File size: 11,362 Bytes
11e6f7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
# SPDX-FileCopyrightText: Copyright (c) 2021-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NvidiaProprietary
#
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
# property and proprietary rights in and to this material, related
# documentation and any modifications thereto. Any use, reproduction,
# disclosure or distribution of this material and related documentation
# without an express license agreement from NVIDIA CORPORATION or
# its affiliates is strictly prohibited.
"""Custom PyTorch ops for efficient bias and activation."""

import os
import numpy as np
import torch
import dnnlib

from .. import custom_ops
from .. import misc

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

activation_funcs = {
    'linear':
    dnnlib.EasyDict(func=lambda x, **_: x,
                    def_alpha=0,
                    def_gain=1,
                    cuda_idx=1,
                    ref='',
                    has_2nd_grad=False),
    'relu':
    dnnlib.EasyDict(func=lambda x, **_: torch.nn.functional.relu(x),
                    def_alpha=0,
                    def_gain=np.sqrt(2),
                    cuda_idx=2,
                    ref='y',
                    has_2nd_grad=False),
    'lrelu':
    dnnlib.EasyDict(
        func=lambda x, alpha, **_: torch.nn.functional.leaky_relu(x, alpha),
        def_alpha=0.2,
        def_gain=np.sqrt(2),
        cuda_idx=3,
        ref='y',
        has_2nd_grad=False),
    'tanh':
    dnnlib.EasyDict(func=lambda x, **_: torch.tanh(x),
                    def_alpha=0,
                    def_gain=1,
                    cuda_idx=4,
                    ref='y',
                    has_2nd_grad=True),
    'sigmoid':
    dnnlib.EasyDict(func=lambda x, **_: torch.sigmoid(x),
                    def_alpha=0,
                    def_gain=1,
                    cuda_idx=5,
                    ref='y',
                    has_2nd_grad=True),
    'elu':
    dnnlib.EasyDict(func=lambda x, **_: torch.nn.functional.elu(x),
                    def_alpha=0,
                    def_gain=1,
                    cuda_idx=6,
                    ref='y',
                    has_2nd_grad=True),
    'selu':
    dnnlib.EasyDict(func=lambda x, **_: torch.nn.functional.selu(x),
                    def_alpha=0,
                    def_gain=1,
                    cuda_idx=7,
                    ref='y',
                    has_2nd_grad=True),
    'softplus':
    dnnlib.EasyDict(func=lambda x, **_: torch.nn.functional.softplus(x),
                    def_alpha=0,
                    def_gain=1,
                    cuda_idx=8,
                    ref='y',
                    has_2nd_grad=True),
    'swish':
    dnnlib.EasyDict(func=lambda x, **_: torch.sigmoid(x) * x,
                    def_alpha=0,
                    def_gain=np.sqrt(2),
                    cuda_idx=9,
                    ref='x',
                    has_2nd_grad=True),
}

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

_plugin = None
_null_tensor = torch.empty([0])


def _init():
    global _plugin
    if _plugin is None:
        _plugin = custom_ops.get_plugin(
            module_name='bias_act_plugin',
            sources=['bias_act.cpp', 'bias_act.cu'],
            headers=['bias_act.h'],
            source_dir=os.path.dirname(__file__),
            extra_cuda_cflags=['--use_fast_math'],
        )
    return True


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


# @torch.autocast(device_type='cuda')
def bias_act(x,
             b=None,
             dim=1,
             act='linear',
             alpha=None,
             gain=None,
             clamp=None,
             impl='cuda'):
    r"""Fused bias and activation function.

    Adds bias `b` to activation tensor `x`, evaluates activation function `act`,
    and scales the result by `gain`. Each of the steps is optional. In most cases,
    the fused op is considerably more efficient than performing the same calculation
    using standard PyTorch ops. It supports first and second order gradients,
    but not third order gradients.

    Args:
        x:      Input activation tensor. Can be of any shape.
        b:      Bias vector, or `None` to disable. Must be a 1D tensor of the same type
                as `x`. The shape must be known, and it must match the dimension of `x`
                corresponding to `dim`.
        dim:    The dimension in `x` corresponding to the elements of `b`.
                The value of `dim` is ignored if `b` is not specified.
        act:    Name of the activation function to evaluate, or `"linear"` to disable.
                Can be e.g. `"relu"`, `"lrelu"`, `"tanh"`, `"sigmoid"`, `"swish"`, etc.
                See `activation_funcs` for a full list. `None` is not allowed.
        alpha:  Shape parameter for the activation function, or `None` to use the default.
        gain:   Scaling factor for the output tensor, or `None` to use default.
                See `activation_funcs` for the default scaling of each activation function.
                If unsure, consider specifying 1.
        clamp:  Clamp the output values to `[-clamp, +clamp]`, or `None` to disable
                the clamping (default).
        impl:   Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).

    Returns:
        Tensor of the same shape and datatype as `x`.
    """
    assert isinstance(x, torch.Tensor)
    assert impl in ['ref', 'cuda']
    if impl == 'cuda' and x.device.type == 'cuda' and _init():
        return _bias_act_cuda(dim=dim,
                              act=act,
                              alpha=alpha,
                              gain=gain,
                              clamp=clamp).apply(x, b)
    return _bias_act_ref(x=x,
                         b=b,
                         dim=dim,
                         act=act,
                         alpha=alpha,
                         gain=gain,
                         clamp=clamp)


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


@misc.profiled_function
def _bias_act_ref(x,
                  b=None,
                  dim=1,
                  act='linear',
                  alpha=None,
                  gain=None,
                  clamp=None):
    """Slow reference implementation of `bias_act()` using standard TensorFlow ops.
    """
    assert isinstance(x, torch.Tensor)
    assert clamp is None or clamp >= 0
    spec = activation_funcs[act]
    alpha = float(alpha if alpha is not None else spec.def_alpha)
    gain = float(gain if gain is not None else spec.def_gain)
    clamp = float(clamp if clamp is not None else -1)

    # Add bias.
    if b is not None:
        assert isinstance(b, torch.Tensor) and b.ndim == 1
        assert 0 <= dim < x.ndim
        assert b.shape[0] == x.shape[dim]
        x = x + b.reshape([-1 if i == dim else 1 for i in range(x.ndim)])

    # Evaluate activation function.
    alpha = float(alpha)
    x = spec.func(x, alpha=alpha)

    # Scale by gain.
    gain = float(gain)
    if gain != 1:
        x = x * gain

    # Clamp.
    if clamp >= 0:
        x = x.clamp(-clamp, clamp)  # pylint: disable=invalid-unary-operand-type
    return x


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

_bias_act_cuda_cache = dict()


# @torch.autocast(device_type='cuda')
def _bias_act_cuda(dim=1, act='linear', alpha=None, gain=None, clamp=None):
    """Fast CUDA implementation of `bias_act()` using custom ops.
    """
    # Parse arguments.
    assert clamp is None or clamp >= 0
    spec = activation_funcs[act]
    alpha = float(alpha if alpha is not None else spec.def_alpha)
    gain = float(gain if gain is not None else spec.def_gain)
    clamp = float(clamp if clamp is not None else -1)

    # Lookup from cache.
    key = (dim, act, alpha, gain, clamp)
    if key in _bias_act_cuda_cache:
        return _bias_act_cuda_cache[key]

    # Forward op.
    class BiasActCuda(torch.autograd.Function):
        @staticmethod
        # @torch.cuda.amp.custom_fwd
        def forward(ctx, x, b):  # pylint: disable=arguments-differ
            ctx.memory_format = torch.channels_last if x.ndim > 2 and x.stride(
                1) == 1 else torch.contiguous_format
            x = x.contiguous(memory_format=ctx.memory_format)
            b = b.contiguous() if b is not None else _null_tensor
            y = x
            if act != 'linear' or gain != 1 or clamp >= 0 or b is not _null_tensor:
                y = _plugin.bias_act(x, b, _null_tensor, _null_tensor,
                                     _null_tensor, 0, dim, spec.cuda_idx,
                                     alpha, gain, clamp)
            ctx.save_for_backward(
                x if 'x' in spec.ref or spec.has_2nd_grad else _null_tensor,
                b if 'x' in spec.ref or spec.has_2nd_grad else _null_tensor,
                y if 'y' in spec.ref else _null_tensor)
            return y

        @staticmethod
        # @torch.cuda.amp.custom_bwd
        def backward(ctx, dy):  # pylint: disable=arguments-differ
            dy = dy.contiguous(memory_format=ctx.memory_format)
            x, b, y = ctx.saved_tensors
            dx = None
            db = None

            if ctx.needs_input_grad[0] or ctx.needs_input_grad[1]:
                dx = dy
                if act != 'linear' or gain != 1 or clamp >= 0:
                    dx = BiasActCudaGrad.apply(dy, x, b, y)

            if ctx.needs_input_grad[1]:
                db = dx.sum([i for i in range(dx.ndim) if i != dim])

            return dx, db

    # Backward op.
    class BiasActCudaGrad(torch.autograd.Function):
        @staticmethod
        # @torch.cuda.amp.custom_fwd
        def forward(ctx, dy, x, b, y):  # pylint: disable=arguments-differ
            ctx.memory_format = torch.channels_last if dy.ndim > 2 and dy.stride(
                1) == 1 else torch.contiguous_format
            dx = _plugin.bias_act(dy, b, x, y, _null_tensor, 1, dim,
                                  spec.cuda_idx, alpha, gain, clamp)
            ctx.save_for_backward(dy if spec.has_2nd_grad else _null_tensor, x,
                                  b, y)
            return dx

        @staticmethod
        # @torch.cuda.amp.custom_bwd
        def backward(ctx, d_dx):  # pylint: disable=arguments-differ
            d_dx = d_dx.contiguous(memory_format=ctx.memory_format)
            dy, x, b, y = ctx.saved_tensors
            d_dy = None
            d_x = None
            d_b = None
            d_y = None

            if ctx.needs_input_grad[0]:
                d_dy = BiasActCudaGrad.apply(d_dx, x, b, y)

            if spec.has_2nd_grad and (ctx.needs_input_grad[1]
                                      or ctx.needs_input_grad[2]):
                d_x = _plugin.bias_act(d_dx, b, x, y, dy, 2, dim,
                                       spec.cuda_idx, alpha, gain, clamp)

            if spec.has_2nd_grad and ctx.needs_input_grad[2]:
                d_b = d_x.sum([i for i in range(d_x.ndim) if i != dim])

            return d_dy, d_x, d_b, d_y

    # Add to cache.
    _bias_act_cuda_cache[key] = BiasActCuda
    return BiasActCuda


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