__all__ = ['block', 'make_clickable_model', 'make_clickable_user', 'get_submissions'] import gradio as gr import pandas as pd import json import pdb import tempfile from constants import * from src.auto_leaderboard.model_metadata_type import ModelType global data_component, filter_component def upload_file(files): file_paths = [file.name for file in files] return file_paths def prediction_analyse(prediction_content): # pdb.set_trace() predictions = prediction_content.split("\n") # 读取 ground_truth JSON 文件 with open("./file/SEED-Bench.json", "r") as file: ground_truth_data = json.load(file)["questions"] # 将 ground_truth 数据转换为以 question_id 为键的字典 ground_truth = {item["question_id"]: item for item in ground_truth_data} # 初始化结果统计字典 results = {i: {"correct": 0, "total": 0} for i in range(1, 13)} # 遍历 predictions，计算每个 question_type_id 的正确预测数和总预测数 for prediction in predictions: # pdb.set_trace() prediction = prediction.strip() if not prediction: continue try: prediction = json.loads(prediction) except json.JSONDecodeError: print(f"Warning: Skipping invalid JSON data in line: {prediction}") continue question_id = prediction["question_id"] gt_item = ground_truth[question_id] question_type_id = gt_item["question_type_id"] if prediction["prediction"] == gt_item["answer"]: results[question_type_id]["correct"] += 1 results[question_type_id]["total"] += 1 return results def add_new_eval( input_file, model_name_textbox: str, revision_name_textbox: str, model_type: str, model_link: str, LLM_type: str, LLM_name_textbox: str, Evaluation_dimension: str, ): if input_file is None: return "Error! Empty file!" else: content = input_file.decode("utf-8") prediction = prediction_analyse(content) each_task_accuracy = {i: round(prediction[i]["correct"] / prediction[i]["total"] * 100, 1) for i in range(1, 13)} # count for average image\video\all total_correct_image = sum(prediction[i]["correct"] for i in range(1, 10)) total_correct_video = sum(prediction[i]["correct"] for i in range(10, 13)) total_image = sum(prediction[i]["total"] for i in range(1, 10)) total_video = sum(prediction[i]["total"] for i in range(10, 13)) average_accuracy_image = round(total_correct_image / total_image * 100, 1) average_accuracy_video = round(total_correct_video / total_video * 100, 1) overall_accuracy = round((total_correct_image + total_correct_video) / (total_image + total_video) * 100, 1) if LLM_type == 'other': LLM_name = LLM_name_textbox else: LLM_name = LLM_type if model_link == '': model_name = model_name_textbox # no url else: model_name = '[' + model_name_textbox + '](' + model_link + ')' # add new data new_data = [ model_type, model_name, LLM_name, each_task_accuracy[1], each_task_accuracy[2], each_task_accuracy[3], each_task_accuracy[4], each_task_accuracy[5], each_task_accuracy[6], each_task_accuracy[7], each_task_accuracy[8], each_task_accuracy[9], average_accuracy_image, each_task_accuracy[10], each_task_accuracy[11], each_task_accuracy[12], average_accuracy_video, overall_accuracy] # pdb.set_trace() csv_data = pd.read_csv(CSV_DIR) # pdb.set_trace() if revision_name_textbox == '': col = csv_data.shape[0] csv_data.loc[col] = new_data csv_data = csv_data.to_csv(CSV_DIR, index=False) else: model_name_list = csv_data['Model'] 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) csv_data.loc[col] = new_data csv_data = csv_data.to_csv(CSV_DIR, index=False) return 0 def get_baseline_df(): df = pd.read_csv(CSV_DIR) return df block = gr.Blocks() with block: gr.Markdown( LEADERBORAD_INTRODUCTION ) with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("🏅 SEED Benchmark", elem_id="seed-benchmark-tab-table", id=0): 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", ).style(show_copy_button=True) gr.Markdown( TABLE_INTRODUCTION ) # selection for column part: checkbox_group = gr.CheckboxGroup( choices=TASK_INFO, value=TASK_INFO, label="Select options", interactive=True, ) # 创建数据帧组件 data_component = gr.components.Dataframe( value=get_baseline_df, headers=COLUMN_NAMES, type="pandas", datatype=DATA_TITILE_TYPE, interactive=False, visible=True, ) def on_checkbox_group_change(selected_columns): # pdb.set_trace() selected_columns = [item for item in TASK_INFO if item in selected_columns] present_columns = MODEL_INFO + selected_columns updated_data = get_baseline_df()[present_columns] updated_headers = present_columns update_datatype = [DATA_TITILE_TYPE[COLUMN_NAMES.index(x)] for x in updated_headers] filter_component = gr.components.Dataframe( value=updated_data, headers=updated_headers, type="pandas", datatype=update_datatype, interactive=False, visible=True, ) # pdb.set_trace() return filter_component.value # 将复选框组关联到处理函数 checkbox_group.change(fn=on_checkbox_group_change, inputs=checkbox_group, outputs=data_component) # table 2 with gr.TabItem("📝 About", elem_id="seed-benchmark-tab-table", id=2): gr.Markdown(LEADERBORAD_INFO, elem_classes="markdown-text") # table 3 with gr.TabItem("🚀 Submit here! ", elem_id="seed-benchmark-tab-table", id=3): gr.Markdown(LEADERBORAD_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="LLaMA-7B" ) revision_name_textbox = gr.Textbox( label="Revision Model Name", placeholder="LLaMA-7B" ) model_type = gr.Dropdown( choices=[ "LLM", "ImageLLM", "VideoLLM", "Other", ], label="Model type", multiselect=False, value="ImageLLM", interactive=True, ) model_link = gr.Textbox( label="Model Link", placeholder="https://huggingface.co/decapoda-research/llama-7b-hf" ) with gr.Column(): LLM_type = gr.Dropdown( choices=["Vicuna-7B", "Flan-T5-XL", "LLaMA-7B", "other"], label="LLM type", multiselect=False, value="LLaMA-7B", interactive=True, ) LLM_name_textbox = gr.Textbox( label="LLM model (for other)", placeholder="LLaMA-13B" ) Evaluation_dimension = gr.Dropdown( choices=["All", "Image", "Video"], label="Evaluation dimension", multiselect=False, value="All", interactive=True, ) with gr.Column(): input_file = gr.inputs.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_type, model_link, LLM_type, LLM_name_textbox, Evaluation_dimension, ], # outputs = submission_result, ) with gr.Row(): data_run = gr.Button("Refresh") data_run.click( get_baseline_df, outputs=data_component ) # block.load(get_baseline_df, outputs=data_title) block.launch()