File size: 7,098 Bytes
f31fe4e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from contextlib import contextmanager, nullcontext

import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import BatchSampler, Subset, DataLoader
from torch.cuda import nvtx

torch_function_class = F.cross_entropy.__class__

_REQUIRES_GRAD_ATTR = '_original_requires_grad'

__all__ = [
    'original_requires_grad', 'record_original_requires_grad',
    'restore_original_requires_grad', 'skip_param_grad', 'im2col_2d',
    'im2col_2d_slow', 'cholesky_inv', 'cholesky_solve', 'smw_inv',
    'PseudoBatchLoaderGenerator', 'nvtx_range', 'has_reduction'
]


def original_requires_grad(module=None, param_name=None, param=None):
    if param is None:
        if module is None or param_name is None:
            raise ValueError('Both module and param_name have to be set.')
        param = getattr(module, param_name, None)
    return param is not None and getattr(param, _REQUIRES_GRAD_ATTR)


def record_original_requires_grad(param):
    setattr(param, _REQUIRES_GRAD_ATTR, param.requires_grad)


def restore_original_requires_grad(param):
    param.requires_grad = getattr(param, _REQUIRES_GRAD_ATTR,
                                  param.requires_grad)


@contextmanager
def skip_param_grad(model, disable=False):
    if not disable:
        for param in model.parameters():
            record_original_requires_grad(param)
            param.requires_grad = False

    yield
    if not disable:
        for param in model.parameters():
            restore_original_requires_grad(param)


def im2col_2d(x: torch.Tensor, conv2d: nn.Module):
    if x.ndim != 4:  # n x c x h_in x w_in
        raise ValueError(f'x.ndim has to be 4. Got {x.ndim}.')
    if not isinstance(conv2d, (nn.Conv2d, nn.ConvTranspose2d)):
        raise TypeError(f'conv2d has to be {nn.Conv2d} or {nn.ConvTranspose2d}. Got {type(conv2d)}.')
    if conv2d.dilation != (1, 1):
        raise ValueError(f'conv2d.dilation has to be (1, 1). Got {conv2d.dilation}.')

    ph, pw = conv2d.padding if conv2d.padding != 'valid' else (0, 0)
    kh, kw = conv2d.kernel_size
    sy, sx = conv2d.stride
    if ph + pw > 0:
        x = F.pad(x, (pw, pw, ph, ph)).data
    x = x.unfold(2, kh, sy)  # n x c x h_out x w_in x kh
    x = x.unfold(3, kw, sx)  # n x c x h_out x w_out x kh x kw
    x = x.permute(0, 1, 4, 5, 2,
                  3).contiguous()  # n x c x kh x kw x h_out x w_out
    x = x.view(x.size(0),
               x.size(1) * x.size(2) * x.size(3),
               x.size(4) * x.size(5))  # n x c(kh)(kw) x (h_out)(w_out)
    return x


def im2col_2d_slow(x: torch.Tensor, conv2d: nn.Module):
    if x.ndim != 4:  # n x c x h_in x w_in
        raise ValueError(f'x.ndim has to be 4. Got {x.ndim}.')
    if not isinstance(conv2d, (nn.Conv2d, nn.ConvTranspose2d)):
        raise TypeError(f'conv2d has to be {nn.Conv2d} or {nn.ConvTranspose2d}. Got {type(conv2d)}.')

    padding = conv2d.padding if conv2d.padding != 'valid' else (0, 0)
    # n x c(k_h)(k_w) x (h_out)(w_out)
    Mx = F.unfold(x,
                  conv2d.kernel_size,
                  dilation=conv2d.dilation,
                  padding=padding,
                  stride=conv2d.stride)

    return Mx


def cholesky_inv(X, damping=1e-7):
    diag = torch.diagonal(X)
    diag += damping
    u = torch.linalg.cholesky(X)
    diag -= damping
    return torch.cholesky_inverse(u)


def cholesky_solve(X, b, damping=1e-7):
    diag = torch.diagonal(X)
    diag += damping
    u = torch.linalg.cholesky(X)
    diag -= damping
    return torch.cholesky_solve(b, u)


def smw_inv(x, damping=1e-7):
    n, d = x.shape  # n x d
    I = torch.eye(d, device=x.device)
    G = x @ x.T  # n x n (Gram matrix)
    diag = torch.diagonal(G)
    diag += damping * n
    Ginv_x = torch.linalg.solve(G, x)  # n x d
    xt_Ginv_x = x.T @ Ginv_x  # d x d
    return (I - xt_Ginv_x) / damping  # d x d


class PseudoBatchLoaderGenerator:
    """
    Example::
    >>> # create a base dataloader
    >>> dataset_size = 10
    >>> x_all = torch.tensor(range(dataset_size))
    >>> dataset = torch.utils.data.TensorDataset(x_all)
    >>> data_loader = torch.utils.data.DataLoader(dataset, shuffle=True)
    >>>
    >>> # create a pseudo-batch loader generator
    >>> pb_loader_generator = PseudoBatchLoaderGenerator(data_loader, 5)
    >>>
    >>> for i, pb_loader in enumerate(pb_loader_generator):
    >>>     print(f'pseudo-batch at step {i}')
    >>>     print(list(pb_loader))

    Outputs:
    ```
    pseudo-batch at step 0
    [[tensor([0])], [tensor([1])], [tensor([3])], [tensor([6])], [tensor([7])]]
    pseudo-batch at step 1
    [[tensor([8])], [tensor([5])], [tensor([4])], [tensor([2])], [tensor([9])]]
    ```
    """
    def __init__(self,
                 base_data_loader,
                 pseudo_batch_size,
                 batch_size=None,
                 drop_last=None):
        if batch_size is None:
            batch_size = base_data_loader.batch_size
        if pseudo_batch_size % batch_size != 0:
            raise ValueError(f'pseudo_batch_size ({pseudo_batch_size}) '
                             f'needs to be divisible by batch_size ({batch_size})')
        if drop_last is None:
            drop_last = base_data_loader.drop_last
        base_dataset = base_data_loader.dataset
        sampler_cls = base_data_loader.sampler.__class__
        pseudo_batch_sampler = BatchSampler(sampler_cls(
            range(len(base_dataset))),
                                            batch_size=pseudo_batch_size,
                                            drop_last=drop_last)
        self.batch_size = batch_size
        self.pseudo_batch_sampler = pseudo_batch_sampler
        self.base_dataset = base_dataset
        self.base_data_loader = base_data_loader

    def __iter__(self):
        loader = self.base_data_loader
        for indices in self.pseudo_batch_sampler:
            subset_in_pseudo_batch = Subset(self.base_dataset, indices)
            data_loader = DataLoader(
                subset_in_pseudo_batch,
                batch_size=self.batch_size,
                shuffle=False,
                num_workers=loader.num_workers,
                collate_fn=loader.collate_fn,
                pin_memory=loader.pin_memory,
                drop_last=False,
                timeout=loader.timeout,
                worker_init_fn=loader.worker_init_fn,
                multiprocessing_context=loader.multiprocessing_context,
                generator=loader.generator,
                prefetch_factor=loader.prefetch_factor,
                persistent_workers=loader.persistent_workers)
            yield data_loader

    def __len__(self) -> int:
        return len(self.pseudo_batch_sampler)


@contextmanager
def nvtx_range(msg, *args, **kwargs):
    if torch.cuda.is_available():
        yield nvtx.range(msg, *args, **kwargs)
    else:
        yield nullcontext()


def has_reduction(func):
    if isinstance(func, nn.Module):
        return hasattr(func, 'reduction')
    elif isinstance(func, torch_function_class):
        return 'reduction' in func.__code__.co_varnames
    return False