sasha's picture
sasha HF staff
filling in missing values
fd31451
raw
history blame
6.57 kB
import gradio as gr
import pandas as pd
from huggingface_hub import list_models
import plotly.express as px
def get_plots(task):
#TO DO : hover text with energy efficiency number, parameters
task_df= pd.read_csv('data/energy/'+task)
params_df = pd.read_csv('data/params/'+task)
params_df= params_df.rename(columns={"Link": "model"})
all_df = pd.merge(task_df, params_df, on='model')
all_df['Total GPU Energy (Wh)'] = all_df['total_gpu_energy']*1000
all_df = all_df.sort_values(by=['Total GPU Energy (Wh)'])
all_df['parameters'] = all_df['parameters'].apply(format_params)
all_df['energy_star'] = pd.cut(all_df['Total GPU Energy (Wh)'], 3, labels=["⭐⭐⭐", "⭐⭐", "⭐"])
fig = px.scatter(all_df, x="model", y='Total GPU Energy (Wh)', custom_data=['parameters'], height= 500, width= 800, color = 'energy_star', color_discrete_map={"⭐": 'red', "⭐⭐": "yellow", "⭐⭐⭐": "green"})
fig.update_traces(
hovertemplate="<br>".join([
"Total Energy: %{y}",
"Parameters: %{customdata[0]}"])
)
return fig
def make_link(mname):
link = "["+ str(mname).split('/')[1] +'](https://huggingface.co/'+str(mname)+")"
return link
def get_model_names(task_data):
#TODO: add link to results in model card of each model
task_df= pd.read_csv('data/params/'+task_data)
energy_df= pd.read_csv('data/energy/'+task_data)
task_df= task_df.rename(columns={"Link": "model"})
all_df = pd.merge(task_df, energy_df, on='model')
all_df=all_df.drop_duplicates(subset=['model'])
all_df['parameters'] = all_df['parameters'].apply(format_params)
all_df['model'] = all_df['model'].apply(make_link)
all_df['Total GPU Energy (Wh)'] = all_df['total_gpu_energy']*1000
all_df['Total GPU Energy (Wh)'] = all_df['Total GPU Energy (Wh)'].round(2)
model_names = all_df[['model','parameters','Total GPU Energy (Wh)']]
return model_names
def format_params(num):
if num > 1000000000:
if not num % 1000000000:
return f'{num // 1000000000}B'
return f'{round(num / 1000000000, 1)}B'
return f'{num // 1000000}M'
demo = gr.Blocks()
with demo:
gr.Markdown(
"""# Energy Star Leaderboard - v.0 (2024) 🌎 πŸ’» 🌟
### Welcome to the leaderboard for the [AI Energy Star Project!](https://huggingface.co/EnergyStarAI)
Click through the tasks below to see how different models measure up in terms of energy efficiency"""
)
gr.Markdown(
"""Test your own models via the [submission portal (TODO)]."""
)
with gr.Tabs():
with gr.TabItem("Text Generation πŸ’¬"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('text_generation.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('text_generation.csv'), datatype="markdown")
with gr.TabItem("Image Generation πŸ“·"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('image_generation.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('image_generation.csv'), datatype="markdown")
with gr.TabItem("Text Classification 🎭"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('text_classification.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('text_classification.csv'), datatype="markdown")
with gr.TabItem("Image Classification πŸ–ΌοΈ"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('image_classification.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('image_classification.csv'), datatype="markdown")
with gr.TabItem("Image Captioning πŸ“"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('question_answering.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('question_answering.csv'), datatype="markdown")
with gr.TabItem("Summarization πŸ“ƒ"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('summarization.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('summarization.csv'), datatype="markdown")
with gr.TabItem("Automatic Speech Recognition πŸ’¬ "):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('asr.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('asr.csv'), datatype="markdown")
with gr.TabItem("Object Detection 🚘"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('object_detection.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('object_detection.csv'), datatype="markdown")
with gr.TabItem("Sentence Similarity πŸ“š"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('sentence_similarity.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('sentence_similarity.csv'), datatype="markdown")
with gr.TabItem("Extractive QA ❔"):
with gr.Row():
with gr.Column():
plot = gr.Plot(get_plots('question_answering.csv'))
with gr.Column():
table = gr.Dataframe(get_model_names('question_answering.csv'), datatype="markdown")
with gr.Accordion("Methodology"):
gr.Markdown(
"""For each of the ten tasks above, we created a custom dataset with 1,000 entries (see all of the datasets on our [org Hub page](https://huggingface.co/EnergyStarAI)).
We then tested each of the models from the leaderboard on the appropriate task, measuring the energy consumed using [Code Carbon](https://mlco2.github.io/codecarbon/), an open-source Python package for tracking the environmental impacts of code.
We developed and used a [Docker container](https://github.com/huggingface/EnergyStarAI/) to maximize the reproducibility of results, and to enable members of the community to benchmark internal models.
Reach out to us if you want to collaborate!
""")
demo.launch()