Initial model commit

README.md

@@ -1,10 +1,10 @@
 ---
-title: ViP-Bench Evaluator
-emoji: 🌖
-colorFrom: indigo
-colorTo: blue
+title: MM-Vet Evaluator
+emoji: 🐨
+colorFrom: gray
+colorTo: gray
 sdk: gradio
-sdk_version: 4.9.0
+sdk_version: 3.39.0
 app_file: app.py
 pinned: false
 license: apache-2.0

app.py

@@ -0,0 +1,325 @@
+import gradio as gr
+import openai
+import json
+import os
+from tqdm import tqdm
+import pandas as pd
+import numpy as np
+from collections import Counter
+import time
+from zipfile import ZipFile
+
+openai.api_key = os.environ.get("AZURE_OPENAI_KEY")
+openai.api_base = os.environ.get("AZURE_OPENAI_ENDPOINT") 
+openai.api_type = 'azure'
+openai.api_version = os.environ.get("AZURE_OPENAI_API_VERSION")
+deployment_id = os.environ.get("AZURE_OPENAI_DEP_ID")
+gpt_model = deployment_id
+
+
+
+
+prompt = """Compare the ground truth and prediction from AI models, to give a correctness score for the prediction. <AND> in the ground truth means it is totally right only when all elements in the ground truth are present in the prediction, and <OR> means it is totally right when any one element in the ground truth is present in the prediction. The correctness score is 0.0 (totally wrong), 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, or 1.0 (totally right). Just complete the last space of the correctness score.
+
+Question | Ground truth | Prediction | Correctness
+--- | --- | --- | ---
+What is x in the equation? | -1 <AND> -5 | x = 3 | 0.0
+What is x in the equation? | -1 <AND> -5 | x = -1 | 0.5
+What is x in the equation? | -1 <AND> -5 | x = -5 | 0.5
+What is x in the equation? | -1 <AND> -5 | x = -5 or 5 | 0.5
+What is x in the equation? | -1 <AND> -5 | x = -1 or x = -5 | 1.0
+Can you explain this meme? | This meme is poking fun at the fact that the names of the countries Iceland and Greenland are misleading. Despite its name, Iceland is known for its beautiful green landscapes, while Greenland is mostly covered in ice and snow. The meme is saying that the person has trust issues because the names of these countries do not accurately represent their landscapes. | The meme talks about Iceland and Greenland. It's pointing out that despite their names, Iceland is not very icy and Greenland isn't very green. | 0.4
+Can you explain this meme? | This meme is poking fun at the fact that the names of the countries Iceland and Greenland are misleading. Despite its name, Iceland is known for its beautiful green landscapes, while Greenland is mostly covered in ice and snow. The meme is saying that the person has trust issues because the names of these countries do not accurately represent their landscapes. | The meme is using humor to point out the misleading nature of Iceland's and Greenland's names. Iceland, despite its name, has lush green landscapes while Greenland is mostly covered in ice and snow. The text 'This is why I have trust issues' is a playful way to suggest that these contradictions can lead to distrust or confusion. The humor in this meme is derived from the unexpected contrast between the names of the countries and their actual physical characteristics. | 1.0
+"""
+
+
+
+def grade(file_obj, progress=gr.Progress()):
+    # load metadata
+    # Download mm-vet.zip and `unzip mm-vet.zip` and change the path below
+    mmvet_path = "mm-vet"
+    use_sub_set = False
+    decimal_places = 1 # number of decimal places to round to
+
+
+    if use_sub_set:
+        bard_set_file = os.path.join(mmvet_path, "bard_set.json")
+        with open(bard_set_file, 'r') as f:
+            sub_set = json.load(f)
+        sub_set_name = 'bardset'
+        sub_set_name = sub_set_name + '_'
+    else:
+        sub_set = None
+        sub_set_name = ''
+
+    mmvet_metadata = os.path.join(mmvet_path, "mm-vet.json")
+    with open(mmvet_metadata, 'r') as f:
+        data = json.load(f)
+
+
+    counter = Counter()
+    cap_set_list = []
+    cap_set_counter = []
+    len_data = 0
+    for id, value in data.items():
+        if sub_set is not None and id not in sub_set:
+            continue
+        question = value["question"]
+        answer = value["answer"]
+        cap = value["capability"]
+        cap = set(cap)
+        counter.update(cap)
+        if cap not in cap_set_list:
+            cap_set_list.append(cap)
+            cap_set_counter.append(1)
+        else:
+            cap_set_counter[cap_set_list.index(cap)] += 1
+        
+        len_data += 1
+
+    sorted_list = counter.most_common()
+    columns = [k for k, v in sorted_list]
+    columns.append("total")
+    columns.append("std")
+    columns.append('runs')
+    df = pd.DataFrame(columns=columns)
+
+
+    cap_set_sorted_indices = np.argsort(-np.array(cap_set_counter))
+    new_cap_set_list = []
+    new_cap_set_counter = []
+    for index in cap_set_sorted_indices:
+        new_cap_set_list.append(cap_set_list[index])
+        new_cap_set_counter.append(cap_set_counter[index])
+
+    cap_set_list = new_cap_set_list
+    cap_set_counter = new_cap_set_counter
+    cap_set_names = ["_".join(list(cap_set)) for cap_set in cap_set_list]
+
+    columns2 = cap_set_names
+    columns2.append("total")
+    columns2.append("std")
+    columns2.append('runs')
+    df2 = pd.DataFrame(columns=columns2)
+
+
+    ###### change your model name ######
+    model = file_obj.name.split("/")[-1][:-5]
+    # result_path = "results"
+    num_run = 1 # we set 5 in the paper
+    # model_results_file = os.path.join(result_path, f"{model}.json")
+    model_results_file = file_obj.name
+
+    # grade results for each sample to svae
+    grade_file = f'{model}_{gpt_model}-grade-{num_run}runs.json'
+    # grade_file = os.path.join(result_path, grade_file)
+
+    # score results regarding capabilities/capability integration to save
+    cap_score_file = f'{model}_{sub_set_name}{gpt_model}-cap-score-{num_run}runs.csv'
+    # cap_score_file = os.path.join(result_path, cap_score_file)
+    cap_int_score_file = f'{model}_{sub_set_name}{gpt_model}-cap-int-score-{num_run}runs.csv'
+    # cap_int_score_file = os.path.join(result_path, cap_int_score_file)
+
+
+
+    with open(model_results_file) as f:
+        results = json.load(f)
+    if os.path.exists(grade_file):
+        with open(grade_file, 'r') as f:
+            grade_results = json.load(f)
+    else:
+        grade_results = {}
+
+
+    def need_more_runs():
+        need_more_runs = False
+        if len(grade_results) > 0:
+            for k, v in grade_results.items():
+                if len(v['score']) < num_run:
+                    need_more_runs = True
+                    break
+        return need_more_runs or len(grade_results) < len_data
+
+
+    while need_more_runs():
+        for j in range(num_run):
+            print(f'eval run {j}')
+            for id, line in progress.tqdm(data.items(), desc="Grading"):
+                if sub_set is not None and id not in sub_set:
+                    continue
+                if id in grade_results and len(grade_results[id]['score']) >= (j + 1):
+                    continue
+
+                model_pred = results[id]
+                
+                question = prompt + '\n' + ' | '.join([line['question'], line['answer'].replace("<AND>", " <AND> ").replace("<OR>", " <OR> "), model_pred, ""])
+                messages = [
+                {"role": "user", "content": question},
+                ]
+
+                if id not in grade_results:
+                    sample_grade = {'model': [], 'content': [], 'score': []}
+                else:
+                    sample_grade = grade_results[id]
+
+                
+                grade_sample_run_complete = False
+                temperature = 0.0
+
+                num_sleep = 0
+                while not grade_sample_run_complete:
+                    try:
+                        response = openai.ChatCompletion.create(
+                            # model=gpt_model,
+                            engine=gpt_model,
+                            max_tokens=3,
+                            temperature=temperature,
+                            messages=messages)
+                        content = response['choices'][0]['message']['content']
+                        flag = True
+                        try_time = 1
+                        while flag:
+                            try:
+                                content = content.split(' ')[0].strip()
+                                score = float(content)
+                                if score > 1.0 or score < 0.0:
+                                    assert False
+                                flag = False
+                            except:
+                                question = prompt + '\n' + ' | '.join([line['question'], line['answer'].replace("<AND>", " <AND> ").replace("<OR>", " <OR> "), model_pred, ""]) + "\nPredict the correctness of the answer (digit): "
+                                messages = [
+                                {"role": "user", "content": question},
+                                ]
+                                response = openai.ChatCompletion.create(
+                                    # model=gpt_model,
+                                    engine=gpt_model,
+                                    max_tokens=3,
+                                    temperature=temperature,
+                                    messages=messages)
+                                content = response['choices'][0]['message']['content']
+                                try_time += 1
+                                temperature += 0.5
+                                print(f"{id} try {try_time} times")
+                                print(content)
+                                if try_time > 5:
+                                    score = 0.0
+                                    flag = False
+                        grade_sample_run_complete = True
+                    except:
+                        # gpt4 may have token rate limit
+                        num_sleep += 1
+                        if num_sleep > 12:
+                            score = 0.0
+                            grade_sample_run_complete = True
+                            num_sleep = 0
+                            continue
+                        print("sleep 5s")
+                        time.sleep(5)
+                        
+
+                if len(sample_grade['model']) >= j + 1:
+                    sample_grade['model'][j] = response['model']
+                    sample_grade['content'][j] = content
+                    sample_grade['score'][j] = score
+                else:
+                    sample_grade['model'].append(response['model'])
+                    sample_grade['content'].append(content)
+                    sample_grade['score'].append(score)
+                grade_results[id] = sample_grade
+
+                with open(grade_file, 'w') as f:
+                    json.dump(grade_results, f, indent=4)
+
+
+    assert not need_more_runs()
+    cap_socres = {k: [0.0]*num_run for k in columns[:-2]}
+    counter['total'] = len_data
+
+    cap_socres2 = {k: [0.0]*num_run for k in columns2[:-2]}
+    counter2 = {columns2[i]:cap_set_counter[i] for i in range(len(cap_set_counter))}
+    counter2['total'] = len_data
+
+    for k, v in grade_results.items():
+        if sub_set is not None and k not in sub_set:
+            continue
+        for i in range(num_run):
+            score = v['score'][i]
+            caps = set(data[k]['capability'])
+            for c in caps:
+                cap_socres[c][i] += score
+            
+            cap_socres['total'][i] += score
+
+            index = cap_set_list.index(caps)
+            cap_socres2[cap_set_names[index]][i] += score
+            cap_socres2['total'][i] += score
+
+    for k, v in cap_socres.items():
+        cap_socres[k] = np.array(v) / counter[k] *100
+
+
+    std = round(cap_socres['total'].std(), decimal_places)
+    total_copy = cap_socres['total'].copy()
+    runs = str(list(np.round(total_copy, decimal_places)))
+
+    for k, v in cap_socres.items():
+        cap_socres[k] = round(v.mean(), decimal_places)
+
+    cap_socres['std'] = std
+    cap_socres['runs'] = runs
+    df.loc[model] = cap_socres
+
+
+    for k, v in cap_socres2.items():
+        cap_socres2[k] = round(np.mean(np.array(v) / counter2[k] *100), decimal_places)
+    cap_socres2['std'] = std
+    cap_socres2['runs'] = runs
+    df2.loc[model] = cap_socres2
+
+    df.to_csv(cap_score_file)
+    df2.to_csv(cap_int_score_file)
+
+    files = [cap_score_file, cap_int_score_file, grade_file]
+    zip_file = f"results.zip"
+    with ZipFile(zip_file, "w") as zipObj:
+        for idx, file in enumerate(files):
+            zipObj.write(file, file)
+    for file in files:
+        os.remove(file)
+    return zip_file
+
+
+
+# demo = gr.Interface(
+#     fn=grade, 
+#     inputs=gr.File(file_types=[".json"]), 
+#     outputs="file")
+
+
+markdown = """
+<p align="center">
+<img src="https://github-production-user-asset-6210df.s3.amazonaws.com/49296856/258254299-29c00dae-8201-4128-b341-dad4663b544a.jpg" width="400"> <br>
+</p>
+
+
+# [MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities](https://github.com/yuweihao/MM-Vet)
+
+In this demo, we offer MM-Vet LLM-based (GPT-4) evaluator to grade open-ended outputs from your models.
+
+Plese upload your json file of your model results containing `{v1_0: ..., v1_1: ..., }`like [this json file](https://raw.githubusercontent.com/yuweihao/MM-Vet/main/results/llava_llama2_13b_chat.json).
+
+The grading may last 5 minutes. Sine we only support 1 queue, the grading time may be longer when you need to wait for other users' grading to finish.
+
+The grading results will be downloaded as a zip file.
+"""
+
+
+with gr.Blocks() as demo:
+    gr.Markdown(markdown)
+    with gr.Row():
+        inp = gr.File(file_types=[".json"])
+        out = gr.File(file_types=[".zip"])
+    inp.change(grade, inp, out)
+
+if __name__ == "__main__":
+    demo.queue().launch()

mm-vet/bard_set.json

@@ -0,0 +1 @@
+["v1_0", "v1_1", "v1_2", "v1_3", "v1_4", "v1_5", "v1_6", "v1_7", "v1_9", "v1_10", "v1_11", "v1_12", "v1_13", "v1_14", "v1_15", "v1_16", "v1_18", "v1_19", "v1_20", "v1_21", "v1_22", "v1_23", "v1_24", "v1_27", "v1_28", "v1_29", "v1_30", "v1_31", "v1_32", "v1_33", "v1_34", "v1_35", "v1_36", "v1_37", "v1_38", "v1_39", "v1_40", "v1_41", "v1_42", "v1_43", "v1_44", "v1_45", "v1_46", "v1_47", "v1_48", "v1_49", "v1_50", "v1_51", "v1_52", "v1_53", "v1_54", "v1_55", "v1_56", "v1_57", "v1_58", "v1_59", "v1_60", "v1_61", "v1_62", "v1_63", "v1_67", "v1_68", "v1_69", "v1_70", "v1_71", "v1_72", "v1_73", "v1_74", "v1_75", "v1_76", "v1_77", "v1_78", "v1_80", "v1_81", "v1_82", "v1_83", "v1_84", "v1_85", "v1_86", "v1_87", "v1_88", "v1_89", "v1_90", "v1_91", "v1_93", "v1_94", "v1_95", "v1_96", "v1_97", "v1_98", "v1_99", "v1_100", "v1_101", "v1_102", "v1_106", "v1_107", "v1_109", "v1_110", "v1_111", "v1_113", "v1_114", "v1_116", "v1_117", "v1_118", "v1_119", "v1_121", "v1_122", "v1_123", "v1_124", "v1_125", "v1_126", "v1_128", "v1_129", "v1_132", "v1_133", "v1_134", "v1_137", "v1_138", "v1_139", "v1_140", "v1_142", "v1_148", "v1_150", "v1_151", "v1_152", "v1_153", "v1_156", "v1_157", "v1_158", "v1_159", "v1_160", "v1_161", "v1_162", "v1_163", "v1_164", "v1_165", "v1_166", "v1_167", "v1_177", "v1_179", "v1_180", "v1_181", "v1_182", "v1_183", "v1_184", "v1_185", "v1_186", "v1_187", "v1_188", "v1_189", "v1_190", "v1_192", "v1_193", "v1_194", "v1_195", "v1_196", "v1_197", "v1_201", "v1_208", "v1_209", "v1_210", "v1_211", "v1_212", "v1_213", "v1_214", "v1_215", "v1_216", "v1_217"]

requirements.txt

@@ -0,0 +1 @@
+openai