m

scripts/convert_gqa_for_eval.py

@@ -0,0 +1,18 @@
+import os
+import json
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--src", type=str)
+parser.add_argument("--dst", type=str)
+args = parser.parse_args()
+
+all_answers = []
+for line_idx, line in enumerate(open(args.src)):
+    res = json.loads(line)
+    question_id = res['question_id']
+    text = res['text'].rstrip('.').lower()
+    all_answers.append({"questionId": question_id, "prediction": text})
+
+with open(args.dst, 'w') as f:
+    json.dump(all_answers, f)

scripts/convert_mmbench_for_submission.py

@@ -0,0 +1,27 @@
+import os
+import json
+import argparse
+import pandas as pd
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--annotation-file", type=str, required=True)
+    parser.add_argument("--result-dir", type=str, required=True)
+    parser.add_argument("--upload-dir", type=str, required=True)
+    parser.add_argument("--experiment", type=str, required=True)
+
+    return parser.parse_args()
+
+if __name__ == "__main__":
+    args = get_args()
+
+    df = pd.read_table(args.annotation_file)
+
+    cur_df = df.copy()
+    cur_df = cur_df.drop(columns=['hint', 'category', 'source', 'image', 'comment', 'l2-category'])
+    cur_df.insert(6, 'prediction', None)
+    for pred in open(os.path.join(args.result_dir, f"{args.experiment}.jsonl")):
+        pred = json.loads(pred)
+        cur_df.loc[df['index'] == pred['question_id'], 'prediction'] = pred['text']
+
+    cur_df.to_excel(os.path.join(args.upload_dir, f"{args.experiment}.xlsx"), index=False, engine='openpyxl')

scripts/convert_mmvet_for_eval.py

@@ -0,0 +1,18 @@
+import os
+import json
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--src", type=str)
+parser.add_argument("--dst", type=str)
+args = parser.parse_args()
+
+cur_result = {}
+
+for line in open(args.src):
+    data = json.loads(line)
+    qid = data['question_id']
+    cur_result[f'v1_{qid}'] = data['text']
+
+with open(args.dst, 'w') as f:
+    json.dump(cur_result, f, indent=2)

scripts/convert_seed_for_submission.py

@@ -0,0 +1,74 @@
+import os
+import json
+import argparse
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--annotation-file", type=str)
+    parser.add_argument("--result-file", type=str)
+    parser.add_argument("--result-upload-file", type=str)
+    return parser.parse_args()
+
+
+def eval_single(result_file, eval_only_type=None):
+    results = {}
+    for line in open(result_file):
+        row = json.loads(line)
+        results[row['question_id']] = row
+
+    type_counts = {}
+    correct_counts = {}
+    for question_data in data['questions']:
+        if eval_only_type is not None and question_data['data_type'] != eval_only_type: continue
+        data_type = question_data['question_type_id']
+        type_counts[data_type] = type_counts.get(data_type, 0) + 1
+        try:
+            question_id = int(question_data['question_id'])
+        except:
+            question_id = question_data['question_id']
+        if question_id not in results:
+            correct_counts[data_type] = correct_counts.get(data_type, 0)
+            continue
+        row = results[question_id]
+        if row['text'] == question_data['answer']:
+            correct_counts[data_type] = correct_counts.get(data_type, 0) + 1
+
+    total_count = 0
+    total_correct = 0
+    for data_type in sorted(type_counts.keys()):
+        accuracy = correct_counts[data_type] / type_counts[data_type] * 100
+        if eval_only_type is None:
+            print(f"{ques_type_id_to_name[data_type]}: {accuracy:.2f}%")
+
+        total_count += type_counts[data_type]
+        total_correct += correct_counts[data_type]
+
+    total_accuracy = total_correct / total_count * 100
+    if eval_only_type is None:
+        print(f"Total accuracy: {total_accuracy:.2f}%")
+    else:
+        print(f"{eval_only_type} accuracy: {total_accuracy:.2f}%")
+
+    return results
+
+if __name__ == "__main__":
+    args = get_args()
+    data = json.load(open(args.annotation_file))
+    ques_type_id_to_name = {id:n for n,id in data['question_type'].items()}
+
+    results = eval_single(args.result_file)
+    eval_single(args.result_file, eval_only_type='image')
+    eval_single(args.result_file, eval_only_type='video')
+
+    with open(args.result_upload_file, 'w') as fp:
+        for question in data['questions']:
+            qid = question['question_id']
+            if qid in results:
+                result = results[qid]
+            else:
+                result = results[int(qid)]
+            fp.write(json.dumps({
+                'question_id': qid,
+                'prediction': result['text']
+            }) + '\n')

scripts/convert_sqa_to_llava.py

@@ -0,0 +1,88 @@
+import json
+import os
+import fire
+import re
+from convert_sqa_to_llava_base_prompt import build_prompt_chatbot
+
+
+def convert_to_llava(base_dir, split, prompt_format="QCM-LEA"):
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+
+    split_problems = build_prompt_chatbot(
+        problems, split_indices, prompt_format,
+        use_caption=False, is_test=False)
+
+    target_format = []
+    for prob_id, (input, output) in split_problems.items():
+        if input.startswith('Question: '):
+            input = input.replace('Question: ', '')
+        if output.startswith('Answer: '):
+            output = output.replace('Answer: ', '')
+
+        raw_prob_data = problems[prob_id]
+        if raw_prob_data['image'] is None:
+            target_format.append({
+                "id": prob_id,
+                "conversations": [
+                    {'from': 'human', 'value': f"{input}"},
+                    {'from': 'gpt', 'value': f"{output}"},
+                ],
+            })
+
+        else:
+            target_format.append({
+                "id": prob_id,
+                "image": os.path.join(prob_id, raw_prob_data['image']),
+                "conversations": [
+                    {'from': 'human', 'value': f"{input}\n<image>"},
+                    {'from': 'gpt', 'value': f"{output}"},
+                ],
+            })
+
+    print(f'Number of samples: {len(target_format)}')
+
+    with open(os.path.join(base_dir, f"llava_{split}_{prompt_format}.json"), "w") as f:
+        json.dump(target_format, f, indent=2)
+
+
+def convert_to_jsonl(base_dir, split, prompt_format="QCM-LEPA"):
+    split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[split]
+    problems = json.load(open(os.path.join(base_dir, "problems.json")))
+
+    split_problems = build_prompt_chatbot(
+        problems, split_indices, prompt_format,
+        use_caption=False, is_test=False)
+
+    writer = open(os.path.join(base_dir, f"scienceqa_{split}_{prompt_format}.jsonl"), "w")
+    for prob_id, (input, output) in split_problems.items():
+        if input.startswith('Question: '):
+            input = input.replace('Question: ', '')
+        if output.startswith('Answer: '):
+            output = output.replace('Answer: ', '')
+
+        raw_prob_data = problems[prob_id]
+        if raw_prob_data['image'] is None:
+            data = {
+                "id": prob_id,
+                "instruction": f"{input}",
+                "output": f"{output}",
+            }
+
+        else:
+            data = {
+                "id": prob_id,
+                "image": os.path.join(prob_id, raw_prob_data['image']),
+                "instruction": f"{input}\n<image>",
+                "output": f"{output}",
+            }
+        writer.write(json.dumps(data) + '\n')
+    writer.close()
+
+
+def main(task, **kwargs):
+    globals()[task](**kwargs)
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

scripts/convert_sqa_to_llava_base_prompt.py

@@ -0,0 +1,334 @@
+def get_question_text(problem):
+    question = problem['question']
+    return question
+
+
+def get_context_text(problem, use_caption):
+    txt_context = problem['hint']
+    img_context = problem['caption'] if use_caption else ""
+    context = " ".join([txt_context, img_context]).strip()
+    if context == "":
+        context = "N/A"
+    return context
+
+
+def get_choice_text(probelm, options):
+    choices = probelm['choices']
+    choice_list = []
+    for i, c in enumerate(choices):
+        choice_list.append("({}) {}".format(options[i], c))
+    choice_txt = " ".join(choice_list)
+    #print(choice_txt)
+    return choice_txt
+
+
+def get_answer(problem, options):
+    return options[problem['answer']]
+
+
+def get_lecture_text(problem):
+    # \\n: GPT-3 can generate the lecture with more tokens.
+    lecture = problem['lecture'].replace("\n", "\\n")
+    return lecture
+
+
+def get_solution_text(problem):
+    # \\n: GPT-3 can generate the solution with more tokens
+    solution = problem['solution'].replace("\n", "\\n")
+    return solution
+
+
+def create_one_example_chatbot(format, question, context, choice, answer, lecture, solution, test_example=True):
+
+    input_format, output_format = format.split("-")
+
+    ## Inputs
+    if input_format == "CQM":
+        input = f"Context: {context}\nQuestion: {question}\nOptions: {choice}\n"
+    elif input_format == "QCM":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\n"
+    # upper bound experiment
+    elif input_format == "QCML":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {lecture}\n"
+    elif input_format == "QCME":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {solution}\n"
+    elif input_format == "QCMLE":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {lecture} {solution}\n"
+
+    elif input_format == "QCLM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {lecture}\nOptions: {choice}\n"
+    elif input_format == "QCEM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {solution}\nOptions: {choice}\n"
+    elif input_format == "QCLEM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {lecture} {solution}\nOptions: {choice}\n"
+
+    # Outputs
+    if test_example:
+        output = "Answer:"
+    elif output_format == 'A':
+        output = f"Answer: The answer is {answer}."
+
+    elif output_format == 'AL':
+        output = f"Answer: The answer is {answer}. BECAUSE: {solution}"
+    elif output_format == 'AE':
+        output = f"Answer: The answer is {answer}. BECAUSE: {lecture}"
+    elif output_format == 'ALE':
+        output = f"Answer: The answer is {answer}. BECAUSE: {lecture} {solution}"
+    elif output_format == 'AEL':
+        output = f"Answer: The answer is {answer}. BECAUSE: {solution} {lecture}"
+
+    elif output_format == 'LA':
+        output = f"Answer: {lecture} The answer is {answer}."
+    elif output_format == 'EA':
+        output = f"Answer: {solution} The answer is {answer}."
+    elif output_format == 'LEA':
+        output = f"Answer: {lecture} {solution} The answer is {answer}."
+    elif output_format == 'ELA':
+        output = f"Answer: {solution} {lecture} The answer is {answer}."
+    elif output_format == 'LEPA':
+        output = ''
+        if len(lecture.strip()) > 0:
+            output += f"LECTURE: {lecture}\n"
+        if len(solution.strip()) > 0:
+            output += f"SOLUTION: {solution}\n"
+        output += '###\n'
+        output += f"ANSWER: {answer}."
+
+    input = input.replace("  ", " ").strip()
+    output = output.replace("  ", " ").strip()
+    if input.endswith("BECAUSE:"):
+        input = input.replace("BECAUSE:", "").strip()
+    if output.endswith("BECAUSE:"):
+        output = output.replace("BECAUSE:", "").strip()
+    return input, output
+
+
+def create_one_example(format, question, context, choice, answer, lecture, solution, test_example=True):
+
+    input_format, output_format = format.split("-")
+
+    ## Inputs
+    if input_format == "CQM":
+        input = f"Context: {context}\nQuestion: {question}\nOptions: {choice}\n"
+    elif input_format == "QCM":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\n"
+    # upper bound experiment
+    elif input_format == "QCML":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {lecture}\n"
+    elif input_format == "QCME":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {solution}\n"
+    elif input_format == "QCMLE":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {lecture} {solution}\n"
+
+    elif input_format == "QCLM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {lecture}\nOptions: {choice}\n"
+    elif input_format == "QCEM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {solution}\nOptions: {choice}\n"
+    elif input_format == "QCLEM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {lecture} {solution}\nOptions: {choice}\n"
+
+    # Outputs
+    if test_example:
+        output = "Answer:"
+    elif output_format == 'A':
+        output = f"Answer: The answer is {answer}."
+
+    elif output_format == 'AL':
+        output = f"Answer: The answer is {answer}. BECAUSE: {solution}"
+    elif output_format == 'AE':
+        output = f"Answer: The answer is {answer}. BECAUSE: {lecture}"
+    elif output_format == 'ALE':
+        output = f"Answer: The answer is {answer}. BECAUSE: {lecture} {solution}"
+    elif output_format == 'AEL':
+        output = f"Answer: The answer is {answer}. BECAUSE: {solution} {lecture}"
+
+    elif output_format == 'LA':
+        output = f"Answer: {lecture} The answer is {answer}."
+    elif output_format == 'EA':
+        output = f"Answer: {solution} The answer is {answer}."
+    elif output_format == 'LEA':
+        output = f"Answer: {lecture} {solution} The answer is {answer}."
+    elif output_format == 'ELA':
+        output = f"Answer: {solution} {lecture} The answer is {answer}."
+
+    text = input + output
+    text = text.replace("  ", " ").strip()
+    if text.endswith("BECAUSE:"):
+        text = text.replace("BECAUSE:", "").strip()
+    return text
+
+
+
+def create_one_example_gpt4(format, question, context, choice, answer, lecture, solution, test_example=True):
+
+    input_format, output_format = format.split("-")
+
+    ## Inputs
+    if input_format == "CQM":
+        input = f"Context: {context}\nQuestion: {question}\nOptions: {choice}\n"
+    elif input_format == "QCM":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\n"
+    # upper bound experiment
+    elif input_format == "QCML":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {lecture}\n"
+    elif input_format == "QCME":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {solution}\n"
+    elif input_format == "QCMLE":
+        input = f"Question: {question}\nContext: {context}\nOptions: {choice}\nBECAUSE: {lecture} {solution}\n"
+
+    elif input_format == "QCLM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {lecture}\nOptions: {choice}\n"
+    elif input_format == "QCEM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {solution}\nOptions: {choice}\n"
+    elif input_format == "QCLEM":
+        input = f"Question: {question}\nContext: {context}\nBECAUSE: {lecture} {solution}\nOptions: {choice}\n"
+
+    # Outputs
+    if test_example:
+        output = "Answer:"
+    elif output_format == 'A':
+        output = f"Answer: The answer is {answer}."
+
+    elif output_format == 'AL':
+        output = f"Answer: The answer is {answer}. BECAUSE: {solution}"
+    elif output_format == 'AE':
+        output = f"Answer: The answer is {answer}. BECAUSE: {lecture}"
+    elif output_format == 'ALE':
+        output = f"Answer: The answer is {answer}. BECAUSE: {lecture} {solution}"
+    elif output_format == 'AEL':
+        output = f"Answer: The answer is {answer}. BECAUSE: {solution} {lecture}"
+
+    elif output_format == 'LA':
+        output = f"Answer: {lecture} The answer is {answer}."
+    elif output_format == 'EA':
+        output = f"Answer: {solution} The answer is {answer}."
+    elif output_format == 'LEA':
+        output = f"Answer: {lecture} {solution} The answer is {answer}."
+    elif output_format == 'ELA':
+        output = f"Answer: {solution} {lecture} The answer is {answer}."
+
+    input = input.replace("  ", " ").strip()
+    output = output.replace("  ", " ").strip()
+    if output.endswith("BECAUSE:"):
+        output = output.replace("BECAUSE:", "").strip()
+
+    user_prompt = {"role": "user", "content": f"Can you explain {input}?"}
+    assistant_prompt = {"role": "assistant", "content": f"{output}"}
+
+    return user_prompt, assistant_prompt
+
+
+def build_prompt_chatbot(problems, shot_qids, prompt_format, use_caption=False, options=["A", "B", "C", "D", "E"], is_test=False):
+    examples = {}
+
+    for qid in shot_qids:
+        question = get_question_text(problems[qid])
+        context = get_context_text(problems[qid], use_caption)
+        choice = get_choice_text(problems[qid], options)
+        answer = get_answer(problems[qid], options)
+        lecture = get_lecture_text(problems[qid]).replace('\\n', '\n')
+        solution = get_solution_text(problems[qid]).replace('\\n', '\n')
+
+        train_example = create_one_example_chatbot(prompt_format,
+                                           question,
+                                           context,
+                                           choice,
+                                           answer,
+                                           lecture,
+                                           solution,
+                                           test_example=is_test)
+        examples[qid] = train_example
+    return examples
+
+
+def build_prompt(problems, shot_qids, test_qid, args):
+
+    examples = []
+
+    # n-shot training examples
+    for qid in shot_qids:
+        question = get_question_text(problems[qid])
+        context = get_context_text(problems[qid], args.use_caption)
+        choice = get_choice_text(problems[qid], args.options)
+        answer = get_answer(problems[qid], args.options)
+        lecture = get_lecture_text(problems[qid])
+        solution = get_solution_text(problems[qid])
+
+        train_example = create_one_example(args.prompt_format,
+                                           question,
+                                           context,
+                                           choice,
+                                           answer,
+                                           lecture,
+                                           solution,
+                                           test_example=False)
+        examples.append(train_example)
+
+    # test example
+    question = get_question_text(problems[test_qid])
+    context = get_context_text(problems[test_qid], args.use_caption)
+    choice = get_choice_text(problems[test_qid], args.options)
+    answer = get_answer(problems[test_qid], args.options)
+    lecture = get_lecture_text(problems[test_qid])
+    solution = get_solution_text(problems[test_qid])
+
+    test_example = create_one_example(args.prompt_format,
+                                      question,
+                                      context,
+                                      choice,
+                                      answer,
+                                      lecture,
+                                      solution,
+                                      test_example=True)
+    examples.append(test_example)
+
+    # create the prompt input
+    prompt_input = '\n\n'.join(examples)
+
+    return prompt_input
+
+
+def build_prompt_gpt4(problems, shot_qids, test_qid, args):
+
+    prompt_array = [{"role": "system", "content": "You are a helpful assistant."}]
+
+    # n-shot training examples
+    for qid in shot_qids:
+        question = get_question_text(problems[qid])
+        context = get_context_text(problems[qid], args.use_caption)
+        choice = get_choice_text(problems[qid], args.options)
+        answer = get_answer(problems[qid], args.options)
+        lecture = get_lecture_text(problems[qid])
+        solution = get_solution_text(problems[qid])
+
+        user_prompt, assistant_prompt = create_one_example_gpt4(args.prompt_format,
+                                           question,
+                                           context,
+                                           choice,
+                                           answer,
+                                           lecture,
+                                           solution,
+                                           test_example=False)
+        prompt_array.append(user_prompt)
+        prompt_array.append(assistant_prompt)
+
+    # test example
+    question = get_question_text(problems[test_qid])
+    context = get_context_text(problems[test_qid], args.use_caption)
+    choice = get_choice_text(problems[test_qid], args.options)
+    answer = get_answer(problems[test_qid], args.options)
+    lecture = get_lecture_text(problems[test_qid])
+    solution = get_solution_text(problems[test_qid])
+
+    user_prompt, assistant_prompt = create_one_example_gpt4(args.prompt_format,
+                                      question,
+                                      context,
+                                      choice,
+                                      answer,
+                                      lecture,
+                                      solution,
+                                      test_example=True)
+    prompt_array.append(user_prompt)
+    prompt_array.append(assistant_prompt)
+
+    return prompt_array

scripts/convert_vizwiz_for_submission.py

@@ -0,0 +1,47 @@
+import os
+import argparse
+import json
+
+from llava.eval.m4c_evaluator import EvalAIAnswerProcessor
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--annotation-file', type=str, required=True)
+    parser.add_argument('--result-file', type=str, required=True)
+    parser.add_argument('--result-upload-file', type=str, required=True)
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+
+    args = parse_args()
+
+    os.makedirs(os.path.dirname(args.result_upload_file), exist_ok=True)
+
+    results = []
+    error_line = 0
+    for line_idx, line in enumerate(open(args.result_file)):
+        try:
+            results.append(json.loads(line))
+        except:
+            error_line += 1
+    results = {x['question_id']: x['text'] for x in results}
+    test_split = [json.loads(line) for line in open(args.annotation_file)]
+    split_ids = set([x['question_id'] for x in test_split])
+
+    print(f'total results: {len(results)}, total split: {len(test_split)}, error_line: {error_line}')
+
+    all_answers = []
+
+    answer_processor = EvalAIAnswerProcessor()
+
+    for x in test_split:
+        assert x['question_id'] in results
+        all_answers.append({
+            'image': x['image'],
+            'answer': answer_processor(results[x['question_id']])
+        })
+
+    with open(args.result_upload_file, 'w') as f:
+        json.dump(all_answers, f)

scripts/convert_vqav2_for_submission.py

@@ -0,0 +1,56 @@
+import os
+import argparse
+import json
+
+from llava.eval.m4c_evaluator import EvalAIAnswerProcessor
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--dir', type=str, default="./playground/data/eval/vqav2")
+    parser.add_argument('--ckpt', type=str, required=True)
+    parser.add_argument('--split', type=str, required=True)
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+
+    args = parse_args()
+
+    src = os.path.join(args.dir, 'answers', args.split, args.ckpt, 'merge.jsonl')
+    test_split = os.path.join(args.dir, 'llava_vqav2_mscoco_test2015.jsonl')
+    dst = os.path.join(args.dir, 'answers_upload', args.split, f'{args.ckpt}.json')
+    os.makedirs(os.path.dirname(dst), exist_ok=True)
+
+    results = []
+    error_line = 0
+    for line_idx, line in enumerate(open(src)):
+        try:
+            results.append(json.loads(line))
+        except:
+            error_line += 1
+
+    results = {x['question_id']: x['text'] for x in results}
+    test_split = [json.loads(line) for line in open(test_split)]
+    split_ids = set([x['question_id'] for x in test_split])
+
+    print(f'total results: {len(results)}, total split: {len(test_split)}, error_line: {error_line}')
+
+    all_answers = []
+
+    answer_processor = EvalAIAnswerProcessor()
+
+    for x in test_split:
+        if x['question_id'] not in results:
+            all_answers.append({
+                'question_id': x['question_id'],
+                'answer': ''
+            })
+        else:
+            all_answers.append({
+                'question_id': x['question_id'],
+                'answer': answer_processor(results[x['question_id']])
+            })
+
+    with open(dst, 'w') as f:
+        json.dump(all_answers, open(dst, 'w'))

scripts/eval_gpt_mmvet.py

@@ -0,0 +1,276 @@
+import argparse
+
+import openai
+import json
+import os
+from tqdm import tqdm
+import pandas as pd
+import numpy as np
+from collections import Counter
+import time
+
+
+
+parser = argparse.ArgumentParser(description='ChatGPT-based QA evaluation.')
+parser.add_argument('--mmvet_path')
+parser.add_argument('--ckpt_name')
+parser.add_argument('--result_path')
+args = parser.parse_args()
+
+
+openai.api_base = "https://api.aiguoguo199.com/v1"
+openai.api_key = 'sk-eionFWpNThMNy4eeFdC25789F60a4cC2A66b2c94D3948bA6'
+
+gpt_model = "gpt-3.5-turbo"
+
+
+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
+"""
+
+# load metadata
+# Download mm-vet.zip and `unzip mm-vet.zip` and change the path below
+mmvet_path = args.mmvet_path
+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 = args.ckpt_name
+result_path = args.result_path
+num_run = 1 # we set it as 5 in the paper
+model_results_file = os.path.join(result_path, f"{model}.json")
+
+# 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 tqdm(data.items()):
+            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
+
+            while not grade_sample_run_complete:
+                try:
+                    response = openai.ChatCompletion.create(
+                        model=gpt_model,
+                        max_tokens=3,
+                        temperature=temperature,
+                        messages=messages)
+                    # print(response['model'])
+                    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,
+                                max_tokens=3,
+                                temperature=temperature,
+                                messages=messages)
+                            # print(response)
+                            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
+                    print("sleep 1s")
+                    time.sleep(1)
+
+            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)
+print(df)
+print(df2)

scripts/finetune.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+# IMPORTANT: this is the training script for the original LLaVA, NOT FOR LLaVA V1.5!
+
+# Uncomment and set the following variables correspondingly to run this script:
+
+################## VICUNA ##################
+# PROMPT_VERSION=v1
+# MODEL_VERSION="vicuna-v1-3-7b"
+################## VICUNA ##################
+
+################## LLaMA-2 ##################
+# PROMPT_VERSION="llava_llama_2"
+# MODEL_VERSION="llama-2-7b-chat"
+################## LLaMA-2 ##################
+
+deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path ./checkpoints/$MODEL_VERSION \
+    --version $PROMPT_VERSION \
+    --data_path ./playground/data/llava_instruct_80k.json \
+    --image_folder /path/to/coco/train2017 \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --pretrain_mm_mlp_adapter ./checkpoints/llava-$MODEL_VERSION-pretrain/mm_projector.bin \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/llava-$MODEL_VERSION-finetune \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 50000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb

scripts/finetune_full_schedule.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+# IMPORTANT: this is the training script for the original LLaVA, NOT FOR LLaVA V1.5!
+
+# Uncomment and set the following variables correspondingly to run this script:
+
+################## VICUNA ##################
+# PROMPT_VERSION=v1
+# MODEL_VERSION="vicuna-v1-3-7b"
+################## VICUNA ##################
+
+################## LLaMA-2 ##################
+# PROMPT_VERSION="llava_llama_2"
+# MODEL_VERSION="llama-2-7b-chat"
+################## LLaMA-2 ##################
+
+deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path ./checkpoints/$MODEL_VERSION \
+    --version $PROMPT_VERSION \
+    --data_path ./playground/data/llava_instruct_158k.json \
+    --image_folder /path/to/coco/train2017 \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --pretrain_mm_mlp_adapter ./checkpoints/llava-$MODEL_VERSION-pretrain/mm_projector.bin \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/llava-$MODEL_VERSION-finetune \
+    --num_train_epochs 3 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 50000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb

scripts/finetune_lora.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+
+# IMPORTANT: this is the training script for the original LLaVA, NOT FOR LLaVA V1.5!
+
+# Uncomment and set the following variables correspondingly to run this script:
+
+################## VICUNA ##################
+# PROMPT_VERSION=v1
+# MODEL_VERSION="vicuna-v1-3-7b"
+################## VICUNA ##################
+
+################## LLaMA-2 ##################
+# PROMPT_VERSION="llava_llama_2"
+# MODEL_VERSION="llama-2-7b-chat"
+################## LLaMA-2 ##################
+
+deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --lora_enable True \
+    --model_name_or_path ./checkpoints/$MODEL_VERSION \
+    --version $PROMPT_VERSION \
+    --data_path ./playground/data/llava_instruct_80k.json \
+    --image_folder /path/to/coco/train2017 \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --pretrain_mm_mlp_adapter ./checkpoints/llava-$MODEL_VERSION-pretrain/mm_projector.bin \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/llava-$MODEL_VERSION-finetune_lora \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 50000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --lazy_preprocess True \
+    --dataloader_num_workers 4 \
+    --report_to wandb

scripts/finetune_qlora.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+# IMPORTANT: this is the training script for the original LLaVA, NOT FOR LLaVA V1.5!
+
+# Uncomment and set the following variables correspondingly to run this script:
+
+################## VICUNA ##################
+# PROMPT_VERSION=v1
+# MODEL_VERSION="vicuna-v1-3-7b"
+################## VICUNA ##################
+
+################## LLaMA-2 ##################
+# PROMPT_VERSION="llava_llama_2"
+# MODEL_VERSION="llama-2-7b-chat"
+################## LLaMA-2 ##################
+
+deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --lora_enable True \
+    --bits 4 \
+    --model_name_or_path ./checkpoints/$MODEL_VERSION \
+    --version $PROMPT_VERSION \
+    --data_path ./playground/data/llava_instruct_80k.json \
+    --image_folder /path/to/coco/train2017 \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --pretrain_mm_mlp_adapter ./checkpoints/llava-$MODEL_VERSION-pretrain/mm_projector.bin \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/llava-$MODEL_VERSION-finetune_lora \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 50000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --lazy_preprocess True \
+    --dataloader_num_workers 4 \
+    --report_to wandb

scripts/finetune_sqa.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+
+# IMPORTANT: this is the training script for the original LLaVA, NOT FOR LLaVA V1.5!
+
+deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path lmsys/vicuna-13b-v1.3 \
+    --version $PROMPT_VERSION \
+    --data_path /Data/ScienceQA/data/scienceqa/llava_train_QCM-LEA.json \
+    --image_folder /Data/ScienceQA/data/scienceqa/images/train \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --pretrain_mm_mlp_adapter ./checkpoints/huggingface/liuhaotian/llava-pretrain-vicuna-13b-v1.3/mm_projector.bin \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/llava-vicuna-13b-v1.3-pretrain_lcs558k_plain-ScienceQA_QCM_LEA-12e \
+    --num_train_epochs 12 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 50000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb

scripts/merge_lora_weights.py

@@ -0,0 +1,22 @@
+import argparse
+from llava.model.builder import load_pretrained_model
+from llava.mm_utils import get_model_name_from_path
+
+
+def merge_lora(args):
+    model_name = get_model_name_from_path(args.model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.model_base, model_name, device_map='cpu')
+
+    model.save_pretrained(args.save_model_path)
+    tokenizer.save_pretrained(args.save_model_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-path", type=str, required=True)
+    parser.add_argument("--model-base", type=str, required=True)
+    parser.add_argument("--save-model-path", type=str, required=True)
+
+    args = parser.parse_args()
+
+    merge_lora(args)

scripts/pretrain.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+
+# IMPORTANT: this is the training script for the original LLaVA, NOT FOR LLaVA V1.5!
+
+# Uncomment and set the following variables correspondingly to run this script:
+
+# MODEL_VERSION=vicuna-v1-3-7b
+# MODEL_VERSION=llama-2-7b-chat
+
+########### DO NOT CHANGE ###########
+########### USE THIS FOR BOTH ###########
+PROMPT_VERSION=plain
+########### DO NOT CHANGE ###########
+
+deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path ./checkpoints/$MODEL_VERSION \
+    --version $PROMPT_VERSION \
+    --data_path /path/to/pretrain_data.json \
+    --image_folder /path/to/images \
+    --vision_tower openai/clip-vit-large-patch14 \
+    --tune_mm_mlp_adapter True \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/llava-$MODEL_VERSION-pretrain \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 24000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-3 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb

scripts/sqa_eval_batch.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+
+CHUNKS=8
+for IDX in {0..7}; do
+    CUDA_VISIBLE_DEVICES=$IDX python -m llava.eval.model_vqa_science \
+        --model-path liuhaotian/llava-lcs558k-scienceqa-vicuna-13b-v1.3 \
+        --question-file ~/haotian/datasets/ScienceQA/data/scienceqa/llava_test_QCM-LEA.json \
+        --image-folder ~/haotian/datasets/ScienceQA/data/scienceqa/images/test \
+        --answers-file ./test_llava-13b-chunk$CHUNKS_$IDX.jsonl \
+        --num-chunks $CHUNKS \
+        --chunk-idx $IDX \
+        --conv-mode llava_v1 &
+done

scripts/sqa_eval_gather.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+CHUNKS=8
+output_file="test_llava-13b.jsonl"
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for idx in $(seq 0 $((CHUNKS-1))); do
+  cat "./test_llava-13b-chunk${idx}.jsonl" >> "$output_file"
+done
+
+python llava/eval/eval_science_qa.py \
+    --base-dir ~/haotian/datasets/ScienceQA/data/scienceqa \
+    --result-file ./test_llava-13b.jsonl \
+    --output-file ./test_llava-13b_output.json \
+    --output-result ./test_llava-13b_result.json

scripts/v1_5/eval/eval_benchmark_1_correctness.sh

@@ -0,0 +1,17 @@
+
+CKPT_NAME="checkpoints/Video-LLaVA-7B"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+pred_path="${Video_5_Benchmark}/${CKPT_NAME}/correctness_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}/gpt3/correctness"
+output_json="${Video_5_Benchmark}/${CKPT_NAME}/results/correctness_qa.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+python3 llava/eval/video/eval_benchmark_1_correctness.py \
+    --pred_path  ${pred_path} \
+    --output_dir  ${output_dir} \
+    --output_json  ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_benchmark_2_detail.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="checkpoints/Video-LLaVA-7B"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+pred_path="${Video_5_Benchmark}/${CKPT_NAME}/detail_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}/gpt3/detail"
+output_json="${Video_5_Benchmark}/${CKPT_NAME}/results/detail_qa.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+python3 llava/eval/video/eval_benchmark_2_detailed_orientation.py \
+    --pred_path  ${pred_path} \
+    --output_dir  ${output_dir} \
+    --output_json  ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_benchmark_3_contextual.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="checkpoints/Video-LLaVA-7B"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+pred_path="${Video_5_Benchmark}/${CKPT_NAME}/contextual_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}/gpt3/contextual"
+output_json="${Video_5_Benchmark}/${CKPT_NAME}/results/contextual_qa.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+python3 llava/eval/video/eval_benchmark_3_context.py \
+    --pred_path  ${pred_path} \
+    --output_dir  ${output_dir} \
+    --output_json  ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_benchmark_4_temporal.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="checkpoints/Video-LLaVA-7B"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+pred_path="${Video_5_Benchmark}/${CKPT_NAME}/temporal_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}/gpt3/temporal"
+output_json="${Video_5_Benchmark}/${CKPT_NAME}/results/temporal_qa.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+python3 llava/eval/video/eval_benchmark_4_temporal.py \
+    --pred_path  ${pred_path} \
+    --output_dir  ${output_dir} \
+    --output_json  ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_benchmark_5_consistency.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="checkpoints/Video-LLaVA-7B"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+pred_path="${Video_5_Benchmark}/${CKPT_NAME}/consistency_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}/gpt3/consistency"
+output_json="${Video_5_Benchmark}/${CKPT_NAME}/results/consistency_qa.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+python3 llava/eval/video/eval_benchmark_5_consistency.py \
+    --pred_path  ${pred_path} \
+    --output_dir  ${output_dir} \
+    --output_json  ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_image_gqa.sh

@@ -0,0 +1,43 @@
+#!/bin/bash
+
+gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
+IFS=',' read -ra GPULIST <<< "$gpu_list"
+
+CHUNKS=${#GPULIST[@]}
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+SPLIT="llava_gqa_testdev_balanced"
+EVAL="eval"
+GQADIR="${EVAL}/gqa/data"
+
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python3 -m llava.eval.model_vqa_loader \
+        --model-path ${CKPT} \
+        --question-file ${EVAL}/gqa/$SPLIT.jsonl \
+        --image-folder ${EVAL}/gqa/data/images \
+        --answers-file ${EVAL}/gqa/answers/$SPLIT/${CKPT_NAME}/${CHUNKS}_${IDX}.jsonl \
+        --num-chunks $CHUNKS \
+        --chunk-idx $IDX \
+        --temperature 0 \
+        --conv-mode vicuna_v1 &
+done
+
+wait
+
+output_file=${EVAL}/gqa/answers/$SPLIT/${CKPT_NAME}/merge.jsonl
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    cat ${EVAL}/gqa/answers/$SPLIT/${CKPT_NAME}/${CHUNKS}_${IDX}.jsonl >> "$output_file"
+done
+
+mkdir -p $GQADIR/$SPLIT/${CKPT_NAME}
+python3 scripts/convert_gqa_for_eval.py --src $output_file --dst $GQADIR/$SPLIT/${CKPT_NAME}/testdev_balanced_predictions.json
+
+cd $GQADIR
+python3 eval/eval_gqa.py --tier $SPLIT/${CKPT_NAME}/testdev_balanced \
+                         --questions /scc_cephfs/yy/lb/LLaVA-Video-YY/questions1.2/testdev_balanced_questions.json

scripts/v1_5/eval/eval_image_llavabench.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/llava-bench-in-the-wild/questions.jsonl \
+    --image-folder ${EVAL}/llava-bench-in-the-wild/images \
+    --answers-file ${EVAL}/llava-bench-in-the-wild/answers/${CKPT_NAME}.jsonl \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+mkdir -p ${EVAL}/llava-bench-in-the-wild/reviews
+
+python3 llava/eval/eval_gpt_review_bench.py \
+    --question ${EVAL}/llava-bench-in-the-wild/questions.jsonl \
+    --context ${EVAL}/llava-bench-in-the-wild/context.jsonl \
+    --rule llava/eval/table/rule.json \
+    --answer-list ${EVAL}/llava-bench-in-the-wild/answers_gpt4.jsonl \
+                  ${EVAL}/llava-bench-in-the-wild/answers/${CKPT_NAME}.jsonl \
+    --output ${EVAL}/llava-bench-in-the-wild/reviews/${CKPT_NAME}.jsonl
+
+python3 llava/eval/summarize_gpt_review.py -f ${EVAL}/llava-bench-in-the-wild/reviews/${CKPT_NAME}.jsonl

scripts/v1_5/eval/eval_image_mmbench.sh

@@ -0,0 +1,22 @@
+#!/bin/bash
+
+SPLIT="mmbench_dev_20230712"
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa_mmbench \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/mmbench/$SPLIT.tsv \
+    --answers-file ${EVAL}/mmbench/answers/$SPLIT/${CKPT_NAME}.jsonl \
+    --single-pred-prompt \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+mkdir -p ${EVAL}/mmbench/answers_upload/$SPLIT
+
+python3 scripts/convert_mmbench_for_submission.py \
+    --annotation-file ${EVAL}/mmbench/$SPLIT.tsv \
+    --result-dir ${EVAL}/mmbench/answers/$SPLIT \
+    --upload-dir ${EVAL}/mmbench/answers_upload/$SPLIT \
+    --experiment ${CKPT_NAME}

scripts/v1_5/eval/eval_image_mmvet.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/mm-vet/llava-mm-vet.jsonl \
+    --image-folder ${EVAL}/mm-vet/images \
+    --answers-file ${EVAL}/mm-vet/answers/${CKPT_NAME}.jsonl \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+mkdir -p ${EVAL}/mm-vet/results
+
+python3 scripts/convert_mmvet_for_eval.py \
+    --src ${EVAL}/mm-vet/answers/${CKPT_NAME}.jsonl \
+    --dst ${EVAL}/mm-vet/results/${CKPT_NAME}.json
+
+
+python3 scripts/eval_gpt_mmvet.py \
+    --mmvet_path ${EVAL}/mm-vet \
+    --ckpt_name ${CKPT_NAME} \
+    --result_path ${EVAL}/mm-vet/results

scripts/v1_5/eval/eval_image_pope.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa_loader \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/pope/llava_pope_test.jsonl \
+    --image-folder ${EVAL}/pope/val2014 \
+    --answers-file ${EVAL}/pope/answers/${CKPT_NAME}.jsonl \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+python3 llava/eval/eval_pope.py \
+    --annotation-dir ${EVAL}/pope/coco \
+    --question-file ${EVAL}/pope/llava_pope_test.jsonl \
+    --result-file ${EVAL}/pope/answers/${CKPT_NAME}.jsonl

scripts/v1_5/eval/eval_image_sqa.sh

@@ -0,0 +1,20 @@
+#!/bin/bash
+
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa_science \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/scienceqa/llava_test_CQM-A.json \
+    --image-folder ${EVAL}/scienceqa/images/test \
+    --answers-file ${EVAL}/scienceqa/answers/${CKPT_NAME}.jsonl \
+    --single-pred-prompt \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+python3 llava/eval/eval_science_qa.py \
+    --base-dir ${EVAL}/scienceqa \
+    --result-file ${EVAL}/scienceqa/answers/${CKPT_NAME}.jsonl \
+    --output-file ${EVAL}/scienceqa/answers/${CKPT_NAME}_output.jsonl \
+    --output-result ${EVAL}/scienceqa/answers/${CKPT_NAME}_result.json

scripts/v1_5/eval/eval_image_textvqa.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa_loader \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/textvqa/llava_textvqa_val_v051_ocr.jsonl \
+    --image-folder ${EVAL}/textvqa/train_images \
+    --answers-file ${EVAL}/textvqa/answers/${CKPT_NAME}.jsonl \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+python3 -m llava.eval.eval_textvqa \
+    --annotation-file ${EVAL}/textvqa/TextVQA_0.5.1_val.json \
+    --result-file ${EVAL}/textvqa/answers/${CKPT_NAME}.jsonl

scripts/v1_5/eval/eval_image_vizwiz.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+EVAL="eval"
+python3 -m llava.eval.model_vqa_loader \
+    --model-path ${CKPT} \
+    --question-file ${EVAL}/vizwiz/llava_test.jsonl \
+    --image-folder ${EVAL}/vizwiz/test \
+    --answers-file ${EVAL}/vizwiz/answers/${CKPT_NAME}.jsonl \
+    --temperature 0 \
+    --conv-mode vicuna_v1
+
+python3 scripts/convert_vizwiz_for_submission.py \
+    --annotation-file ${EVAL}/vizwiz/llava_test.jsonl \
+    --result-file ${EVAL}/vizwiz/answers/${CKPT_NAME}.jsonl \
+    --result-upload-file ${EVAL}/vizwiz/answers_upload/${CKPT_NAME}.json

scripts/v1_5/eval/eval_image_vqav2.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
+IFS=',' read -ra GPULIST <<< "$gpu_list"
+
+CHUNKS=${#GPULIST[@]}
+
+CKPT_NAME="Video-LLaVA-7B"
+CKPT="checkpoints/${CKPT_NAME}"
+SPLIT="llava_vqav2_mscoco_test-dev2015"
+EVAL="eval"
+
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python3 -m llava.eval.model_vqa_loader \
+        --model-path ${CKPT} \
+        --question-file ${EVAL}/vqav2/$SPLIT.jsonl \
+        --image-folder ${EVAL}/vqav2/test2015 \
+        --answers-file ${EVAL}/vqav2/answers/$SPLIT/${CKPT_NAME}/${CHUNKS}_${IDX}.jsonl \
+        --num-chunks $CHUNKS \
+        --chunk-idx $IDX \
+        --temperature 0 \
+        --conv-mode vicuna_v1 &
+done
+
+wait
+
+output_file=${EVAL}/vqav2/answers/$SPLIT/${CKPT_NAME}/merge.jsonl
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    cat ${EVAL}/vqav2/answers/$SPLIT/${CKPT_NAME}/${CHUNKS}_${IDX}.jsonl >> "$output_file"
+done
+
+python3 scripts/convert_vqav2_for_submission.py --split $SPLIT --ckpt ${CKPT_NAME} --dir ${EVAL}/vqav2
+

scripts/v1_5/eval/eval_qa_activitynet.sh

@@ -0,0 +1,20 @@
+
+
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+output_name="Video-LLaVA-7B"
+pred_path="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/${output_name}/merge.jsonl"
+output_dir="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/${output_name}/gpt3-0.25"
+output_json="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/${output_name}/results.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+
+
+python3 llava/eval/video/eval_video_qa.py \
+    --pred_path ${pred_path} \
+    --output_dir ${output_dir} \
+    --output_json ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_qa_msrvtt.sh

@@ -0,0 +1,20 @@
+
+
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+output_name="Video-LLaVA-7B"
+pred_path="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/${output_name}/merge.jsonl"
+output_dir="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/${output_name}/gpt"
+output_json="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/${output_name}/results.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+
+
+python3 llava/eval/video/eval_video_qa.py \
+    --pred_path ${pred_path} \
+    --output_dir ${output_dir} \
+    --output_json ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_qa_msvd.sh

@@ -0,0 +1,20 @@
+
+
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+output_name="Video-LLaVA-7B"
+pred_path="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/${output_name}/merge.jsonl"
+output_dir="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/${output_name}/gpt3.5"
+output_json="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/${output_name}/results.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+
+
+python3 llava/eval/video/eval_video_qa.py \
+    --pred_path ${pred_path} \
+    --output_dir ${output_dir} \
+    --output_json ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/eval_qa_tgif.sh

@@ -0,0 +1,20 @@
+
+
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+output_name="Video-LLaVA-7B"
+pred_path="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/${output_name}/merge.jsonl"
+output_dir="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/${output_name}/gpt3.5-0.0"
+output_json="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/${output_name}/results.json"
+api_key=""
+api_base=""
+num_tasks=8
+
+
+
+python3 llava/eval/video/eval_video_qa.py \
+    --pred_path ${pred_path} \
+    --output_dir ${output_dir} \
+    --output_json ${output_json} \
+    --api_key ${api_key} \
+    --api_base ${api_base} \
+    --num_tasks ${num_tasks}

scripts/v1_5/eval/run_benchmark_1_correctness.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+video_dir="${Video_5_Benchmark}/Test_Videos"
+gt_file="${Video_5_Benchmark}/Benchmarking_QA/generic_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}"
+output_name="correctness_qa"
+
+python3 llava/eval/video/run_inference_benchmark_general.py \
+    --model_path ${model_path} \
+    --cache_dir ${cache_dir} \
+    --video_dir ${video_dir} \
+    --gt_file ${gt_file} \
+    --output_dir ${output_dir} \
+    --output_name ${output_name}

scripts/v1_5/eval/run_benchmark_2_detail.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+video_dir="${Video_5_Benchmark}/Test_Videos"
+gt_file="${Video_5_Benchmark}/Benchmarking_QA/generic_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}"
+output_name="detail_qa"
+
+python3 llava/eval/video/run_inference_benchmark_general.py \
+    --model_path ${model_path} \
+    --cache_dir ${cache_dir} \
+    --video_dir ${video_dir} \
+    --gt_file ${gt_file} \
+    --output_dir ${output_dir} \
+    --output_name ${output_name}

scripts/v1_5/eval/run_benchmark_3_contextual.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+video_dir="${Video_5_Benchmark}/Test_Videos"
+gt_file="${Video_5_Benchmark}/Benchmarking_QA/generic_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}"
+output_name="contextual_qa"
+
+python3 llava/eval/video/run_inference_benchmark_general.py \
+    --model_path ${model_path} \
+    --cache_dir ${cache_dir} \
+    --video_dir ${video_dir} \
+    --gt_file ${gt_file} \
+    --output_dir ${output_dir} \
+    --output_name ${output_name}

scripts/v1_5/eval/run_benchmark_4_temporal.sh

@@ -0,0 +1,17 @@
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+video_dir="${Video_5_Benchmark}/Test_Videos"
+gt_file="${Video_5_Benchmark}/Benchmarking_QA/temporal_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}"
+output_name="temporal_qa"
+
+python3 llava/eval/video/run_inference_benchmark_general.py \
+    --model_path ${model_path} \
+    --cache_dir ${cache_dir} \
+    --video_dir ${video_dir} \
+    --gt_file ${gt_file} \
+    --output_dir ${output_dir} \
+    --output_name ${output_name}

scripts/v1_5/eval/run_benchmark_5_consistency.sh

@@ -0,0 +1,18 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+Video_5_Benchmark="eval/Video_5_Benchmark"
+video_dir="${Video_5_Benchmark}/Test_Videos"
+gt_file="${Video_5_Benchmark}/Benchmarking_QA/consistency_qa.json"
+output_dir="${Video_5_Benchmark}/${CKPT_NAME}"
+output_name="consistency_qa"
+
+python3 llava/eval/video/run_inference_benchmark_consistency.py \
+    --model_path ${model_path} \
+    --cache_dir ${cache_dir} \
+    --video_dir ${video_dir} \
+    --gt_file ${gt_file} \
+    --output_dir ${output_dir} \
+    --output_name ${output_name}

scripts/v1_5/eval/run_qa_activitynet.sh

@@ -0,0 +1,42 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+video_dir="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/all_test"
+gt_file_question="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/test_q.json"
+gt_file_answers="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/test_a.json"
+output_dir="${GPT_Zero_Shot_QA}/Activitynet_Zero_Shot_QA/${CKPT_NAME}"
+
+
+gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
+IFS=',' read -ra GPULIST <<< "$gpu_list"
+
+CHUNKS=${#GPULIST[@]}
+
+
+for IDX in $(seq 0 $((CHUNKS-1))); do
+  CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python3 llava/eval/video/run_inference_video_qa_act.py \
+      --model_path ${model_path} \
+      --cache_dir ${cache_dir} \
+      --video_dir ${video_dir} \
+      --gt_file_question ${gt_file_question} \
+      --gt_file_answers ${gt_file_answers} \
+      --output_dir ${output_dir} \
+      --output_name ${CHUNKS}_${IDX} \
+      --num_chunks $CHUNKS \
+      --chunk_idx $IDX &
+done
+
+wait
+
+output_file=${output_dir}/merge.jsonl
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    cat ${output_dir}/${CHUNKS}_${IDX}.json >> "$output_file"
+done

scripts/v1_5/eval/run_qa_msrvtt.sh

@@ -0,0 +1,43 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+video_dir="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/videos/all"
+gt_file_question="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/test_q.json"
+gt_file_answers="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/test_a.json"
+output_dir="${GPT_Zero_Shot_QA}/MSRVTT_Zero_Shot_QA/${CKPT_NAME}"
+
+
+
+gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
+IFS=',' read -ra GPULIST <<< "$gpu_list"
+
+CHUNKS=${#GPULIST[@]}
+
+
+for IDX in $(seq 0 $((CHUNKS-1))); do
+  CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python3 llava/eval/video/run_inference_video_qa.py \
+      --model_path ${model_path} \
+      --cache_dir ${cache_dir} \
+      --video_dir ${video_dir} \
+      --gt_file_question ${gt_file_question} \
+      --gt_file_answers ${gt_file_answers} \
+      --output_dir ${output_dir} \
+      --output_name ${CHUNKS}_${IDX} \
+      --num_chunks $CHUNKS \
+      --chunk_idx $IDX &
+done
+
+wait
+
+output_file=${output_dir}/merge.jsonl
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    cat ${output_dir}/${CHUNKS}_${IDX}.json >> "$output_file"
+done

scripts/v1_5/eval/run_qa_msvd.sh

@@ -0,0 +1,42 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+video_dir="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/videos"
+gt_file_question="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/test_q.json"
+gt_file_answers="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/test_a.json"
+output_dir="${GPT_Zero_Shot_QA}/MSVD_Zero_Shot_QA/${CKPT_NAME}"
+
+
+gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
+IFS=',' read -ra GPULIST <<< "$gpu_list"
+
+CHUNKS=${#GPULIST[@]}
+
+
+for IDX in $(seq 0 $((CHUNKS-1))); do
+  CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python3 llava/eval/video/run_inference_video_qa.py \
+      --model_path ${model_path} \
+      --cache_dir ${cache_dir} \
+      --video_dir ${video_dir} \
+      --gt_file_question ${gt_file_question} \
+      --gt_file_answers ${gt_file_answers} \
+      --output_dir ${output_dir} \
+      --output_name ${CHUNKS}_${IDX} \
+      --num_chunks $CHUNKS \
+      --chunk_idx $IDX &
+done
+
+wait
+
+output_file=${output_dir}/merge.jsonl
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    cat ${output_dir}/${CHUNKS}_${IDX}.json >> "$output_file"
+done

scripts/v1_5/eval/run_qa_tgif.sh

@@ -0,0 +1,42 @@
+
+
+CKPT_NAME="Video-LLaVA-7B"
+model_path="checkpoints/${CKPT_NAME}"
+cache_dir="./cache_dir"
+GPT_Zero_Shot_QA="eval/GPT_Zero_Shot_QA"
+video_dir="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/mp4"
+gt_file_question="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/test_q.json"
+gt_file_answers="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/test_a.json"
+output_dir="${GPT_Zero_Shot_QA}/TGIF_Zero_Shot_QA/${CKPT_NAME}"
+
+
+gpu_list="${CUDA_VISIBLE_DEVICES:-0}"
+IFS=',' read -ra GPULIST <<< "$gpu_list"
+
+CHUNKS=${#GPULIST[@]}
+
+
+for IDX in $(seq 0 $((CHUNKS-1))); do
+  CUDA_VISIBLE_DEVICES=${GPULIST[$IDX]} python3 llava/eval/video/run_inference_video_qa.py \
+      --model_path ${model_path} \
+      --cache_dir ${cache_dir} \
+      --video_dir ${video_dir} \
+      --gt_file_question ${gt_file_question} \
+      --gt_file_answers ${gt_file_answers} \
+      --output_dir ${output_dir} \
+      --output_name ${CHUNKS}_${IDX} \
+      --num_chunks $CHUNKS \
+      --chunk_idx $IDX &
+done
+
+wait
+
+output_file=${output_dir}/merge.jsonl
+
+# Clear out the output file if it exists.
+> "$output_file"
+
+# Loop through the indices and concatenate each file.
+for IDX in $(seq 0 $((CHUNKS-1))); do
+    cat ${output_dir}/${CHUNKS}_${IDX}.json >> "$output_file"
+done

scripts/v1_5/finetune.sh

@@ -0,0 +1,42 @@
+
+
+DATA_ROOT="llava_all_image_video"
+HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 CUDA_VISIBLE_DEVICES=0,1,2,3 deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path lmsys/vicuna-7b-v1.5 \
+    --version v1 \
+    --data_path train_json/videochatgpt_llavaimage_tune.json \
+    --video_folder ${DATA_ROOT} \
+    --image_folder ${DATA_ROOT} \
+    --X "Video" "Image" \
+    --video_tower LanguageBind/LanguageBind_Video \
+    --image_tower LanguageBind/LanguageBind_Image \
+    --pretrain_mm_mlp_adapter checkpoints/Video-LLaVA-Pretrain-7B/mm_projector.bin \
+    --mm_projector_type mlp2x_gelu \
+    --mm_vision_select_layer -2 \
+    --mm_use_x_start_end False \
+    --mm_use_x_patch_token False \
+    --image_aspect_ratio pad \
+    --group_by_modality_length True \
+    --bf16 True \
+    --output_dir ./checkpoints/Video-LLaVA-7B \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 2 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 50000 \
+    --save_total_limit 1 \
+    --learning_rate 2e-5 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 8 \
+    --lazy_preprocess True \
+    --report_to tensorboard \
+    --cache_dir "./cache_dir"

scripts/v1_5/pretrain.sh

@@ -0,0 +1,39 @@
+
+DATA_ROOT="llava_all_image_video"
+CUDA_VISIBLE_DEVICES=0,1,2,3 deepspeed llava/train/train_mem.py \
+    --deepspeed ./scripts/zero2.json \
+    --model_name_or_path lmsys/vicuna-7b-v1.5 \
+    --version plain \
+    --data_path train_json/valley_llavaimage.json \
+    --video_folder ${DATA_ROOT} \
+    --image_folder ${DATA_ROOT} \
+    --X "Video" "Image" \
+    --video_tower LanguageBind/LanguageBind_Video \
+    --image_tower LanguageBind/LanguageBind_Image \
+    --mm_projector_type mlp2x_gelu \
+    --tune_mm_mlp_adapter True \
+    --mm_vision_select_layer -2 \
+    --mm_use_x_start_end False \
+    --mm_use_x_patch_token False \
+    --bf16 True \
+    --output_dir ./checkpoints/Video-LLaVA-Pretrain-7B \
+    --num_train_epochs 1 \
+    --per_device_train_batch_size 32 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 2 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 24000 \
+    --save_total_limit 1 \
+    --learning_rate 1e-3 \
+    --weight_decay 0. \
+    --warmup_ratio 0.03 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 2048 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 8 \
+    --lazy_preprocess True \
+    --report_to tensorboard \
+    --cache_dir "./cache_dir"

scripts/zero2.json

@@ -0,0 +1,23 @@
+{
+    "fp16": {
+        "enabled": "auto",
+        "loss_scale": 0,
+        "loss_scale_window": 1000,
+        "initial_scale_power": 16,
+        "hysteresis": 2,
+        "min_loss_scale": 1
+    },
+    "bf16": {
+        "enabled": "auto"
+    },
+    "train_micro_batch_size_per_gpu": "auto",
+    "train_batch_size": "auto",
+    "gradient_accumulation_steps": "auto",
+    "zero_optimization": {
+        "stage": 2,
+        "overlap_comm": true,
+        "contiguous_gradients": true,
+        "sub_group_size": 1e9,
+        "reduce_bucket_size": "auto"
+    }
+}

scripts/zero3.json

@@ -0,0 +1,28 @@
+{
+    "fp16": {
+        "enabled": "auto",
+        "loss_scale": 0,
+        "loss_scale_window": 1000,
+        "initial_scale_power": 16,
+        "hysteresis": 2,
+        "min_loss_scale": 1
+    },
+    "bf16": {
+        "enabled": "auto"
+    },
+    "train_micro_batch_size_per_gpu": "auto",
+    "train_batch_size": "auto",
+    "gradient_accumulation_steps": "auto",
+    "zero_optimization": {
+        "stage": 3,
+        "overlap_comm": true,
+        "contiguous_gradients": true,
+        "sub_group_size": 1e9,
+        "reduce_bucket_size": "auto",
+        "stage3_prefetch_bucket_size": "auto",
+        "stage3_param_persistence_threshold": "auto",
+        "stage3_max_live_parameters": 1e9,
+        "stage3_max_reuse_distance": 1e9,
+        "stage3_gather_16bit_weights_on_model_save": true
+    }
+}

scripts/zero3_offload.json

@@ -0,0 +1,56 @@
+{
+  "fp16": {
+    "enabled": "auto",
+    "loss_scale": 0,
+    "loss_scale_window": 1000,
+    "initial_scale_power": 16,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "bf16": {
+    "enabled": "auto"
+  },
+  "optimizer": {
+    "type": "AdamW",
+    "params": {
+      "lr": "auto",
+      "betas": "auto",
+      "eps": "auto",
+      "weight_decay": "auto"
+    }
+  },
+  "scheduler": {
+    "type": "WarmupLR",
+    "params": {
+      "warmup_min_lr": "auto",
+      "warmup_max_lr": "auto",
+      "warmup_num_steps": "auto"
+    }
+  },
+  "zero_optimization": {
+    "stage": 3,
+    "offload_optimizer": {
+      "device": "cpu",
+      "pin_memory": true
+    },
+    "offload_param": {
+      "device": "cpu",
+      "pin_memory": true
+    },
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 1e9,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 1e9,
+    "stage3_max_reuse_distance": 1e9,
+    "gather_16bit_weights_on_model_save": true
+  },
+  "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "steps_per_print": 1e5,
+  "wall_clock_breakdown": false
+}