File size: 6,737 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
from typing import List, Optional

import torch
from transformers import PreTrainedModel

from xtuner.utils import IGNORE_INDEX, IMAGE_TOKEN_INDEX


def prepare_inputs_labels_for_multimodal(
        llm: PreTrainedModel,
        input_ids: torch.LongTensor = None,
        position_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        labels: Optional[torch.LongTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        region_masks = None,
):
    if pixel_values is None:
        return {
            'input_ids': input_ids,
            'position_ids': position_ids,
            'attention_mask': attention_mask,
            'past_key_values': past_key_values,
            'inputs_embeds': None,
            'labels': labels
        }

    # pixel_values (b, n, c)
    new_pixel_values = []
    assert len(pixel_values) == len(region_masks)
    for batch_pixel_values, batch_region_masks in zip(pixel_values, region_masks):
        batch_region_masks = batch_region_masks.flatten(1).to(torch.bool)
        for batch_region_mask in batch_region_masks:
            new_pixel_values.append(batch_pixel_values[batch_region_mask])
    pixel_values = new_pixel_values

    _labels = labels
    _position_ids = position_ids
    _attention_mask = attention_mask
    if attention_mask is None:
        attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
    else:
        attention_mask = attention_mask.bool()
    if position_ids is None:
        position_ids = torch.arange(
            0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
    if labels is None:
        labels = torch.full_like(input_ids, IGNORE_INDEX)

    # remove the padding using attention_mask -- TODO: double check
    input_ids = [
        cur_input_ids[cur_attention_mask]
        for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)
    ]
    labels = [
        cur_labels[cur_attention_mask]
        for cur_labels, cur_attention_mask in zip(labels, attention_mask)
    ]

    new_inputs_embeds = []
    new_labels = []
    cur_image_idx = 0
    for batch_idx, cur_input_ids in enumerate(input_ids):
        num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
        if num_images == 0:
            cur_pixel_values = pixel_values[cur_image_idx]
            cur_inputs_embeds_1 = llm.get_input_embeddings()(cur_input_ids)
            cur_inputs_embeds = torch.cat(
                [cur_inputs_embeds_1, cur_pixel_values[0:0]], dim=0)
            new_inputs_embeds.append(cur_inputs_embeds)
            new_labels.append(labels[batch_idx])
            cur_image_idx += 1
            continue

        image_token_indices = [-1] + torch.where(
            cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [
                cur_input_ids.shape[0]
            ]
        cur_input_ids_noim = []
        cur_labels = labels[batch_idx]
        cur_labels_noim = []
        for i in range(len(image_token_indices) - 1):
            cur_input_ids_noim.append(cur_input_ids[image_token_indices[i] +
                                                    1:image_token_indices[i +
                                                                          1]])
            cur_labels_noim.append(cur_labels[image_token_indices[i] +
                                              1:image_token_indices[i + 1]])
        split_sizes = [x.shape[0] for x in cur_labels_noim]
        cur_inputs_embeds = llm.get_input_embeddings()(
            torch.cat(cur_input_ids_noim))
        cur_inputs_embeds_no_im = torch.split(
            cur_inputs_embeds, split_sizes, dim=0)
        cur_new_inputs_embeds = []
        cur_new_labels = []

        for i in range(num_images + 1):
            cur_new_inputs_embeds.append(cur_inputs_embeds_no_im[i])
            cur_new_labels.append(cur_labels_noim[i])
            if i < num_images:
                cur_pixel_values = pixel_values[cur_image_idx]
                cur_image_idx += 1
                cur_new_inputs_embeds.append(cur_pixel_values)
                cur_new_labels.append(
                    torch.full((cur_pixel_values.shape[0], ),
                               IGNORE_INDEX,
                               device=cur_labels.device,
                               dtype=cur_labels.dtype))

        cur_new_inputs_embeds = torch.cat(cur_new_inputs_embeds)
        cur_new_labels = torch.cat(cur_new_labels)

        new_inputs_embeds.append(cur_new_inputs_embeds)
        new_labels.append(cur_new_labels)

    # Combine them
    max_len = max(x.shape[0] for x in new_inputs_embeds)
    batch_size = len(new_inputs_embeds)

    new_inputs_embeds_padded = []
    new_labels_padded = torch.full((batch_size, max_len),
                                   IGNORE_INDEX,
                                   dtype=new_labels[0].dtype,
                                   device=new_labels[0].device)
    attention_mask = torch.zeros((batch_size, max_len),
                                 dtype=attention_mask.dtype,
                                 device=attention_mask.device)
    position_ids = torch.zeros((batch_size, max_len),
                               dtype=position_ids.dtype,
                               device=position_ids.device)

    for i, (cur_new_embed,
            cur_new_labels) in enumerate(zip(new_inputs_embeds, new_labels)):
        cur_len = cur_new_embed.shape[0]
        new_inputs_embeds_padded.append(
            torch.cat((cur_new_embed,
                       torch.zeros((max_len - cur_len, cur_new_embed.shape[1]),
                                   dtype=cur_new_embed.dtype,
                                   device=cur_new_embed.device)),
                      dim=0))
        if cur_len > 0:
            new_labels_padded[i, :cur_len] = cur_new_labels
            attention_mask[i, :cur_len] = True
            position_ids[i, :cur_len] = torch.arange(
                0,
                cur_len,
                dtype=position_ids.dtype,
                device=position_ids.device)

    new_inputs_embeds = torch.stack(new_inputs_embeds_padded, dim=0)

    if _labels is None:
        new_labels = None
    else:
        new_labels = new_labels_padded

    if _attention_mask is None:
        attention_mask = None
    else:
        attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

    if _position_ids is None:
        position_ids = None
    return {
        'input_ids': None,
        'position_ids': position_ids,
        'attention_mask': attention_mask,
        'past_key_values': past_key_values,
        'inputs_embeds': new_inputs_embeds,
        'labels': new_labels
    }