File size: 10,179 Bytes
2f85de4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# python3.7

# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.

"""Custom replacement for convolution operators.

Operators in this file support arbitrarily high order gradients with zero
performance penalty. Please set `impl` as `cuda` to use faster customized
operators, OR as `ref` to use native `torch.nn.functional.conv2d` and
`torch.nn.functional.conv_transpose2d`.

Please refer to https://github.com/NVlabs/stylegan3
"""

# pylint: disable=redefined-builtin
# pylint: disable=arguments-differ
# pylint: disable=protected-access
# pylint: disable=line-too-long
# pylint: disable=global-statement
# pylint: disable=missing-class-docstring
# pylint: disable=missing-function-docstring

import contextlib
import torch

#----------------------------------------------------------------------------

enabled = True                      # Enable the custom op by setting this to true.
weight_gradients_disabled = False   # Forcefully disable computation of gradients with respect to the weights.

@contextlib.contextmanager
def no_weight_gradients(disable=True):
    global weight_gradients_disabled
    old = weight_gradients_disabled
    if disable:
        weight_gradients_disabled = True
    yield
    weight_gradients_disabled = old

#----------------------------------------------------------------------------

def conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1, impl='cuda'):
    if impl == 'cuda' and _should_use_custom_op(input):
        return _conv2d_gradfix(transpose=False, weight_shape=weight.shape, stride=stride, padding=padding, output_padding=0, dilation=dilation, groups=groups).apply(input, weight, bias)
    return torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, dilation=dilation, groups=groups)

def conv_transpose2d(input, weight, bias=None, stride=1, padding=0, output_padding=0, groups=1, dilation=1, impl='cuda'):
    if impl == 'cuda' and _should_use_custom_op(input):
        return _conv2d_gradfix(transpose=True, weight_shape=weight.shape, stride=stride, padding=padding, output_padding=output_padding, groups=groups, dilation=dilation).apply(input, weight, bias)
    return torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, stride=stride, padding=padding, output_padding=output_padding, groups=groups, dilation=dilation)

#----------------------------------------------------------------------------

def _should_use_custom_op(input):
    assert isinstance(input, torch.Tensor)
    if (not enabled) or (not torch.backends.cudnn.enabled):
        return False
    if input.device.type != 'cuda':
        return False
    return True

def _tuple_of_ints(xs, ndim):
    xs = tuple(xs) if isinstance(xs, (tuple, list)) else (xs,) * ndim
    assert len(xs) == ndim
    assert all(isinstance(x, int) for x in xs)
    return xs

#----------------------------------------------------------------------------

_conv2d_gradfix_cache = dict()
_null_tensor = torch.empty([0])

def _conv2d_gradfix(transpose, weight_shape, stride, padding, output_padding, dilation, groups):
    # Parse arguments.
    ndim = 2
    weight_shape = tuple(weight_shape)
    stride = _tuple_of_ints(stride, ndim)
    padding = _tuple_of_ints(padding, ndim)
    output_padding = _tuple_of_ints(output_padding, ndim)
    dilation = _tuple_of_ints(dilation, ndim)

    # Lookup from cache.
    key = (transpose, weight_shape, stride, padding, output_padding, dilation, groups)
    if key in _conv2d_gradfix_cache:
        return _conv2d_gradfix_cache[key]

    # Validate arguments.
    assert groups >= 1
    assert len(weight_shape) == ndim + 2
    assert all(stride[i] >= 1 for i in range(ndim))
    assert all(padding[i] >= 0 for i in range(ndim))
    assert all(dilation[i] >= 0 for i in range(ndim))
    if not transpose:
        assert all(output_padding[i] == 0 for i in range(ndim))
    else: # transpose
        assert all(0 <= output_padding[i] < max(stride[i], dilation[i]) for i in range(ndim))

    # Helpers.
    common_kwargs = dict(stride=stride, padding=padding, dilation=dilation, groups=groups)
    def calc_output_padding(input_shape, output_shape):
        if transpose:
            return [0, 0]
        return [
            input_shape[i + 2]
            - (output_shape[i + 2] - 1) * stride[i]
            - (1 - 2 * padding[i])
            - dilation[i] * (weight_shape[i + 2] - 1)
            for i in range(ndim)
        ]

    # Forward & backward.
    class Conv2d(torch.autograd.Function):
        @staticmethod
        def forward(ctx, input, weight, bias):
            assert weight.shape == weight_shape
            ctx.save_for_backward(
                input if weight.requires_grad else _null_tensor,
                weight if input.requires_grad else _null_tensor,
            )
            ctx.input_shape = input.shape

            # Simple 1x1 convolution => cuBLAS (only on Volta, not on Ampere).
            if weight_shape[2:] == stride == dilation == (1, 1) and padding == (0, 0) and torch.cuda.get_device_capability(input.device) < (8, 0):
                a = weight.reshape(groups, weight_shape[0] // groups, weight_shape[1])
                b = input.reshape(input.shape[0], groups, input.shape[1] // groups, -1)
                c = (a.transpose(1, 2) if transpose else a) @ b.permute(1, 2, 0, 3).flatten(2)
                c = c.reshape(-1, input.shape[0], *input.shape[2:]).transpose(0, 1)
                c = c if bias is None else c + bias.unsqueeze(0).unsqueeze(2).unsqueeze(3)
                return c.contiguous(memory_format=(torch.channels_last if input.stride(1) == 1 else torch.contiguous_format))

            # General case => cuDNN.
            if transpose:
                return torch.nn.functional.conv_transpose2d(input=input, weight=weight, bias=bias, output_padding=output_padding, **common_kwargs)
            return torch.nn.functional.conv2d(input=input, weight=weight, bias=bias, **common_kwargs)

        @staticmethod
        def backward(ctx, grad_output):
            input, weight = ctx.saved_tensors
            input_shape = ctx.input_shape
            grad_input = None
            grad_weight = None
            grad_bias = None

            if ctx.needs_input_grad[0]:
                p = calc_output_padding(input_shape=input_shape, output_shape=grad_output.shape)
                op = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs)
                grad_input = op.apply(grad_output, weight, None)
                assert grad_input.shape == input_shape

            if ctx.needs_input_grad[1] and not weight_gradients_disabled:
                grad_weight = Conv2dGradWeight.apply(grad_output, input)
                assert grad_weight.shape == weight_shape

            if ctx.needs_input_grad[2]:
                grad_bias = grad_output.sum([0, 2, 3])

            return grad_input, grad_weight, grad_bias

    # Gradient with respect to the weights.
    class Conv2dGradWeight(torch.autograd.Function):
        @staticmethod
        def forward(ctx, grad_output, input):
            ctx.save_for_backward(
                grad_output if input.requires_grad else _null_tensor,
                input if grad_output.requires_grad else _null_tensor,
            )
            ctx.grad_output_shape = grad_output.shape
            ctx.input_shape = input.shape

            # Simple 1x1 convolution => cuBLAS (on both Volta and Ampere).
            if weight_shape[2:] == stride == dilation == (1, 1) and padding == (0, 0):
                a = grad_output.reshape(grad_output.shape[0], groups, grad_output.shape[1] // groups, -1).permute(1, 2, 0, 3).flatten(2)
                b = input.reshape(input.shape[0], groups, input.shape[1] // groups, -1).permute(1, 2, 0, 3).flatten(2)
                c = (b @ a.transpose(1, 2) if transpose else a @ b.transpose(1, 2)).reshape(weight_shape)
                return c.contiguous(memory_format=(torch.channels_last if input.stride(1) == 1 else torch.contiguous_format))

            # General case => cuDNN.
            name = 'aten::cudnn_convolution_transpose_backward_weight' if transpose else 'aten::cudnn_convolution_backward_weight'
            flags = [torch.backends.cudnn.benchmark, torch.backends.cudnn.deterministic, torch.backends.cudnn.allow_tf32]
            return torch._C._jit_get_operation(name)(weight_shape, grad_output, input, padding, stride, dilation, groups, *flags)

        @staticmethod
        def backward(ctx, grad2_grad_weight):
            grad_output, input = ctx.saved_tensors
            grad_output_shape = ctx.grad_output_shape
            input_shape = ctx.input_shape
            grad2_grad_output = None
            grad2_input = None

            if ctx.needs_input_grad[0]:
                grad2_grad_output = Conv2d.apply(input, grad2_grad_weight, None)
                assert grad2_grad_output.shape == grad_output_shape

            if ctx.needs_input_grad[1]:
                p = calc_output_padding(input_shape=input_shape, output_shape=grad_output_shape)
                op = _conv2d_gradfix(transpose=(not transpose), weight_shape=weight_shape, output_padding=p, **common_kwargs)
                grad2_input = op.apply(grad_output, grad2_grad_weight, None)
                assert grad2_input.shape == input_shape

            return grad2_grad_output, grad2_input

    _conv2d_gradfix_cache[key] = Conv2d
    return Conv2d

#----------------------------------------------------------------------------

# pylint: enable=redefined-builtin
# pylint: enable=arguments-differ
# pylint: enable=protected-access
# pylint: enable=line-too-long
# pylint: enable=global-statement
# pylint: enable=missing-class-docstring
# pylint: enable=missing-function-docstring