File size: 8,148 Bytes
3d49622
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import pdb
from os import path
import torch
import torch.distributed as dist
import torch.autograd as autograd
import torch.cuda.comm as comm
from torch.autograd.function import once_differentiable
from torch.utils.cpp_extension import load

_src_path = path.join(path.dirname(path.abspath(__file__)), "src")
_backend = load(name="inplace_abn",
                extra_cflags=["-O3"],
                sources=[path.join(_src_path, f) for f in [
                    "inplace_abn.cpp",
                    "inplace_abn_cpu.cpp",
                    "inplace_abn_cuda.cu",
                    "inplace_abn_cuda_half.cu"
                ]],
                extra_cuda_cflags=["--expt-extended-lambda"])

# Activation names
ACT_RELU = "relu"
ACT_LEAKY_RELU = "leaky_relu"
ACT_ELU = "elu"
ACT_NONE = "none"


def _check(fn, *args, **kwargs):
    success = fn(*args, **kwargs)
    if not success:
        raise RuntimeError("CUDA Error encountered in {}".format(fn))


def _broadcast_shape(x):
    out_size = []
    for i, s in enumerate(x.size()):
        if i != 1:
            out_size.append(1)
        else:
            out_size.append(s)
    return out_size


def _reduce(x):
    if len(x.size()) == 2:
        return x.sum(dim=0)
    else:
        n, c = x.size()[0:2]
        return x.contiguous().view((n, c, -1)).sum(2).sum(0)


def _count_samples(x):
    count = 1
    for i, s in enumerate(x.size()):
        if i != 1:
            count *= s
    return count


def _act_forward(ctx, x):
    if ctx.activation == ACT_LEAKY_RELU:
        _backend.leaky_relu_forward(x, ctx.slope)
    elif ctx.activation == ACT_ELU:
        _backend.elu_forward(x)
    elif ctx.activation == ACT_NONE:
        pass


def _act_backward(ctx, x, dx):
    if ctx.activation == ACT_LEAKY_RELU:
        _backend.leaky_relu_backward(x, dx, ctx.slope)
    elif ctx.activation == ACT_ELU:
        _backend.elu_backward(x, dx)
    elif ctx.activation == ACT_NONE:
        pass


class InPlaceABN(autograd.Function):
    @staticmethod
    def forward(ctx, x, weight, bias, running_mean, running_var,
                training=True, momentum=0.1, eps=1e-05, activation=ACT_LEAKY_RELU, slope=0.01):
        # Save context
        ctx.training = training
        ctx.momentum = momentum
        ctx.eps = eps
        ctx.activation = activation
        ctx.slope = slope
        ctx.affine = weight is not None and bias is not None

        # Prepare inputs
        count = _count_samples(x)
        x = x.contiguous()
        weight = weight.contiguous() if ctx.affine else x.new_empty(0)
        bias = bias.contiguous() if ctx.affine else x.new_empty(0)

        if ctx.training:
            mean, var = _backend.mean_var(x)

            # Update running stats
            running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean)
            running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * var * count / (count - 1))

            # Mark in-place modified tensors
            ctx.mark_dirty(x, running_mean, running_var)
        else:
            mean, var = running_mean.contiguous(), running_var.contiguous()
            ctx.mark_dirty(x)

        # BN forward + activation
        _backend.forward(x, mean, var, weight, bias, ctx.affine, ctx.eps)
        _act_forward(ctx, x)

        # Output
        ctx.var = var
        ctx.save_for_backward(x, var, weight, bias)
        ctx.mark_non_differentiable(running_mean, running_var)
        return x, running_mean, running_var

    @staticmethod
    @once_differentiable
    def backward(ctx, dz, _drunning_mean, _drunning_var):
        z, var, weight, bias = ctx.saved_tensors
        dz = dz.contiguous()

        # Undo activation
        _act_backward(ctx, z, dz)

        if ctx.training:
            edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps)
        else:
            # TODO: implement simplified CUDA backward for inference mode
            edz = dz.new_zeros(dz.size(1))
            eydz = dz.new_zeros(dz.size(1))

        dx = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps)
        # dweight = eydz * weight.sign() if ctx.affine else None
        dweight = eydz if ctx.affine else None
        if dweight is not None:
            dweight[weight < 0] *= -1
        dbias = edz if ctx.affine else None

        return dx, dweight, dbias, None, None, None, None, None, None, None


class InPlaceABNSync(autograd.Function):
    @classmethod
    def forward(cls, ctx, x, weight, bias, running_mean, running_var,
                training=True, momentum=0.1, eps=1e-05, activation=ACT_LEAKY_RELU, slope=0.01, equal_batches=True):
        # Save context
        ctx.training = training
        ctx.momentum = momentum
        ctx.eps = eps
        ctx.activation = activation
        ctx.slope = slope
        ctx.affine = weight is not None and bias is not None

        # Prepare inputs
        ctx.world_size = dist.get_world_size() if dist.is_initialized() else 1

        # count = _count_samples(x)
        batch_size = x.new_tensor([x.shape[0]], dtype=torch.long)

        x = x.contiguous()
        weight = weight.contiguous() if ctx.affine else x.new_empty(0)
        bias = bias.contiguous() if ctx.affine else x.new_empty(0)

        if ctx.training:
            mean, var = _backend.mean_var(x)
            if ctx.world_size > 1:
                # get global batch size
                if equal_batches:
                    batch_size *= ctx.world_size
                else:
                    dist.all_reduce(batch_size, dist.ReduceOp.SUM)

                ctx.factor = x.shape[0] / float(batch_size.item())

                mean_all = mean.clone() * ctx.factor
                dist.all_reduce(mean_all, dist.ReduceOp.SUM)

                var_all = (var + (mean - mean_all) ** 2) * ctx.factor
                dist.all_reduce(var_all, dist.ReduceOp.SUM)

                mean = mean_all
                var = var_all

            # Update running stats
            running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean)
            count = batch_size.item() * x.view(x.shape[0], x.shape[1], -1).shape[-1]
            running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * var * (float(count) / (count - 1)))

            # Mark in-place modified tensors
            ctx.mark_dirty(x, running_mean, running_var)
        else:
            mean, var = running_mean.contiguous(), running_var.contiguous()
            ctx.mark_dirty(x)

        # BN forward + activation
        _backend.forward(x, mean, var, weight, bias, ctx.affine, ctx.eps)
        _act_forward(ctx, x)

        # Output
        ctx.var = var
        ctx.save_for_backward(x, var, weight, bias)
        ctx.mark_non_differentiable(running_mean, running_var)
        return x, running_mean, running_var

    @staticmethod
    @once_differentiable
    def backward(ctx, dz, _drunning_mean, _drunning_var):
        z, var, weight, bias = ctx.saved_tensors
        dz = dz.contiguous()

        # Undo activation
        _act_backward(ctx, z, dz)

        if ctx.training:
            edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps)
            edz_local = edz.clone()
            eydz_local = eydz.clone()

            if ctx.world_size > 1:
                edz *= ctx.factor
                dist.all_reduce(edz, dist.ReduceOp.SUM)

                eydz *= ctx.factor
                dist.all_reduce(eydz, dist.ReduceOp.SUM)
        else:
            edz_local = edz = dz.new_zeros(dz.size(1))
            eydz_local = eydz = dz.new_zeros(dz.size(1))

        dx = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps)
        # dweight = eydz_local * weight.sign() if ctx.affine else None
        dweight = eydz_local if ctx.affine else None
        if dweight is not None:
            dweight[weight < 0] *= -1
        dbias = edz_local if ctx.affine else None

        return dx, dweight, dbias, None, None, None, None, None, None, None


inplace_abn = InPlaceABN.apply
inplace_abn_sync = InPlaceABNSync.apply

__all__ = ["inplace_abn", "inplace_abn_sync", "ACT_RELU", "ACT_LEAKY_RELU", "ACT_ELU", "ACT_NONE"]