File size: 7,759 Bytes
6073e55
23fdbc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2025 Ye Liu. Licensed under the BSD-3-Clause License.

import math
import random
from collections import defaultdict
from itertools import accumulate

import nncore
import numpy as np
import termplotlib as tpl
import torch
from tabulate import tabulate
from torch.utils.data import Dataset

from videomind.constants import IGNORE_INDEX
from videomind.dataset.utils import preprocess, process_vision_info
from videomind.utils.parser import parse_span

DATASETS = nncore.Registry('datasets')


class HybridDataset(Dataset):

    def __init__(self, processor, model_config, model_args, data_args, training_args):
        super().__init__()

        datasets = []
        for key in data_args.datasets.split(','):
            datasets.append(DATASETS.get(key)(processor, model_args, data_args, training_args))

        data_types = [a['data_type'] for d in datasets for a in d.annos]

        cum_length = [0] + list(accumulate([len(d) for d in datasets]))
        idx_ranges = [[cum_length[i], cum_length[i + 1]] for i in range(len(cum_length) - 1)]

        if training_args.local_rank in (0, -1):
            raw_length = sum(d.raw_length for d in datasets)
            cur_length = idx_ranges[-1][-1]

            ratio = round(cur_length / raw_length * 100, 2)
            print(f'Number of samples: {raw_length} (original) -> {cur_length} (filtered) {ratio}%')

            data_type_cnt = ' '.join([f'{data_types.count(t)} ({t})' for t in list(set(data_types))])
            print(f'Data types: {data_type_cnt}')

            tab = defaultdict(int)
            for dataset in datasets:
                for anno in dataset.annos:
                    tab[anno.get('source', 'unknown')] += 1

            tab = [[k, v, round(v / cur_length, 3)] for k, v in tab.items()]
            print(tabulate(tab, headers=['Source', '#Samples', 'Ratio'], tablefmt='pretty', stralign='left'))

            d, _ = torch.Tensor([a['duration'] for d in datasets for a in d.annos if 'duration' in a]).sort()
            if d.size(0) > 0:
                n, r = min(d.size(0), 10), d.flip(0)
                print(f'Top-{n} max video durations: {[round(r[i].item(), 1) for i in range(n)]}')
                print(f'Top-{n} min video durations: {[round(d[i].item(), 1) for i in range(n)]}')
                print(f'Average video duration ({d.size(0)} samples): {round(d.mean().item(), 1)}s')

                print('Video duration histogram:')
                counts, edges = np.histogram(d)
                labels = [f'{edges[i]:.2f}s - {edges[i + 1]:.2f}s' for i in range(len(edges) - 1)]
                fig = tpl.figure()
                fig.barh(counts, labels)
                fig.show()

            d, _ = torch.Tensor([abs(b[0] - b[1]) for d in datasets for a in d.annos if 'span' in a
                                 for b in a['span']]).sort()
            if d.size(0) > 0:
                n, r = min(d.size(0), 10), d.flip(0)
                print(f'Top-{n} max span durations: {[round(r[i].item(), 1) for i in range(n)]}')
                print(f'Top-{n} min span durations: {[round(d[i].item(), 1) for i in range(n)]}')
                print(f'Average span duration ({d.size(0)} samples): {round(d.mean().item(), 1)}s')

                print('Span duration histogram:')
                counts, edges = np.histogram(d)
                labels = [f'{edges[i]:.2f}s - {edges[i + 1]:.2f}s' for i in range(len(edges) - 1)]
                fig = tpl.figure()
                fig.barh(counts, labels)
                fig.show()

        self.datasets = datasets
        self.data_types = data_types
        self.idx_ranges = idx_ranges
        self.processor = processor
        self.model_config = model_config
        self.model_args = model_args
        self.data_args = data_args
        self.training_args = training_args

    def __len__(self):
        return self.idx_ranges[-1][-1]

    def __getitem__(self, idx):
        for retry in range(self.data_args.max_retries + 1):
            try:
                return self.fetch_data(idx)
            except Exception as e:
                print(f'Error in loading {idx}: {type(e).__name__}({e})')
                idx = random.choice([i for i, t in enumerate(self.data_types) if t == self.data_types[idx]])

        raise RuntimeError(f'Data loading failed after {retry} retries')

    def map(self, *args, **kwargs):
        return self

    def fetch_data(self, idx):
        for (s, e), dataset in zip(self.idx_ranges, self.datasets):
            if s <= idx < e:
                meta = dataset[idx - s]
                break

        text = self.processor.apply_chat_template(meta['messages'])
        text = [text.strip()]

        images, videos = process_vision_info(meta['messages'], sanity_check=True)

        data = self.processor(text=text, images=images, videos=videos, return_tensors='pt')
        assert data['input_ids'].size(0) == 1

        data['input_ids'] = data['input_ids'][0]
        data['labels'] = preprocess(data['input_ids'], text[0], self.processor.tokenizer, self.model_args.conv_type)

        # insert segment start/end tokens
        if 'ss' in meta and 'se' in meta:
            video_grid_thw = data['video_grid_thw'][0]
            num_frames, window = int(video_grid_thw[0]), int(video_grid_thw[1] * video_grid_thw[2] / 4)
            assert num_frames * window * 4 == data['pixel_values_videos'].size(0)

            pos_s, pos_e = round(meta['ss'] * num_frames), round(meta['se'] * num_frames)
            pos_s, pos_e = min(max(0, pos_s), num_frames), min(max(0, pos_e), num_frames)
            assert pos_s <= pos_e, (num_frames, meta['ss'], meta['se'])

            base_idx = torch.nonzero(data['input_ids'] == self.model_config.vision_start_token_id).item()
            pos_s, pos_e = pos_s * window + base_idx + 1, pos_e * window + base_idx + 2

            input_ids = data['input_ids'].tolist()
            input_ids.insert(pos_s, self.model_config.seg_s_token_id)
            input_ids.insert(pos_e, self.model_config.seg_e_token_id)
            data['input_ids'] = torch.LongTensor(input_ids)

            labels = data['labels'].tolist()
            labels.insert(pos_s, IGNORE_INDEX)
            labels.insert(pos_e, IGNORE_INDEX)
            data['labels'] = torch.LongTensor(labels)

        if 'span' in meta:
            span, duration = meta['span'], meta['duration']

            pixel_values_videos, video_grid_thw = data['pixel_values_videos'], data['video_grid_thw']
            num_frames = int(video_grid_thw[0][0])

            assert video_grid_thw.size(0) == 1
            assert video_grid_thw.prod() == pixel_values_videos.size(0)

            # actual fps would be 1/2 of config (temporal patch size = 2)
            fps = num_frames / duration

            safe_span = [parse_span(b, duration, 1 / fps) for b in span]

            # num_reg_tokens -> num_bnds -> s & e
            timestamps = [[[s / duration, e / duration] for s, e in safe_span]]

            saliency, pos_inds = torch.zeros(num_frames), []
            for s, e in safe_span:
                span_ind = max(0, s * fps), min(e * fps, num_frames)
                pos_inds = list(range(math.ceil(span_ind[0]), math.ceil(span_ind[1])))
                assert len(pos_inds) > 0, f'empty pos_inds ({idx}): {fps} {num_frames} {duration} {span}'
                saliency[pos_inds] = 1

            assert saliency.any(), f'empty saliency ({idx}): {pos_inds} {fps} {num_frames} {duration} {span}'
            pos_clip = random.sample(saliency.nonzero()[:, 0].tolist(), 1)
            pos_clip = torch.LongTensor(pos_clip)

            data['timestamps'] = timestamps
            data['saliency'] = saliency
            data['pos_clip'] = pos_clip

        return data