app.py

import gradio as gr
import pandas as pd
import os
import plotly.express as px
import numpy as np

datadir = 'data/emissions/complete'

model_param_df = pd.read_csv('data/model_parameters.csv', header=0)
model_performance_df = pd.read_csv('data/performance.csv', header=0)
emissions_df = pd.read_csv('data/co2_data.csv',header=0)
modalities_df = pd.read_csv('data/modalities_data.csv',header=0)
finetuned_df = emissions_df[~emissions_df['task'].str.contains('zero')]
finetuned_df['task'] = finetuned_df['task'].str.replace('_',' ')

fig0 = px.scatter(emissions_df, x="num_params", y="query emissions (g)", color="model", log_x=True, log_y=True)
fig0.update_layout(xaxis={'categoryorder':'mean ascending'})
fig0.update_layout(yaxis_title='Total carbon emitted (g)')
fig0.update_layout(xaxis_title='Number of Parameters')


fig1 = px.box(finetuned_df, x="task", y="query_energy (kWh)", color="task", log_y=True)
fig1.update_layout(xaxis={'categoryorder':'mean ascending'})
fig1.update_layout(yaxis_title='Total energy used (Wh)')
fig1.update_layout(xaxis_title='Task')

fig2 = px.scatter(modalities_df, x="num_params", y="query emissions (g)", color="modality",
             log_x=True, log_y=True, custom_data=['model','task'])

fig2.update_traces(
    hovertemplate="<br>".join([
        "Model: %{customdata[0]}",
        "Task: %{customdata[1]}",
    ])
)
fig2.update_layout(xaxis_title='Model size (number of parameters)')
fig2.update_layout(yaxis_title='Model emissions (g of CO<sub>2</sub>)')


demo = gr.Blocks()

with demo:
    gr.Markdown("# CO2 Inference Demo")
    gr.Markdown("### TL;DR - We ran a series of experiments to measure the energy efficiency and carbon emissions of different\
    models from the HuggingFace Hub, and to see how different tasks and models compare.\
        We found that multi-purpose, generative models are orders of magnitude more energy-intensive than task-specific systems\
        for a variety of tasks, even for models with a similar number of parameters")
    gr.Markdown("### Explore the plots below to get more insights about the different models and tasks from our study.")
    with gr.Accordion("More details about our methodology:", open=False):
        gr.Markdown("We chose ten ML tasks: text classification, token classification, question answering, \
        ), masked language modeling, text generation, summarization, image classification, object detection, \
         image captioning and image generation. For each of the taks, we chose three of the most downloaded datasets and 8 of the most \
        downloaded models from the Hugging Face Hub. We ran each of the models ten times over a 1,000 sample from each of the models and measured the energy consumed and carbon emitted.")
    with gr.Row():
        with gr.Column():
            gr.Markdown("## All models from our study (carbon)")
            gr.Markdown("### Double click on the model name in the list on the right to isolate its datapoints:")
            gr.Markdown("The axes of the plot are in logarithmic scale, meaning that the difference between the least carbon-intensive and the most carbon-intensive models is over 9,000 times!")
            gr.Plot(fig0)
    with gr.Row():
        with gr.Column():
            gr.Markdown("## Task-by-task comparison (energy)")
            gr.Markdown("### Grouping the models by task, we can see different patterns emerge:")
            gr.Markdown("Image generation is by far the most energy- and carbon-intensive task from the ones studied, and text classification \
            is the least.")
            gr.Plot(fig1)
        with gr.Column():
            gr.Markdown("##  Modality comparison (carbon)")
            gr.Markdown("### Grouping the models by their modality shows the different characteristics of each one:")
            gr.Markdown("We can see that tasks involving images (image-to-text, image-to-category) require more energy and emit more carbon\
            than ones inolving text.")
            gr.Plot(fig2)






demo.launch()