File size: 7,440 Bytes
310b910
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
import math
import torch
import torch.nn as nn
import torch.distributed as dist
import torch.nn.functional as F
import numpy as np
from einops import rearrange
from torch.utils.data import Sampler

# -- # Common Functions
class InputPadder:
    """ Pads images such that dimensions are divisible by ds """
    def __init__(self, dims, mode='leftend', ds=32):
        self.ht, self.wd = dims[-2:]
        pad_ht = (((self.ht // ds) + 1) * ds - self.ht) % ds
        pad_wd = (((self.wd // ds) + 1) * ds - self.wd) % ds
        if mode == 'leftend':
            self._pad = [0, pad_wd, 0, pad_ht]
        else:
            self._pad = [pad_wd // 2, pad_wd - pad_wd // 2, pad_ht // 2, pad_ht - pad_ht // 2]

        self.mode = mode

    def pad(self, *inputs):
        return [F.pad(x, self._pad, mode='replicate') for x in inputs]

    def unpad(self, x):
        ht, wd = x.shape[-2:]
        c = [self._pad[2], ht - self._pad[3], self._pad[0], wd - self._pad[1]]
        return x[..., c[0]:c[1], c[2]:c[3]]

def coords_gridN(batch, ht, wd, device):
    coords = torch.meshgrid(
        (
            torch.linspace(-1 + 1 / ht, 1 - 1 / ht, ht, device=device),
            torch.linspace(-1 + 1 / wd, 1 - 1 / wd, wd, device=device),
        )
    )

    coords = torch.stack((coords[1], coords[0]), dim=0)[
        None
    ].repeat(batch, 1, 1, 1)
    return coords

def to_cuda(batch):
    for key, value in batch.items():
        if isinstance(value, torch.Tensor):
            batch[key] = value.cuda()
    return batch

def rename_ckpt(ckpt):
    renamed_ckpt = dict()
    for k in ckpt.keys():
        if 'module.' in k:
            renamed_ckpt[k.replace('module.', '')] = torch.clone(ckpt[k])
        else:
            renamed_ckpt[k] = torch.clone(ckpt[k])
    return renamed_ckpt

def resample_rgb(rgb, scaleM, batch, ht, wd, device):
    coords = coords_gridN(batch, ht, wd, device)
    x, y = torch.split(coords, 1, dim=1)
    x = (x + 1) / 2 * wd
    y = (y + 1) / 2 * ht

    scaleM = scaleM.squeeze()

    x = x * scaleM[0, 0] + scaleM[0, 2]
    y = y * scaleM[1, 1] + scaleM[1, 2]

    _, _, orgh, orgw = rgb.shape
    x = x / orgw * 2 - 1.0
    y = y / orgh * 2 - 1.0

    coords = torch.stack([x.squeeze(1), y.squeeze(1)], dim=3)
    rgb_resized = torch.nn.functional.grid_sample(rgb, coords, mode='bilinear', align_corners=True)

    return rgb_resized

def intrinsic2incidence(K, b, h, w, device):
    coords = coords_gridN(b, h, w, device)

    x, y = torch.split(coords, 1, dim=1)
    x = (x + 1) / 2.0 * w
    y = (y + 1) / 2.0 * h

    pts3d = torch.cat([x, y, torch.ones_like(x)], dim=1)
    pts3d = rearrange(pts3d, 'b d h w -> b h w d')
    pts3d = pts3d.unsqueeze(dim=4)

    K_ex = K.view([b, 1, 1, 3, 3])
    pts3d = torch.linalg.inv(K_ex) @ pts3d
    pts3d = torch.nn.functional.normalize(pts3d, dim=3)
    return pts3d

def apply_augmentation(rgb, K, seed=None, augscale=2.0, no_change_prob=0.0):
    _, h, w = rgb.shape

    if seed is not None:
        np.random.seed(seed)

    if np.random.uniform(0, 1) < no_change_prob:
        extension_rx, extension_ry = 1.0, 1.0
    else:
        extension_rx, extension_ry = np.random.uniform(1, augscale), np.random.uniform(1, augscale)

    hs, ws = int(np.ceil(h * extension_ry)), int(np.ceil(w * extension_rx))

    stx = float(np.random.randint(0, int(ws - w + 1), 1).item() + 0.5)
    edx = float(stx + w - 1)
    sty = float(np.random.randint(0, int(hs - h + 1), 1).item() + 0.5)
    edy = float(sty + h - 1)

    stx = stx / ws * w
    edx = edx / ws * w

    sty = sty / hs * h
    edy = edy / hs * h

    ptslt, ptslt_ = np.array([stx, sty, 1]), np.array([0.5, 0.5, 1])
    ptsrt, ptsrt_ = np.array([edx, sty, 1]), np.array([w-0.5, 0.5, 1])
    ptslb, ptslb_ = np.array([stx, edy, 1]), np.array([0.5, h-0.5, 1])
    ptsrb, ptsrb_ = np.array([edx, edy, 1]), np.array([w-0.5, h-0.5, 1])

    pts1 = np.stack([ptslt, ptsrt, ptslb, ptsrb], axis=1)
    pts2 = np.stack([ptslt_, ptsrt_, ptslb_, ptsrb_], axis=1)

    T_num = pts1 @ pts2.T @ np.linalg.inv(pts2 @ pts2.T)
    T = np.eye(3)
    T[0, 0] = T_num[0, 0]
    T[0, 2] = T_num[0, 2]
    T[1, 1] = T_num[1, 1]
    T[1, 2] = T_num[1, 2]
    T = torch.from_numpy(T).float()

    K_trans = torch.inverse(T) @ K

    b = 1
    _, h, w = rgb.shape
    device = rgb.device
    rgb_trans = resample_rgb(rgb.unsqueeze(0), T, b, h, w, device).squeeze(0)
    return rgb_trans, K_trans, T



class IncidenceLoss(nn.Module):
    def __init__(self, loss='cosine'):
        super(IncidenceLoss, self).__init__()
        self.loss = loss
        self.smoothl1 = torch.nn.SmoothL1Loss(beta=0.2)

    def forward(self, incidence, K):
        b, _, h, w = incidence.shape
        device = incidence.device

        incidence_gt = intrinsic2incidence(K, b, h, w, device)
        incidence_gt = incidence_gt.squeeze(4)
        incidence_gt = rearrange(incidence_gt, 'b h w d -> b d h w')

        if self.loss == 'cosine':
            loss = 1 - torch.cosine_similarity(incidence, incidence_gt, dim=1)
        elif self.loss == 'absolute':
            loss = self.smoothl1(incidence, incidence_gt)

        loss = loss.mean()
        return loss


class DistributedSamplerNoEvenlyDivisible(Sampler):
    """Sampler that restricts data loading to a subset of the dataset.

    It is especially useful in conjunction with
    :class:`torch.nn.parallel.DistributedDataParallel`. In such case, each
    process can pass a DistributedSampler instance as a DataLoader sampler,
    and load a subset of the original dataset that is exclusive to it.

    .. note::
        Dataset is assumed to be of constant size.

    Arguments:
        dataset: Dataset used for sampling.
        num_replicas (optional): Number of processes participating in
            distributed training.
        rank (optional): Rank of the current process within num_replicas.
        shuffle (optional): If true (default), sampler will shuffle the indices
    """

    def __init__(self, dataset, num_replicas=None, rank=None, shuffle=True):
        if num_replicas is None:
            if not dist.is_available():
                raise RuntimeError("Requires distributed package to be available")
            num_replicas = dist.get_world_size()
        if rank is None:
            if not dist.is_available():
                raise RuntimeError("Requires distributed package to be available")
            rank = dist.get_rank()
        self.dataset = dataset
        self.num_replicas = num_replicas
        self.rank = rank
        self.epoch = 0
        num_samples = int(math.floor(len(self.dataset) * 1.0 / self.num_replicas))
        rest = len(self.dataset) - num_samples * self.num_replicas
        if self.rank < rest:
            num_samples += 1
        self.num_samples = num_samples
        self.total_size = len(dataset)
        self.shuffle = shuffle

    def __iter__(self):
        # deterministically shuffle based on epoch
        g = torch.Generator()
        g.manual_seed(self.epoch)
        if self.shuffle:
            indices = torch.randperm(len(self.dataset), generator=g).tolist()
        else:
            indices = list(range(len(self.dataset)))

        # subsample
        indices = indices[self.rank:self.total_size:self.num_replicas]
        self.num_samples = len(indices)

        return iter(indices)

    def __len__(self):
        return self.num_samples

    def set_epoch(self, epoch):
        self.epoch = epoch