benchmark / app.py
cdminix's picture
Update app.py
63e255e verified
raw
history blame
11.9 kB
from pathlib import Path
import json
import os
import gradio as gr
from huggingface_hub import snapshot_download
from gradio_leaderboard import Leaderboard, SelectColumns
import pandas as pd
from apscheduler.schedulers.background import BackgroundScheduler
from ttsds.benchmarks.benchmark import BenchmarkCategory
from ttsds import BenchmarkSuite
from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN, TAGS
from src.texts import LLM_BENCHMARKS_TEXT, EVALUATION_QUEUE_TEXT, CITATION_TEXT
from src.css_html_js import custom_css
def filter_dfs(tags, lb):
global f_b_df, f_a_df
is_agg = False
if "Environment" in lb.columns:
is_agg = True
if is_agg:
lb = f_a_df.copy()
else:
lb = f_b_df.copy()
if tags and len(lb) > 0:
lb = lb[lb["Tags"].apply(lambda x: any(tag in x for tag in tags))]
lb = rounded_df(lb)
return lb
def change_mean(env, lb):
global f_b_df, f_a_df
lb = f_a_df.copy()
if env:
mean_cols = [col for col in lb.columns if str(col) not in ["Mean", "Environment", "Model", "Tags"]]
else:
mean_cols = [col for col in lb.columns if str(col) not in ["Mean", "Model", "Tags"]]
lb["Mean"] = lb[mean_cols].mean(axis=1)
lb = rounded_df(lb)
return lb
def restart_space():
API.restart_space(repo_id=REPO_ID)
def submit_eval(model_name, model_tags, web_url, hf_url, code_url, paper_url, inference_details, file_path):
model_id = model_name.lower().replace(" ", "_")
# check if model already exists
if Path(f"{EVAL_REQUESTS_PATH}/{model_id}.json").exists():
return "Model already exists in the evaluation queue"
# check which urls are valid
if web_url and not web_url.startswith("http"):
return "Please enter a valid URL"
if hf_url and not hf_url.startswith("http"):
return "Please enter a valid URL"
if code_url and not code_url.startswith("http"):
return "Please enter a valid URL"
if paper_url and not paper_url.startswith("http"):
return "Please enter a valid URL"
# move file to correct location
if not file_path.endswith(".tar.gz"):
return "Please upload a .tar.gz file"
Path(file_path).rename(f"{EVAL_REQUESTS_PATH}/{model_id}.tar.gz")
# build display name - use web_url to link text if available, and emojis for the other urls
display_name = model_name + " "
if web_url:
display_name = f"[{display_name}]({web_url}) "
if hf_url:
display_name += f"[πŸ€—]({hf_url})"
if code_url:
display_name += f"[πŸ’»]({code_url})"
if paper_url:
display_name += f"[πŸ“„]({paper_url})"
request_obj = {
"model_name": model_name,
"display_name": display_name,
"model_tags": model_tags,
"web_url": web_url,
"hf_url": hf_url,
"code_url": code_url,
"paper_url": paper_url,
"inference_details": inference_details,
"status": "pending",
}
try:
with open(f"{EVAL_REQUESTS_PATH}/{model_id}.json", "w") as f:
json.dump(request_obj, f)
API.upload_file(
path_or_fileobj=f"{EVAL_REQUESTS_PATH}/{model_id}.json",
path_in_repo=f"{model_id}.json",
repo_id=QUEUE_REPO,
repo_type="dataset",
commit_message=f"Add {model_name} to evaluation queue",
)
API.upload_file(
path_or_fileobj=f"{EVAL_REQUESTS_PATH}/{model_id}.tar.gz",
path_in_repo=f"{model_id}.tar.gz",
repo_id=QUEUE_REPO,
repo_type="dataset",
commit_message=f"Add {model_name} to evaluation queue",
)
except error as e:
os.remove(f"{EVAL_REQUESTS_PATH}/{model_id}.json")
return f"Error: {e}"
return "Model submitted successfully πŸŽ‰"
### Space initialisation
try:
print(EVAL_REQUESTS_PATH)
snapshot_download(
repo_id=QUEUE_REPO,
local_dir=EVAL_REQUESTS_PATH,
repo_type="dataset",
tqdm_class=None,
etag_timeout=30,
token=TOKEN,
)
except Exception:
restart_space()
try:
print(EVAL_RESULTS_PATH)
snapshot_download(
repo_id=RESULTS_REPO,
local_dir=EVAL_RESULTS_PATH,
repo_type="dataset",
tqdm_class=None,
etag_timeout=30,
token=TOKEN,
)
except Exception:
restart_space()
def rounded_df(df):
df = df.copy()
for col in df.columns:
if isinstance(df[col].values[0], float):
df[col] = df[col].apply(lambda x: round(x, 2))
return df
results_df = pd.read_csv(EVAL_RESULTS_PATH + "/results.csv")
agg_df = BenchmarkSuite.aggregate_df(results_df)
agg_df = agg_df.pivot(index="dataset", columns="benchmark_category", values="score")
agg_df.rename(columns={"OVERALL": "General"}, inplace=True)
agg_df.columns = [x.capitalize() for x in agg_df.columns]
mean_cols = [col for col in agg_df.columns if str(col) not in ["Mean", "Environment", "Model", "Tags"]]
agg_df["Mean"] = agg_df[mean_cols].mean(axis=1)
# make sure mean is the first column
agg_df = agg_df[["Mean"] + [col for col in agg_df.columns if col != "Mean"]]
agg_df["Tags"] = ""
agg_df.reset_index(inplace=True)
agg_df.rename(columns={"dataset": "Model"}, inplace=True)
agg_df.sort_values("Mean", ascending=False, inplace=True)
benchmark_df = results_df.pivot(index="dataset", columns="benchmark_name", values="score")
# get benchmark name order by category
benchmark_order = list(results_df.sort_values("benchmark_category")["benchmark_name"].unique())
benchmark_df = benchmark_df[benchmark_order]
benchmark_df = benchmark_df.reset_index()
benchmark_df.rename(columns={"dataset": "Model"}, inplace=True)
# set index
benchmark_df.set_index("Model", inplace=True)
benchmark_df["Mean"] = benchmark_df.mean(axis=1)
# make sure mean is the first column
benchmark_df = benchmark_df[["Mean"] + [col for col in benchmark_df.columns if col != "Mean"]]
benchmark_df["Tags"] = ""
benchmark_df.reset_index(inplace=True)
benchmark_df.sort_values("Mean", ascending=False, inplace=True)
# get details for each model
model_detail_files = Path(EVAL_REQUESTS_PATH).glob("*.json")
model_details = {}
for model_detail_file in model_detail_files:
with open(model_detail_file) as f:
model_detail = json.load(f)
model_details[model_detail_file.stem] = model_detail
# replace .tar.gz
benchmark_df["Model"] = benchmark_df["Model"].apply(lambda x: x.replace(".tar.gz", ""))
agg_df["Model"] = agg_df["Model"].apply(lambda x: x.replace(".tar.gz", ""))
benchmark_df["Tags"] = benchmark_df["Model"].apply(lambda x: model_details.get(x, {}).get("model_tags", ""))
agg_df["Tags"] = agg_df["Model"].apply(lambda x: model_details.get(x, {}).get("model_tags", ""))
benchmark_df["Model"] = benchmark_df["Model"].apply(lambda x: model_details.get(x, {}).get("display_name", x))
agg_df["Model"] = agg_df["Model"].apply(lambda x: model_details.get(x, {}).get("display_name", x))
f_b_df = benchmark_df.copy()
f_a_df = agg_df.copy()
def init_leaderboard(dataframe):
if dataframe is None or dataframe.empty:
raise ValueError("Leaderboard DataFrame is empty or None.")
df_types = []
for col in dataframe.columns:
if col == "Model":
df_types.append("markdown")
elif col == "Tags":
df_types.append("markdown")
else:
df_types.append("number")
cols = list(dataframe.columns)
cols.remove("Tags")
return Leaderboard(
value=rounded_df(dataframe),
select_columns=SelectColumns(
default_selection=cols,
cant_deselect=["Model", "Mean"],
label="Select Columns to Display:",
),
search_columns=["Model", "Tags"],
filter_columns=[],
interactive=False,
datatype=df_types,
)
app = gr.Blocks(css=custom_css, title="TTS Benchmark Leaderboard")
with app:
with gr.Tabs(elem_classes="tab-buttons") as tabs:
with gr.TabItem("πŸ… TTSDS Scores", elem_id="llm-benchmark-tab-table", id=0):
with gr.Group():
env = gr.Checkbox(value=True, label="Exclude environment from mean.")
gr.Markdown("**Environment** measures how well the system can reproduce noise in the training data. This doesn't correlate with human judgements for 'naturalness'")
tags = gr.Dropdown(
TAGS,
value=[],
multiselect=True,
label="Tags",
info="Select tags to filter the leaderboard. You can suggest new tags here: https://huggingface.co/spaces/ttsds/benchmark/discussions/1",
)
leaderboard = init_leaderboard(f_a_df)
tags.change(filter_dfs, [tags, leaderboard], [leaderboard])
env.change(change_mean, [env, leaderboard], [leaderboard])
with gr.TabItem("πŸ… Individual Benchmarks", elem_id="llm-benchmark-tab-table", id=1):
tags = gr.Dropdown(
TAGS,
value=[],
multiselect=True,
label="Tags",
info="Select tags to filter the leaderboard",
)
leaderboard = init_leaderboard(f_b_df)
tags.change(filter_dfs, [tags, leaderboard], [leaderboard])
with gr.TabItem("πŸ“ About", elem_id="llm-benchmark-tab-table", id=2):
gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text")
with gr.TabItem("πŸš€ Submit here!", elem_id="llm-benchmark-tab-table", id=3):
with gr.Column():
with gr.Row():
gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text")
with gr.Row():
gr.Markdown("# βœ‰οΈβœ¨ Submit a TTS dataset here!", elem_classes="markdown-text")
with gr.Row():
with gr.Column():
model_name_textbox = gr.Textbox(label="Model name")
model_tags_dropdown = gr.Dropdown(
label="Model tags",
choices=TAGS,
multiselect=True,
)
website_url_textbox = gr.Textbox(label="Website URL (optional)")
hf_url_textbox = gr.Textbox(label="Huggingface URL (optional)")
code_url_textbox = gr.Textbox(label="Code URL (optional)")
paper_url_textbox = gr.Textbox(label="Paper URL (optional)")
inference_details_textbox = gr.TextArea(label="Inference details (optional)")
file_input = gr.File(file_types=[".gz"], interactive=True, label=".tar.gz TTS dataset")
submit_button = gr.Button("Submit Eval")
submission_result = gr.Markdown()
submit_button.click(
submit_eval,
[
model_name_textbox,
model_tags_dropdown,
website_url_textbox,
hf_url_textbox,
code_url_textbox,
paper_url_textbox,
inference_details_textbox,
file_input,
],
submission_result,
)
with gr.Row():
with gr.Accordion("Citation", open=False):
gr.Markdown(f"Copy the BibTeX citation to cite this source:\n\n```bibtext\n{CITATION_TEXT}\n```")
scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "interval", seconds=5*86400)
scheduler.start()
app.queue(default_concurrency_limit=40).launch()