ynhe's picture
Update app.py
446db64 verified
raw
history blame
14.8 kB
__all__ = ['block', 'make_clickable_model', 'make_clickable_user', 'get_submissions']
import os
import gradio as gr
import pandas as pd
import json
import tempfile
from constants import *
from huggingface_hub import Repository
HF_TOKEN = os.environ.get("HF_TOKEN")
global data_component, filter_component
def upload_file(files):
file_paths = [file.name for file in files]
return file_paths
def add_new_eval(
input_file,
model_name_textbox: str,
revision_name_textbox: str,
model_link: str,
):
if input_file is None:
return "Error! Empty file!"
upload_data=json.loads(input_file)
submission_repo = Repository(local_dir=SUBMISSION_NAME, clone_from=SUBMISSION_URL, use_auth_token=HF_TOKEN, repo_type="dataset")
submission_repo.git_pull()
shutil.copyfile(CSV_DIR, os.path.join(SUBMISSION_NAME, f"{input_file}"))
csv_data = pd.read_csv(CSV_DIR)
if revision_name_textbox == '':
col = csv_data.shape[0]
model_name = model_name_textbox
else:
model_name = revision_name_textbox
model_name_list = csv_data['Model Name (clickable)']
name_list = [name.split(']')[0][1:] for name in model_name_list]
if revision_name_textbox not in name_list:
col = csv_data.shape[0]
else:
col = name_list.index(revision_name_textbox)
if model_link == '':
model_name = model_name # no url
else:
model_name = '[' + model_name + '](' + model_link + ')'
# add new data
new_data = [
model_name
]
for key in TASK_INFO:
if key in upload_data:
new_data.append(upload_data[key][0])
else:
new_data.append(0)
csv_data.loc[col] = new_data
csv_data = csv_data.to_csv(CSV_DIR, index=False)
submission_repo.push_to_hub()
return 0
def get_normalized_df(df):
# final_score = df.drop('name', axis=1).sum(axis=1)
# df.insert(1, 'Overall Score', final_score)
normalize_df = df.copy().fillna(0.0)
for column in normalize_df.columns[1:]:
min_val = NORMALIZE_DIC[column]['Min']
max_val = NORMALIZE_DIC[column]['Max']
print(normalize_df[column], min_val, max_val)
normalize_df[column] = (normalize_df[column] - min_val) / (max_val - min_val)
return normalize_df
def calculate_selected_score(df, selected_columns):
# selected_score = df[selected_columns].sum(axis=1)
selected_QUALITY = [i for i in selected_columns if i in QUALITY_LIST]
selected_SEMANTIC = [i for i in selected_columns if i in SEMANTIC_LIST]
selected_quality_score = df[selected_QUALITY].sum(axis=1)/sum([DIM_WEIGHT[i] for i in selected_QUALITY])
selected_semantic_score = df[selected_SEMANTIC].sum(axis=1)/sum([DIM_WEIGHT[i] for i in selected_SEMANTIC ])
if selected_quality_score.isna().any().any() and selected_semantic_score.isna().any().any():
selected_score = (selected_quality_score * QUALITY_WEIGHT + selected_semantic_score * SEMANTIC_WEIGHT) / (QUALITY_WEIGHT + SEMANTIC_WEIGHT)
return selected_score.fillna(0.0)
if selected_quality_score.isna().any().any():
return selected_semantic_score
if selected_semantic_score.isna().any().any():
return selected_quality_score
# print(selected_semantic_score,selected_quality_score )
selected_score = (selected_quality_score * QUALITY_WEIGHT + selected_semantic_score * SEMANTIC_WEIGHT) / (QUALITY_WEIGHT + SEMANTIC_WEIGHT)
return selected_score.fillna(0.0)
def get_final_score(df, selected_columns):
normalize_df = get_normalized_df(df)
#final_score = normalize_df.drop('name', axis=1).sum(axis=1)
for name in normalize_df.drop('Model Name (clickable)', axis=1):
normalize_df[name] = normalize_df[name]*DIM_WEIGHT[name]
quality_score = normalize_df[QUALITY_LIST].sum(axis=1)/sum([DIM_WEIGHT[i] for i in QUALITY_LIST])
semantic_score = normalize_df[SEMANTIC_LIST].sum(axis=1)/sum([DIM_WEIGHT[i] for i in SEMANTIC_LIST ])
final_score = (quality_score * QUALITY_WEIGHT + semantic_score * SEMANTIC_WEIGHT) / (QUALITY_WEIGHT + SEMANTIC_WEIGHT)
if 'Total Score' in df:
df['Total Score'] = final_score
else:
df.insert(1, 'Total Score', final_score)
if 'Semantic Score' in df:
df['Semantic Score'] = semantic_score
else:
df.insert(2, 'Semantic Score', semantic_score)
if 'Quality Score' in df:
df['Quality Score'] = quality_score
else:
df.insert(3, 'Quality Score', quality_score)
selected_score = calculate_selected_score(normalize_df, selected_columns)
if 'Selected Score' in df:
df['Selected Score'] = selected_score
else:
df.insert(1, 'Selected Score', selected_score)
return df
def get_final_score_quality(df, selected_columns):
normalize_df = get_normalized_df(df)
quality_score = normalize_df[QUALITY_TAB].sum(axis=1) / len(QUALITY_TAB)
if 'Quality Score' in df:
df['Quality Score'] = quality_score
else:
df.insert(1, 'Quality Score', quality_score)
selected_score = normalize_df[selected_columns].sum(axis=1) / len(selected_columns)
if 'Selected Score' in df:
df['Selected Score'] = selected_score
else:
df.insert(1, 'Selected Score', selected_score)
return df
def get_baseline_df():
submission_repo = Repository(local_dir=SUBMISSION_NAME, clone_from=SUBMISSION_URL, use_auth_token=HF_TOKEN, repo_type="dataset")
submission_repo.git_pull()
df = pd.read_csv(CSV_DIR)
df = get_final_score(df, checkbox_group.value)
df = df.sort_values(by="Selected Score", ascending=False)
present_columns = MODEL_INFO + checkbox_group.value
df = df[present_columns]
df = convert_scores_to_percentage(df)
return df
def get_baseline_df_quality():
submission_repo = Repository(local_dir=SUBMISSION_NAME, clone_from=SUBMISSION_URL, use_auth_token=HF_TOKEN, repo_type="dataset")
submission_repo.git_pull()
df = pd.read_csv(QUALITY_DIR)
df = get_final_score_quality(df, checkbox_group_quality.value)
df = df.sort_values(by="Selected Score", ascending=False)
present_columns = MODEL_INFO_TAB_QUALITY + checkbox_group_quality.value
df = df[present_columns]
df = convert_scores_to_percentage(df)
return df
def get_all_df(selected_columns):
submission_repo = Repository(local_dir=SUBMISSION_NAME, clone_from=SUBMISSION_URL, use_auth_token=HF_TOKEN, repo_type="dataset")
submission_repo.git_pull()
df = pd.read_csv(CSV_DIR)
df = get_final_score(df, selected_columns)
df = df.sort_values(by="Selected Score", ascending=False)
return df
def convert_scores_to_percentage(df):
# 对DataFrame中的每一列(除了'name'列)进行操作
for column in df.columns[1:]: # 假设第一列是'name'
df[column] = round(df[column] * 100,2) # 将分数转换为百分数
df[column] = df[column].astype(str) + '%'
return df
def choose_all_quailty():
return gr.update(value=QUALITY_LIST)
def choose_all_semantic():
return gr.update(value=SEMANTIC_LIST)
def disable_all():
return gr.update(value=[])
def enable_all():
return gr.update(value=TASK_INFO)
def on_filter_model_size_method_change(selected_columns):
updated_data = get_all_df(selected_columns)
#print(updated_data)
# columns:
selected_columns = [item for item in TASK_INFO if item in selected_columns]
present_columns = MODEL_INFO + selected_columns
updated_data = updated_data[present_columns]
updated_data = updated_data.sort_values(by="Selected Score", ascending=False)
updated_data = convert_scores_to_percentage(updated_data)
updated_headers = present_columns
update_datatype = [DATA_TITILE_TYPE[COLUMN_NAMES.index(x)] for x in updated_headers]
# print(updated_data,present_columns,update_datatype)
filter_component = gr.components.Dataframe(
value=updated_data,
headers=updated_headers,
type="pandas",
datatype=update_datatype,
interactive=False,
visible=True,
)
return filter_component#.value
def on_filter_model_size_method_change_quality(selected_columns):
updated_data = get_baseline_df_quality()
#print(updated_data)
# columns:
selected_columns = [item for item in QUALITY_TAB if item in selected_columns]
present_columns = MODEL_INFO_TAB_QUALITY + selected_columns
updated_data = updated_data[present_columns]
updated_data = updated_data.sort_values(by="Selected Score", ascending=False)
updated_data = convert_scores_to_percentage(updated_data)
updated_headers = present_columns
update_datatype = [DATA_TITILE_TYPE[COLUMN_NAMES.index(x)] for x in updated_headers]
# print(updated_data,present_columns,update_datatype)
filter_component = gr.components.Dataframe(
value=updated_data,
headers=updated_headers,
type="pandas",
datatype=update_datatype,
interactive=False,
visible=True,
)
return filter_component#.value
block = gr.Blocks()
with block:
gr.Markdown(
LEADERBORAD_INTRODUCTION
)
with gr.Tabs(elem_classes="tab-buttons") as tabs:
# Table 0
with gr.TabItem("📊 VBench", elem_id="vbench-tab-table", id=1):
with gr.Row():
with gr.Accordion("Citation", open=False):
citation_button = gr.Textbox(
value=CITATION_BUTTON_TEXT,
label=CITATION_BUTTON_LABEL,
elem_id="citation-button",
lines=10,
)
gr.Markdown(
TABLE_INTRODUCTION
)
with gr.Row():
with gr.Column(scale=0.2):
choosen_q = gr.Button("Select Quality Dimensions")
choosen_s = gr.Button("Select Semantic Dimensions")
# enable_b = gr.Button("Select All")
disable_b = gr.Button("Deselect All")
with gr.Column(scale=0.8):
# selection for column part:
checkbox_group = gr.CheckboxGroup(
choices=TASK_INFO,
value=DEFAULT_INFO,
label="Evaluation Dimension",
interactive=True,
)
data_component = gr.components.Dataframe(
value=get_baseline_df,
headers=COLUMN_NAMES,
type="pandas",
datatype=DATA_TITILE_TYPE,
interactive=False,
visible=True,
)
choosen_q.click(choose_all_quailty, inputs=None, outputs=[checkbox_group]).then(fn=on_filter_model_size_method_change, inputs=[ checkbox_group], outputs=data_component)
choosen_s.click(choose_all_semantic, inputs=None, outputs=[checkbox_group]).then(fn=on_filter_model_size_method_change, inputs=[ checkbox_group], outputs=data_component)
# enable_b.click(enable_all, inputs=None, outputs=[checkbox_group]).then(fn=on_filter_model_size_method_change, inputs=[ checkbox_group], outputs=data_component)
disable_b.click(disable_all, inputs=None, outputs=[checkbox_group]).then(fn=on_filter_model_size_method_change, inputs=[ checkbox_group], outputs=data_component)
checkbox_group.change(fn=on_filter_model_size_method_change, inputs=[ checkbox_group], outputs=data_component)
with gr.TabItem("Video Quaity", elem_id="vbench-tab-table", id=2):
with gr.Accordion("INSTRUCTION", open=False):
citation_button = gr.Textbox(
value=QUALITY_CLAIM_TEXT,
label="",
elem_id="quality-button",
lines=10,
)
with gr.Row():
with gr.Column(scale=1.0):
# selection for column part:
checkbox_group_quality = gr.CheckboxGroup(
choices=QUALITY_TAB,
value=QUALITY_TAB,
label="Evaluation Quality Dimension",
interactive=True,
)
data_component_quality = gr.components.Dataframe(
value=get_baseline_df_quality,
headers=COLUMN_NAMES_QUALITY,
type="pandas",
datatype=DATA_TITILE_TYPE,
interactive=False,
visible=True,
)
checkbox_group_quality.change(fn=on_filter_model_size_method_change_quality, inputs=[checkbox_group_quality], outputs=data_component_quality)
# table 2
with gr.TabItem("📝 About", elem_id="mvbench-tab-table", id=3):
gr.Markdown(LEADERBORAD_INFO, elem_classes="markdown-text")
# table 3
with gr.TabItem("🚀 Submit here! ", elem_id="mvbench-tab-table", id=4):
gr.Markdown(LEADERBORAD_INTRODUCTION, elem_classes="markdown-text")
with gr.Row():
gr.Markdown(SUBMIT_INTRODUCTION, elem_classes="markdown-text")
with gr.Row():
gr.Markdown("# ✉️✨ Submit your model evaluation json file here!", elem_classes="markdown-text")
with gr.Row():
with gr.Column():
model_name_textbox = gr.Textbox(
label="Model name", placeholder="LaVie"
)
revision_name_textbox = gr.Textbox(
label="Revision Model Name", placeholder="LaVie"
)
with gr.Column():
model_link = gr.Textbox(
label="Model Link", placeholder="https://huggingface.co/decapoda-research/llama-7b-hf"
)
with gr.Column():
input_file = gr.components.File(label = "Click to Upload a json File", file_count="single", type='binary')
submit_button = gr.Button("Submit Eval")
submission_result = gr.Markdown()
submit_button.click(
add_new_eval,
inputs = [
input_file,
model_name_textbox,
revision_name_textbox,
model_link,
],
)
def refresh_data():
value1 = get_baseline_df()
return value1
with gr.Row():
data_run = gr.Button("Refresh")
data_run.click(on_filter_model_size_method_change, inputs=[checkbox_group], outputs=data_component)
block.launch()