app_bakeoff.py

import sys
sys.path.append('/Users/alexdeane/Downloads')

import random
import time

import pandas as pd
from faicons import icon_svg
from shiny import reactive, req
from shiny.express import input, module, render, ui
from shinymedia import audio_spinner, input_video_clip

from query_bakeoff import chat

ui.tags.style("body { background-color: var(--bs-light); }")


@module
def candidate(input, output, session, model_id, model_name, clip, current_winner):
    # Each model gets its own message history
    messages = []
    model_wins = reactive.value(0)

    ui.h5(model_name)

    # Process the video and chat with the model.
    @reactive.extended_task
    async def process(video_input, messages, session):
        req(video_input)
        with ui.Progress(session=session) as p:
            p.set(message=model_name)
            start_time = time.time_ns()
            return (
                await chat(model_id, video_input, messages, p),
                (time.time_ns() - start_time) / 1e9,
            )

    # Kick off the video processing whenever a new video is uploaded.
    @reactive.effect
    def start_process():
        process.cancel()
        process(clip(), messages, session)

    # When a result is available, show the audio player.
    @render.express
    def audio_output():
        result, time_secs = process.result()
        ui.markdown(f"Completed in **{time_secs:.2f}** secs")
        with ui.card(class_="bg-secondary", height="100px"):
            audio_spinner(src=result, height="100%", autoplay=False, autodismiss=False)

    # Show the voting button
    @render.express
    def vote_button():
        if process.status() != "success":
            return

        if current_winner() is None:
            # Voting hasn't happened yet, show the voting button
            ui.input_action_button(
                "vote",
                [icon_svg("thumbs-up"), " Vote for ", model_name],
                class_="btn-primary",
            )
        elif current_winner() == model_id:
            # Voting already happened, and this model won. Show a disabled
            # success button.
            success_labels = [
                "Booya",
                "Nailed it",
                "Suck it, other models",
                "I knew it",
                "Victory is mine",
                "Hell yeah",
                "Obviously",
            ]
            random_label = success_labels[random.randint(0, len(success_labels) - 1)]
            ui.input_action_button(
                "unvote",
                [icon_svg("thumbs-up"), f" {random_label} "],
                class_="btn-success",
            )

    @reactive.effect
    @reactive.event(input.vote)
    def vote():
        # The user voted for this model! Record that result.
        current_winner.set(model_id)
        model_wins.set(model_wins() + 1)

    @reactive.effect
    @reactive.event(input.unvote)
    def unvote():
        # The user made a mistake, undo the vote.
        current_winner.set(None)
        model_wins.set(model_wins() - 1)

    # Make the current number of wins for this model available to the rest of
    # the app.
    return model_wins


with ui.layout_columns(class_="my-4"):
    input_video_clip("clip", reset_on_record=True)
    with ui.card(fill=True, height="250px"):

        @render.plot
        def plot_leader():
            df = results_df()
            # Make a horizontal bar chart of the results using matplotlib, setting the x axis to start at 0 and end at 5 or the max value
            ax = df.plot.barh(x="Model", y="Wins", legend=False)
            ax.set_xlim(0, max(5, df["Wins"].max()))
            return ax


current_winner = reactive.value(None)
candidates = {}


@reactive.effect
def reset_winner():
    if input.clip() is None:
        current_winner.set(None)


with ui.panel_conditional("input.clip"):
    with ui.layout_columns(class_="my-4"):
        with ui.card():
            candidates["GPT-4o"] = candidate(
                "one", "gpt-4o", "GPT-4o", input.clip, current_winner
            )
        with ui.card():
            candidates["GPT-4"] = candidate(
                "two", "gpt-4-vision-preview", "GPT-4", input.clip, current_winner
            )
        with ui.card():
            candidates["LLaVA (7B)"] = candidate(
                "three", "llava:7b", "LLaVA (7B)", input.clip, current_winner
            )


@reactive.calc
def results_df():
    snapshot = {k: v() for k, v in candidates.items()}
    return pd.DataFrame(snapshot.items(), columns=["Model", "Wins"])