LargeFeatureFDS/fds.py at main · MSomnia/LargeFeatureFDS · GitHub

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
import logging
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from scipy.ndimage import gaussian_filter1d
from scipy.signal.windows import triang
from util import calibrate_mean_var


class FDS(nn.Module):

    def __init__(self, feature_dim, bucket_num=100, bucket_start=7, start_update=0, start_smooth=1,
                 kernel='gaussian', ks=5, sigma=2, momentum=0.9):
        super(FDS, self).__init__()
        self.feature_dim = feature_dim
        self.bucket_num = bucket_num
        self.bucket_start = bucket_start
        self.kernel_window = self._get_kernel_window(kernel, ks, sigma)
        self.half_ks = (ks - 1) // 2
        self.momentum = momentum
        self.start_update = start_update
        self.start_smooth = start_smooth

        self.register_buffer('epoch', torch.zeros(1).fill_(start_update))
        self.register_buffer('running_mean', torch.zeros(bucket_num - bucket_start, feature_dim))
        self.register_buffer('running_var', torch.ones(bucket_num - bucket_start, feature_dim))
        self.register_buffer('running_mean_last_epoch', torch.zeros(bucket_num - bucket_start, feature_dim))
        self.register_buffer('running_var_last_epoch', torch.ones(bucket_num - bucket_start, feature_dim))
        self.register_buffer('smoothed_mean_last_epoch', torch.zeros(bucket_num - bucket_start, feature_dim))
        self.register_buffer('smoothed_var_last_epoch', torch.ones(bucket_num - bucket_start, feature_dim))
        self.register_buffer('num_samples_tracked', torch.zeros(bucket_num - bucket_start))

    @staticmethod
    def _get_kernel_window(kernel, ks, sigma):
        assert kernel in ['gaussian', 'triang', 'laplace']
        half_ks = (ks - 1) // 2
        if kernel == 'gaussian':
            base_kernel = [0.] * half_ks + [1.] + [0.] * half_ks
            base_kernel = np.array(base_kernel, dtype=np.float32)
            kernel_window = gaussian_filter1d(base_kernel, sigma=sigma) / sum(gaussian_filter1d(base_kernel, sigma=sigma))
        elif kernel == 'triang':
            kernel_window = triang(ks) / sum(triang(ks))
        else:
            laplace = lambda x: np.exp(-abs(x) / sigma) / (2. * sigma)
            kernel_window = list(map(laplace, np.arange(-half_ks, half_ks + 1))) / sum(map(laplace, np.arange(-half_ks, half_ks + 1)))

        logging.info(f'Using FDS: [{kernel.upper()}] ({ks}/{sigma})')
        return torch.tensor(kernel_window, dtype=torch.float32).cuda()

    def _get_bucket_idx(self, label):
        label = np.float32(label.cpu())
        return max(min(int(label * np.float32(10)), self.bucket_num - 1), self.bucket_start)

    def _update_last_epoch_stats(self):
        self.running_mean_last_epoch = self.running_mean
        self.running_var_last_epoch = self.running_var

        self.smoothed_mean_last_epoch = F.conv1d(
            input=F.pad(self.running_mean_last_epoch.unsqueeze(1).permute(2, 1, 0),
                        pad=(self.half_ks, self.half_ks), mode='reflect'),
            weight=self.kernel_window.view(1, 1, -1), padding=0
        ).permute(2, 1, 0).squeeze(1)
        self.smoothed_var_last_epoch = F.conv1d(
            input=F.pad(self.running_var_last_epoch.unsqueeze(1).permute(2, 1, 0),
                        pad=(self.half_ks, self.half_ks), mode='reflect'),
            weight=self.kernel_window.view(1, 1, -1), padding=0
        ).permute(2, 1, 0).squeeze(1)

        assert self.smoothed_mean_last_epoch.shape == self.running_mean_last_epoch.shape, \
            "Smoothed shape is not aligned with running shape!"

    def reset(self):
        self.running_mean.zero_()
        self.running_var.fill_(1)
        self.running_mean_last_epoch.zero_()
        self.running_var_last_epoch.fill_(1)
        self.smoothed_mean_last_epoch.zero_()
        self.smoothed_var_last_epoch.fill_(1)
        self.num_samples_tracked.zero_()

    def update_last_epoch_stats(self, epoch):
        if epoch == self.epoch + 1:
            self.epoch += 1
            self._update_last_epoch_stats()
            logging.info(f"Updated smoothed statistics of last epoch on Epoch [{epoch}]!")

    def _running_stats_to_device(self, device):
        if device == 'cpu':
            self.num_samples_tracked = self.num_samples_tracked.cpu()
            self.running_mean = self.running_mean.cpu()
            self.running_var = self.running_var.cpu()
        else:
            self.num_samples_tracked = self.num_samples_tracked.cuda()
            self.running_mean = self.running_mean.cuda()
            self.running_var = self.running_var.cuda()

    def update_running_stats(self, features, labels, epoch):
        if epoch < self.epoch:
            return

        assert self.feature_dim == features.size(1), "Input feature dimension is not aligned!"
        assert features.size(0) == labels.size(0), "Dimensions of features and labels are not aligned!"

        self._running_stats_to_device('cpu')

        labels = labels.squeeze(1).view(-1)
        features = features.permute(0, 2, 3, 1).contiguous().view(-1, self.feature_dim)

        buckets = np.array([self._get_bucket_idx(label) for label in labels])
        for bucket in np.unique(buckets):
            curr_feats = features[torch.tensor((buckets == bucket).astype(np.bool))]
            curr_num_sample = curr_feats.size(0)
            curr_mean = torch.mean(curr_feats, 0)
            curr_var = torch.var(curr_feats, 0, unbiased=True if curr_feats.size(0) != 1 else False)

            self.num_samples_tracked[bucket - self.bucket_start] += curr_num_sample
            factor = self.momentum if self.momentum is not None else \
                (1 - curr_num_sample / float(self.num_samples_tracked[bucket - self.bucket_start]))
            factor = 0 if epoch == self.start_update else factor
            self.running_mean[bucket - self.bucket_start] = \
                (1 - factor) * curr_mean + factor * self.running_mean[bucket - self.bucket_start]
            self.running_var[bucket - self.bucket_start] = \
                (1 - factor) * curr_var + factor * self.running_var[bucket - self.bucket_start]

        self._running_stats_to_device('cuda')
        logging.info(f"Updated running statistics with Epoch [{epoch}] features!")

    def smooth(self, features, labels, epoch):
        if epoch < self.start_smooth:
            return features

        sp = labels.squeeze(1).shape

        labels = labels.squeeze(1).view(-1)
        features = features.permute(0, 2, 3, 1).contiguous().view(-1, self.feature_dim)

        buckets = torch.max(torch.stack([torch.min(torch.stack([torch.floor(labels * torch.tensor([10.]).cuda()).int(),
            torch.zeros(labels.size(0)).fill_(self.bucket_num - 1).int().cuda()], 0), 0)[0], torch.zeros(labels.size(0)).fill_(self.bucket_start).int().cuda()], 0), 0)[0]
        for bucket in torch.unique(buckets):
            features[buckets.eq(bucket)] = calibrate_mean_var(
                features[buckets.eq(bucket)],
                self.running_mean_last_epoch[bucket.item() - self.bucket_start],
                self.running_var_last_epoch[bucket.item() - self.bucket_start],
                self.smoothed_mean_last_epoch[bucket.item() - self.bucket_start],
                self.smoothed_var_last_epoch[bucket.item() - self.bucket_start]
            )

        return features.view(*sp, self.feature_dim).permute(0, 3, 1, 2)