data-preprocessing/preprocess_datasets.py at main · DriveFusion/data-preprocessing · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
import pandas as pd

import os
import json

from paths import *
from drivefusion_preprocessing.dataset_preprocessors import (
    LingoDatasetPreprocessor,
    DriveGPT4BDDXDatasetPreprocessor,
    DriveLMDatasetPreprocessor,
)


def preprocess_lingoqa(system_path: str):
    # Action Part
    print('\n\n--------------------LingoQA Action Part--------------------')
    os.makedirs(OUR_LINGO_ACTION_DATASET_PATH, exist_ok=True)
    df_action = pd.read_parquet(os.path.join(LINGO_ACTION_PATH, 'train.parquet'))

    LINGO_ACTION_IMAGES_PATH = f'{system_path}/datasets/lingoqa_dataset/action' #For datacrunch
    # LINGO_ACTION_IMAGES_PATH = '/content/gdrive/MyDrive/drivefusion_project/datasets/lingoqa_dataset/action' # For Colab

    action_lingo_preprocessor = LingoDatasetPreprocessor(train_data=df_action, eval_data=None, is_single=True)
    action_lingo_preprocessor.llama_train_preprocess(tokens=['<image>'],
                                                     images_path=LINGO_ACTION_IMAGES_PATH,
                                                     output_path=OUR_LINGO_ACTION_DATASET_PATH,
                                                     multi_qa_output_filename='lingoqa_action_llama_format_multi_question.json',
                                                     single_qa_output_filename='lingoqa_action_llama_format_single_question.json')
    action_lingo_preprocessor.llava_train_preprocess(tokens=['<image>\n'],
                                                     images_path=LINGO_ACTION_IMAGES_PATH,
                                                     output_path=OUR_LINGO_ACTION_DATASET_PATH,
                                                     multi_qa_output_filename='lingoqa_action_llava_format_multi_question.json',
                                                     single_qa_output_filename='lingoqa_action_llava_format_single_question.json'
                                                     )

    # Scenery Part
    print('\n\n--------------------LingoQA Scenery Part--------------------')
    os.makedirs(OUR_LINGO_SCENERY_DATASET_PATH, exist_ok=True)
    df_scenery = pd.read_parquet(os.path.join(LINGO_SCENERY_PATH, 'train.parquet'))

    LINGO_SCENERY_IMAGES_PATH = f'{system_path}/datasets/lingoqa_dataset/scenery'
    # LINGO_SCENERY_IMAGES_PATH = '/content/gdrive/MyDrive/drivefusion_project/datasets/lingoqa_dataset/scenery'
    scenery_lingo_preprocessor = LingoDatasetPreprocessor(train_data=df_scenery, eval_data=None, is_single=True)
    scenery_lingo_preprocessor.llama_train_preprocess(tokens=['<image>'],
                                                      images_path=LINGO_SCENERY_IMAGES_PATH,
                                                      output_path=OUR_LINGO_SCENERY_DATASET_PATH,
                                                      multi_qa_output_filename='lingoqa_scenery_llama_format_multi_question.json',
                                                      single_qa_output_filename='lingoqa_scenery_llama_format_single_question.json')
    action_lingo_preprocessor.llava_train_preprocess(tokens=['<image>\n'],
                                                     images_path=LINGO_SCENERY_IMAGES_PATH,
                                                     output_path=OUR_LINGO_SCENERY_DATASET_PATH,
                                                     multi_qa_output_filename='lingoqa_scenery_llava_format_multi_question.json',
                                                     single_qa_output_filename='lingoqa_scenery_llava_format_single_question.json')

    # Eval Part
    print('\n\n--------------------LingoQA Eval Part--------------------')
    os.makedirs(OUR_LINGO_EVAL_DATASET_PATH, exist_ok=True)
    df_eval = pd.read_parquet(os.path.join(LINGO_EVAL_PATH, 'val.parquet'))
    eval_lingo_preprocessor = LingoDatasetPreprocessor(train_data=None, eval_data=df_eval, is_single=False)
    eval_lingo_preprocessor.eval_preprocess(tokens=[''],
                                            output_path=OUR_LINGO_EVAL_DATASET_PATH,
                                            eval_output_filename='lingoqa_eval.json'
                                            )

def preprocess_drivegpt4_bddx(system_path: str):
    # Training Part
    print('\n\n--------------------DriveGPT4 BDDX Training Part--------------------')
    os.makedirs(OUR_DRIVE_BDDX_TRAIN_DATASET_PATH, exist_ok=True)
    with open(os.path.join(DRIVE_BDDX_DATASET_PATH, 'BDD_X_training_label.json'), 'r') as f:
        drive_bddx_train_data = json.load(f)
    with open(os.path.join(DRIVE_BDDX_DATASET_PATH, 'BDD_X_testing_label.json'), 'r') as f:
        drive_bddx_test_data = json.load(f)


    DRIVE_BDDX_IMAGES_PATH = f'{system_path}/datasets/drivegpt4_dataset/BDD_X_imgs_select'

    drivegpt4_preprocessor = DriveGPT4BDDXDatasetPreprocessor(train_data=drive_bddx_train_data, eval_data=drive_bddx_test_data)
    drivegpt4_preprocessor.llama_train_preprocess(tokens=['<image>', '\n<video>'],
                                                  images_path=DRIVE_BDDX_IMAGES_PATH,
                                                  output_path=OUR_DRIVE_BDDX_TRAIN_DATASET_PATH,
                                                  multi_qa_output_filename='drivegpt_bddx_llama_format_training_multi_question.json',
                                                  single_qa_output_filename='drivegpt_bddx_llama_format_training_single_question.json'
                                                  )
    drivegpt4_preprocessor.llava_train_preprocess(tokens=['<image>\n', '\n<video>'],
                                                  images_path=DRIVE_BDDX_IMAGES_PATH,
                                                  output_path=OUR_DRIVE_BDDX_TRAIN_DATASET_PATH,
                                                  multi_qa_output_filename='drivegpt_bddx_llava_format_training_multi_question.json',
                                                  single_qa_output_filename='drivegpt_bddx_llava_format_training_single_question.json'
                                                  )

    # Testing Part
    print('\n\n--------------------DriveGPT4 BDDX Testing Part--------------------')
    os.makedirs(OUR_DRIVE_BDDX_TEST_DATASET_PATH, exist_ok=True)
    drivegpt4_preprocessor.eval_preprocess(tokens=[''],
                                           output_path=OUR_DRIVE_BDDX_TEST_DATASET_PATH,
                                           eval_output_filename='drivegpt_bddx_testing.json'
                                           )

def preprocess_drivelm(system_path: str):
    print('\n\n--------------------DriveLM Training Part--------------------')
    os.makedirs(OUR_DRIVELM_DATASET_PATH, exist_ok=True)
    with open(os.path.join(DRIVELM_DATASET_PATH, 'v1_1_train_nus.json'), 'r') as f:
        drive_lm_data = json.load(f)

    def remove_dots_img_path(json_data: dict) -> dict:
        for _, scene_data in json_data.items():
            # Iterate through each keyframe
            for key_frame_id, key_frame_data in scene_data.get("key_frames", {}).items():
                key_frame_data["image_paths"]["CAM_FRONT"] = key_frame_data["image_paths"]["CAM_FRONT"].replace("../", "")
                key_frame_data["image_paths"]["CAM_FRONT_LEFT"] = key_frame_data["image_paths"]["CAM_FRONT_LEFT"].replace("../", "")
                key_frame_data["image_paths"]["CAM_FRONT_RIGHT"] = key_frame_data["image_paths"]["CAM_FRONT_RIGHT"].replace("../", "")
                key_frame_data["image_paths"]["CAM_BACK"] = key_frame_data["image_paths"]["CAM_BACK"].replace("../", "")
                key_frame_data["image_paths"]["CAM_BACK_LEFT"] = key_frame_data["image_paths"]["CAM_BACK_LEFT"].replace("../", "")
                key_frame_data["image_paths"]["CAM_BACK_RIGHT"] = key_frame_data["image_paths"]["CAM_BACK_RIGHT"].replace("../", "")

        return json_data

    drive_lm_data = remove_dots_img_path(drive_lm_data)
    DRIVELM_IMAGES_PATH = f"{system_path}/datasets/drivelm_dataset/drivelm_nus_imgs_final_train"

    drivelm_preprocessor = DriveLMDatasetPreprocessor(train_data=drive_lm_data, eval_data=None)
    drivelm_preprocessor.llama_train_preprocess(
        tokens=['<image>'],
        images_path=DRIVELM_IMAGES_PATH,
        output_path=OUR_DRIVELM_DATASET_PATH,
        multi_qa_output_filename='drivelm_llama_format_multi_question.json',
        single_qa_output_filename='drivelm_llama_format_single_question.json'
    )
    drivelm_preprocessor.llava_train_preprocess(
        tokens=['<image>\n'],
        images_path=DRIVELM_IMAGES_PATH,
        output_path=OUR_DRIVELM_DATASET_PATH,
        multi_qa_output_filename='drivelm_llava_format_multi_question.json',
        single_qa_output_filename='drivelm_llava_format_single_question.json'
    )

def merge_data():
    with open(os.path.join(OUR_LINGO_ACTION_DATASET_PATH, 'lingoqa_action_llama_format_multi_question.json'), 'r') as f:
        lingo_action_data = json.load(f)
    with open(os.path.join(OUR_LINGO_SCENERY_DATASET_PATH, 'lingoqa_scenery_llama_format_multi_question.json'), 'r') as f:
        lingo_scenery_data = json.load(f)
    print(f"The number of examples in LingoQA Dataset: {len(lingo_action_data) + len(lingo_scenery_data)}")
    with open(os.path.join(OUR_DRIVE_BDDX_TRAIN_DATASET_PATH, 'drivegpt_bddx_llama_format_training_multi_question.json'), 'r') as f:
        drive_bddx_data = json.load(f)
    print(f"The number of examples in DriveGPT4 Dataset: {len(drive_bddx_data)}")
    with open(os.path.join(OUR_DRIVELM_DATASET_PATH, 'drivelm_llama_format_multi_question.json'), 'r') as f:
        drivelm_data = json.load(f)
    print(f"The number of examples in DriveLM Dataset: {len(drivelm_data)}")

    new_data = lingo_action_data + lingo_scenery_data + drive_bddx_data + drivelm_data
    print(f'total training examples: {len(new_data)}')

    with open(os.path.join(OUR_DATASET_PATH, 'training_multi_qa_all_examples.json'), 'w') as f:
        json.dump(new_data, f, indent=4)


def main():
    SYSTEM_PATH = DATACRUNCH_LINUX_SYSTEM_PATH
    preprocess_lingoqa(system_path=SYSTEM_PATH)
    preprocess_drivegpt4_bddx(system_path=SYSTEM_PATH)
    preprocess_drivelm(system_path=SYSTEM_PATH)

    merge_data()
    # check_images_existence(os.path.join(OUR_DATASET_PATH, 'training_multi_qa_all_examples.json'))


if __name__ == '__main__':
    main()