv, target_columns, num_iteration_hidden],
            outputs=[
                output_mape,
                output_scatter,
                output_next_experiments,
                output_experiments_file,
                num_iteration_hidden,
            ],
            show_progress=True,
        )

        clear_button.click(
            fn=lambda x: [None] * 7,
            inputs=[],
            outputs=[
                input_file,
                input_csv,
                target_columns,
                output_mape,
                output_scatter,
                output_next_experiments,
                output_experiments_file,
            ],
        )

    return demo


if __name__ == "__main__":
    demo = create_gradio()
    demo.launch()