File size: 8,227 Bytes
032e687
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
# Copyright (c) OpenMMLab. All rights reserved.
import random
from xtuner.utils import DEFAULT_IMAGE_TOKEN

SEG_QUESTIONS = [
    "Can you segment the {class_name} in this image?",
    "Please segment {class_name} in this image.",
    "What is {class_name} in this image? Please respond with segmentation mask.",
    "What is {class_name} in this image? Please output segmentation mask.",

    "Can you segment the {class_name} in this image",
    "Please segment {class_name} in this image",
    "What is {class_name} in this image? Please respond with segmentation mask",
    "What is {class_name} in this image? Please output segmentation mask",

    "Could you provide a segmentation mask for the {class_name} in this image?",
    "Please identify and segment the {class_name} in this image.",
    "Where is the {class_name} in this picture? Please respond with a segmentation mask.",
    "Can you highlight the {class_name} in this image with a segmentation mask?",

    "Could you provide a segmentation mask for the {class_name} in this image",
    "Please identify and segment the {class_name} in this image",
    "Where is the {class_name} in this picture? Please respond with a segmentation mask",
    "Can you highlight the {class_name} in this image with a segmentation mask",
]

ANSWER_LIST = [
    "It is [SEG].",
    "Sure, [SEG].",
    "Sure, it is [SEG].",
    "Sure, the segmentation result is [SEG].",
    "[SEG].",
]

ANSWER_LIST_GCG_FORMAT = [
    "<p> {} </p> [SEG].",
]

def semantic_seg_conversations(labels):
    ret = []
    for i, label in enumerate(labels):
        label = label.strip()
        assert len(label.split("||")) == 1
        for question_template in SEG_QUESTIONS:
            for answer_template in ANSWER_LIST:
                item = {}
                item['conversations'] = [{'from': 'human', 'value': DEFAULT_IMAGE_TOKEN+question_template.format(class_name=label.lower())},
                                         {'from': 'gpt', 'value': answer_template}]
                item['class_id'] = i
                ret.append(item)
    return ret

def semantic_seg_map_fn(example):
    # example {'conversations', 'class_id'}
    messages = example['conversations']
    input = ''
    conversation = []
    while messages and messages[0]['from'] == 'gpt':
        # Skip the first one if it is from gpt
        messages = messages[1:]
    for msg in messages:
        if msg['from'] == 'human':
            if DEFAULT_IMAGE_TOKEN in msg['value']:
                msg['value'] = msg['value'].replace(DEFAULT_IMAGE_TOKEN,
                                                    '').strip()
                msg['value'] = DEFAULT_IMAGE_TOKEN + '\n' + msg['value']
                msg['value'] = msg['value'].strip()
            input += msg['value']

        elif msg['from'] == 'gpt':
            conversation.append({'input': input, 'output': msg['value']})
            input = ''
        else:
            raise NotImplementedError
    example.update({'conversation': conversation})
    return example

def pascal_part_conversation(selected_labels):
    conversations = []
    for i, selected_label in enumerate(selected_labels):
        question = random.choice(SEG_QUESTIONS).format(class_name=selected_label.lower()).strip()
        answer = random.choice(ANSWER_LIST)
        if i == 0:
            question = DEFAULT_IMAGE_TOKEN + question
        conversations.append({'from': 'human', 'value': question})
        conversations.append({'from': 'gpt', 'value': answer})
    return conversations

def pascal_part_preprocess(example):
    selected_labels = example["selected_labels"]
    conversations = pascal_part_conversation(selected_labels)
    example['conversations'] = conversations
    return example

def pascal_part_map_fn(example):
    example = pascal_part_preprocess(example)
    example['image'] = example["file_name"]
    # do llava preprocess
    messages = example['conversations']
    input = ''
    conversation = []
    while messages and messages[0]['from'] == 'gpt':
        # Skip the first one if it is from gpt
        messages = messages[1:]
    for msg in messages:
        if msg['from'] == 'human':
            if DEFAULT_IMAGE_TOKEN in msg['value']:
                msg['value'] = msg['value'].replace(DEFAULT_IMAGE_TOKEN,
                                                    '').strip()
                msg['value'] = DEFAULT_IMAGE_TOKEN + '\n' + msg['value']
                msg['value'] = msg['value'].strip()
            input += msg['value']

        elif msg['from'] == 'gpt':
            conversation.append({'input': input, 'output': msg['value']})
            input = ''
        else:
            raise NotImplementedError
    example.update({'conversation': conversation})
    return example


def semantic_seg_gcg_format_conversations(labels):
    ret = []
    for i, label in enumerate(labels):
        label = label.strip()
        assert len(label.split("||")) == 1
        for question_template in SEG_QUESTIONS:
            for answer_template in ANSWER_LIST_GCG_FORMAT:
                item = {}
                item['conversations'] = [{'from': 'human', 'value': DEFAULT_IMAGE_TOKEN+question_template.format(class_name=label.lower())},
                                         {'from': 'gpt', 'value': answer_template.format(label.lower().capitalize())}]
                item['class_id'] = i
                ret.append(item)
    return ret

def semantic_seg_gcg_format_map_fn(example):
    # example {'conversations', 'class_id'}
    messages = example['conversations']
    input = ''
    conversation = []
    while messages and messages[0]['from'] == 'gpt':
        # Skip the first one if it is from gpt
        messages = messages[1:]
    for msg in messages:
        if msg['from'] == 'human':
            if DEFAULT_IMAGE_TOKEN in msg['value']:
                msg['value'] = msg['value'].replace(DEFAULT_IMAGE_TOKEN,
                                                    '').strip()
                msg['value'] = DEFAULT_IMAGE_TOKEN + '\n' + msg['value']
                msg['value'] = msg['value'].strip()
            input += msg['value']

        elif msg['from'] == 'gpt':
            conversation.append({'input': input, 'output': msg['value']})
            input = ''
        else:
            raise NotImplementedError
    example.update({'conversation': conversation})
    return example

def pascal_part_gcg_format_conversation(selected_labels):
    conversations = []
    for i, selected_label in enumerate(selected_labels):
        question = random.choice(SEG_QUESTIONS).format(class_name=selected_label.lower()).strip()
        answer = random.choice(ANSWER_LIST).format(selected_label.lower().capitalize())
        if i == 0:
            question = DEFAULT_IMAGE_TOKEN + question
        conversations.append({'from': 'human', 'value': question})
        conversations.append({'from': 'gpt', 'value': answer})
    return conversations

def pascal_part_gcg_format_preprocess(example):
    selected_labels = example["selected_labels"]
    conversations = pascal_part_gcg_format_conversation(selected_labels)
    example['conversations'] = conversations
    return example

def pascal_part_gcg_format_map_fn(example):
    example = pascal_part_gcg_format_preprocess(example)
    example['image'] = example["file_name"]
    # do llava preprocess
    messages = example['conversations']
    input = ''
    conversation = []
    while messages and messages[0]['from'] == 'gpt':
        # Skip the first one if it is from gpt
        messages = messages[1:]
    for msg in messages:
        if msg['from'] == 'human':
            if DEFAULT_IMAGE_TOKEN in msg['value']:
                msg['value'] = msg['value'].replace(DEFAULT_IMAGE_TOKEN,
                                                    '').strip()
                msg['value'] = DEFAULT_IMAGE_TOKEN + '\n' + msg['value']
                msg['value'] = msg['value'].strip()
            input += msg['value']

        elif msg['from'] == 'gpt':
            conversation.append({'input': input, 'output': msg['value']})
            input = ''
        else:
            raise NotImplementedError
    example.update({'conversation': conversation})
    return example