Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
data.py 4.73 KB
Newer Older
Vijay Korthikanti's avatar
Fourth phase vision merge: classification and segmentation tasks  
Vijay Korthikanti committed
1
2
3
4
5
6
7
8
9 
import random
import os
import math
import mmcv
import torch
import numpy as np
import torchvision.transforms as T
from torchvision import datasets
from torch.utils.data import Dataset
xingjinliang's avatar
升级megatron_v0.10  
xingjinliang committed
10 
from megatron.legacy.data.autoaugment import ImageNetPolicy
Vijay Korthikanti's avatar
Fourth phase vision merge: classification and segmentation tasks  
Vijay Korthikanti committed
11
12 
from tasks.vision.segmentation.cityscapes import Cityscapes
import tasks.vision.segmentation.transforms as ET
xingjinliang's avatar
升级megatron_v0.10  
xingjinliang committed
13
14 
from megatron.legacy.data.autoaugment import ImageNetPolicy
from megatron.training import get_args
Vijay Korthikanti's avatar
Fourth phase vision merge: classification and segmentation tasks  
Vijay Korthikanti committed
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 
from PIL import Image, ImageOps


class VitSegmentationJointTransform():
    def __init__(self, train=True, resolution=None):
        self.train = train
        if self.train:
            self.transform0 = ET.RandomSizeAndCrop(resolution)
            self.transform1 = ET.RandomHorizontallyFlip()

    def __call__(self, img, mask):
        if self.train:
            img, mask = self.transform0(img, mask)
            img, mask = self.transform1(img, mask)
        return img, mask


class VitSegmentationImageTransform():
    def __init__(self, train=True, resolution=None):
        args = get_args()
        self.train = train
        assert args.fp16 or args.bf16
        self.data_type = torch.half if args.fp16 else torch.bfloat16
        self.mean_std = args.mean_std
        if self.train:
            assert resolution is not None
            self.transform = T.Compose([
                ET.PhotoMetricDistortion(),
                T.ToTensor(),
                T.Normalize(*self.mean_std),
                T.ConvertImageDtype(self.data_type)
            ])
        else:
            self.transform = T.Compose([
                T.ToTensor(),
                T.Normalize(*self.mean_std),
                T.ConvertImageDtype(self.data_type)
            ])

    def __call__(self, input):
        output = self.transform(input)
        return output


class VitSegmentationTargetTransform():
    def __init__(self, train=True, resolution=None):
        self.train = train

    def __call__(self, input):
        output = torch.from_numpy(np.array(input, dtype=np.int32)).long()
        return output


class RandomSeedSegmentationDataset(Dataset):
    def __init__(self,
                 dataset,
                 joint_transform,
                 image_transform,
                 target_transform):

        args = get_args()
        self.base_seed = args.seed
        self.curr_seed = self.base_seed
        self.dataset = dataset
        self.joint_transform = joint_transform
        self.image_transform = image_transform
        self.target_transform = target_transform

    def __len__(self):
        return len(self.dataset)

    def set_epoch(self, epoch):
        self.curr_seed = self.base_seed + 100 * epoch

    def __getitem__(self, idx):
        seed = idx + self.curr_seed
        img, mask = self.dataset[idx]

        torch.manual_seed(seed)
        random.seed(seed)
        np.random.seed(seed)
        img, mask = self.joint_transform(img, mask)
        img = self.image_transform(img)
        mask = self.target_transform(mask)

        return img, mask


def build_cityscapes_train_valid_datasets(data_path, image_size):
    args = get_args()
    args.num_classes = Cityscapes.num_classes
    args.ignore_index = Cityscapes.ignore_index
    args.color_table = Cityscapes.color_table
    args.mean_std = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])

    train_joint_transform = \
        VitSegmentationJointTransform(train=True, resolution=image_size)
    val_joint_transform = \
        VitSegmentationJointTransform(train=False, resolution=image_size)
    train_image_transform = \
        VitSegmentationImageTransform(train=True, resolution=image_size)
    val_image_transform = \
        VitSegmentationImageTransform(train=False, resolution=image_size)
    train_target_transform = \
        VitSegmentationTargetTransform(train=True, resolution=image_size)
    val_target_transform = \
        VitSegmentationTargetTransform(train=False, resolution=image_size)

    # training dataset
    train_data = Cityscapes(
        root=data_path[0],
        split='train',
        mode='fine',
        resolution=image_size
    )
    train_data = RandomSeedSegmentationDataset(
        train_data,
        joint_transform=train_joint_transform,
        image_transform=train_image_transform,
        target_transform=train_target_transform)

    # validation dataset
    val_data = Cityscapes(
        root=data_path[0],
        split='val',
        mode='fine',
        resolution=image_size
    )

    val_data = RandomSeedSegmentationDataset(
        val_data,
        joint_transform=val_joint_transform,
        image_transform=val_image_transform,
        target_transform=val_target_transform)

    return train_data, val_data


def build_train_valid_datasets(data_path, image_size):
    return build_cityscapes_train_valid_datasets(data_path, image_size)