File size: 9,197 Bytes
2875fe6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import numpy as np
import torch.nn.functional as F


def costume_collate(data, max_len=None, mask_compensation=False):
    """Build mini-batch tensors from a list of (X, mask) tuples. Mask input. Create

    Args:

        data: len(batch_size) list of tuples (X, mask).

            - X: torch tensor of shape (seq_length, feat_dim); variable seq_length.

            - mask: boolean torch tensor of shape (seq_length, feat_dim); variable seq_length.

        max_len: global fixed sequence length. Used for architectures requiring fixed length input,

            where the batch length cannot vary dynamically. Longer sequences are clipped, shorter are padded with 0s

    Returns:

        X: (batch_size, padded_length, feat_dim) torch tensor of masked features (input)

        targets: (batch_size, padded_length, feat_dim) torch tensor of unmasked features (output)

        target_masks: (batch_size, padded_length, feat_dim) boolean torch tensor

            0 indicates masked values to be predicted, 1 indicates unaffected/"active" feature values

        padding_masks: (batch_size, padded_length) boolean tensor, 1 means keep vector at this position, 0 ignore (padding)

    """

    batch_size = len(data)
    features, masks = zip(*data)

    # Stack and pad features and masks (convert 2D to 3D tensors, i.e. add batch dimension)
    lengths = [
        X.shape[0] for X in features
    ]  # original sequence length for each time series
    if max_len is None:
        max_len = max(lengths)
    X = torch.zeros(
        batch_size, max_len, features[0].shape[-1]
    )  # (batch_size, padded_length, feat_dim)
    target_masks = torch.zeros_like(
        X, dtype=torch.bool
    )  # (batch_size, padded_length, feat_dim) masks related to objective
    for i in range(batch_size):
        end = min(lengths[i], max_len)
        X[i, :end, :] = features[i][:end, :]
        target_masks[i, :end, :] = masks[i][:end, :]

    targets = X.clone()
    X = X * target_masks  # mask input
    if mask_compensation:
        X = compensate_masking(X, target_masks)

    padding_masks = padding_mask(
        torch.tensor(lengths, dtype=torch.int16), max_len=max_len
    )  # (batch_size, padded_length) boolean tensor, "1" means keep
    # target_masks = ~target_masks  # inverse logic: 0 now means ignore, 1 means predict
    return X, targets, target_masks, padding_masks


def compensate_masking(X, mask):
    """

    Compensate feature vectors after masking values, in a way that the matrix product W @ X would not be affected on average.

    If p is the proportion of unmasked (active) elements, X' = X / p = X * feat_dim/num_active

    Args:

        X: (batch_size, seq_length, feat_dim) torch tensor

        mask: (batch_size, seq_length, feat_dim) torch tensor: 0s means mask and predict, 1s: unaffected (active) input

    Returns:

        (batch_size, seq_length, feat_dim) compensated features

    """

    # number of unmasked elements of feature vector for each time step
    num_active = torch.sum(mask, dim=-1).unsqueeze(-1)  # (batch_size, seq_length, 1)
    # to avoid division by 0, set the minimum to 1
    num_active = torch.max(
        num_active, torch.ones(num_active.shape, dtype=torch.int16)
    )  # (batch_size, seq_length, 1)
    return X.shape[-1] * X / num_active


def padding_mask(lengths, max_len=None):
    """

    Used to mask padded positions: creates a (batch_size, max_len) boolean mask from a tensor of sequence lengths,

    where 1 means keep element at this position (time step)

    """
    batch_size = lengths.numel()
    max_len = (
        max_len or lengths.max_val()
    )  # trick works because of overloading of 'or' operator for non-boolean types
    return (
        torch.arange(0, max_len, device=lengths.device)
        .type_as(lengths)
        .repeat(batch_size, 1)
        .lt(lengths.unsqueeze(1))
    )


def noise_mask(

    X,

    masking_ratio,

    lm=3,

    mode="separate",

    distribution="geometric",

    exclude_feats=None,

):
    """

    Creates a random boolean mask of the same shape as X, with 0s at places where a feature should be masked.

    Args:

        X: (seq_length, feat_dim) numpy array of features corresponding to a single sample

        masking_ratio: proportion of seq_length to be masked. At each time step, will also be the proportion of

            feat_dim that will be masked on average

        lm: average length of masking subsequences (streaks of 0s). Used only when `distribution` is 'geometric'.

        mode: whether each variable should be masked separately ('separate'), or all variables at a certain positions

            should be masked concurrently ('concurrent')

        distribution: whether each mask sequence element is sampled independently at random, or whether

            sampling follows a markov chain (and thus is stateful), resulting in geometric distributions of

            masked squences of a desired mean length `lm`

        exclude_feats: iterable of indices corresponding to features to be excluded from masking (i.e. to remain all 1s)



    Returns:

        boolean numpy array with the same shape as X, with 0s at places where a feature should be masked

    """
    if exclude_feats is not None:
        exclude_feats = set(exclude_feats)

    if distribution == "geometric":  # stateful (Markov chain)
        if mode == "separate":  # each variable (feature) is independent
            mask = np.ones(X.shape, dtype=bool)
            for m in range(X.shape[1]):  # feature dimension
                if exclude_feats is None or m not in exclude_feats:
                    mask[:, m] = geom_noise_mask_single(
                        X.shape[0], lm, masking_ratio
                    )  # time dimension
        else:  # replicate across feature dimension (mask all variables at the same positions concurrently)
            mask = np.tile(
                np.expand_dims(
                    geom_noise_mask_single(X.shape[0], lm, masking_ratio), 1
                ),
                X.shape[1],
            )
    else:  # each position is independent Bernoulli with p = 1 - masking_ratio
        if mode == "separate":
            mask = np.random.choice(
                np.array([True, False]),
                size=X.shape,
                replace=True,
                p=(1 - masking_ratio, masking_ratio),
            )
        else:
            mask = np.tile(
                np.random.choice(
                    np.array([True, False]),
                    size=(X.shape[0], 1),
                    replace=True,
                    p=(1 - masking_ratio, masking_ratio),
                ),
                X.shape[1],
            )

    return mask


def geom_noise_mask_single(L, lm, masking_ratio):
    """

    Randomly create a boolean mask of length `L`, consisting of subsequences of average length lm, masking with 0s a `masking_ratio`

    proportion of the sequence L. The length of masking subsequences and intervals follow a geometric distribution.

    Args:

        L: length of mask and sequence to be masked

        lm: average length of masking subsequences (streaks of 0s)

        masking_ratio: proportion of L to be masked



    Returns:

        (L,) boolean numpy array intended to mask ('drop') with 0s a sequence of length L

    """
    keep_mask = np.ones(L, dtype=bool)
    p_m = (
        1 / lm
    )  # probability of each masking sequence stopping. parameter of geometric distribution.
    p_u = p_m * masking_ratio / (1 - masking_ratio)
    # probability of each unmasked sequence stopping. parameter of geometric distribution.
    p = [p_m, p_u]

    # Start in state 0 with masking_ratio probability
    state = int(
        np.random.rand() > masking_ratio
    )  # state 0 means masking, 1 means not masking
    for i in range(L):
        keep_mask[i] = (
            state  # here it happens that state and masking value corresponding to state are identical
        )
        if np.random.rand() < p[state]:
            state = 1 - state

    return keep_mask


def uniform_noise_mask_single(L, lm, masking_ratio):
    """

    Randomly create a boolean mask of length `L`, consisting of subsequences of average length lm, masking with 0s a `masking_ratio`

    proportion of the sequence L. The length of masking subsequences and intervals follow a uniform distribution.

    Args:

        L: length of mask and sequence to be masked

        lm: average length of masking subsequences (streaks of 0s)

        masking_ratio: proportion of L to be masked

    """
    keep_mask = np.ones(L, dtype=bool)
    # Start in state 0 with masking_ratio probability
    state = int(
        np.random.rand() > masking_ratio
    )  # state 0 means masking, 1 means not masking
    for i in range(L):
        keep_mask[i] = (
            state  # here it happens that state and masking value corresponding to state are identical
        )
        if np.random.rand() < 1 / lm:
            state = 1 - state

    return keep_mask