-
Notifications
You must be signed in to change notification settings - Fork 1.4k
/
softmax_cross_entropy.py
298 lines (265 loc) · 12 KB
/
softmax_cross_entropy.py
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
import numpy
import six
import chainer
from chainer import cuda
from chainer import function
from chainer.functions.activation import log_softmax
from chainer.utils import type_check
from chainer import variable
def _broadcast_to(array, shape):
if hasattr(numpy, "broadcast_to"):
return numpy.broadcast_to(array, shape)
dummy = numpy.empty(shape, array.dtype)
return numpy.broadcast_arrays(array, dummy)[0]
class SoftmaxCrossEntropy(function.Function):
"""Softmax activation followed by a cross entropy loss."""
normalize = True
def __init__(self, normalize=True, cache_score=True, class_weight=None,
ignore_label=-1, reduce='mean'):
self.normalize = normalize
self.cache_score = cache_score
self.class_weight = class_weight
if class_weight is not None:
if self.class_weight.ndim != 1:
raise ValueError('class_weight.ndim should be 1')
if self.class_weight.dtype.kind != 'f':
raise ValueError('The dtype of class_weight should be \'f\'')
if isinstance(self.class_weight, variable.Variable):
raise ValueError('class_weight should be a numpy.ndarray or '
'cupy.ndarray, not a chainer.Variable')
self.ignore_label = ignore_label
if reduce not in ('mean', 'no'):
raise ValueError(
"only 'mean' and 'no' are valid for 'reduce', but '%s' is "
'given' % reduce)
self.reduce = reduce
def check_type_forward(self, in_types):
type_check.expect(in_types.size() == 2)
x_type, t_type = in_types
type_check.expect(
x_type.dtype.kind == 'f',
t_type.dtype == numpy.int32,
t_type.ndim == x_type.ndim - 1,
x_type.shape[0] == t_type.shape[0],
x_type.shape[2:] == t_type.shape[1:],
)
def _check_input_values(self, x, t):
if not (((0 <= t) &
(t < x.shape[1])) |
(t == self.ignore_label)).all():
msg = ('Each label `t` need to satisfy '
'`0 <= t < x.shape[1] or t == %d`' % self.ignore_label)
raise ValueError(msg)
def forward_cpu(self, inputs):
x, t = inputs
if chainer.is_debug():
self._check_input_values(x, t)
log_y = log_softmax._log_softmax(x)
if self.cache_score:
self.y = numpy.exp(log_y)
if self.class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
log_y *= _broadcast_to(self.class_weight.reshape(shape), x.shape)
log_yd = numpy.rollaxis(log_y, 1)
log_yd = log_yd.reshape(len(log_yd), -1)
log_p = log_yd[numpy.maximum(t.ravel(), 0), numpy.arange(t.size)]
log_p *= (t.ravel() != self.ignore_label)
if self.reduce == 'mean':
# deal with the case where the SoftmaxCrossEntropy is
# unpickled from the old version
if self.normalize:
count = (t != self.ignore_label).sum()
else:
count = len(x)
self._coeff = 1.0 / max(count, 1)
y = log_p.sum(keepdims=True) * (-self._coeff)
return y.reshape(()),
else:
return -log_p.reshape(t.shape),
def forward_gpu(self, inputs):
cupy = cuda.cupy
x, t = inputs
if chainer.is_debug():
self._check_input_values(x, t)
log_y = log_softmax._log_softmax(x)
if self.cache_score:
self.y = cupy.exp(log_y)
if self.class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
log_y *= cupy.broadcast_to(
self.class_weight.reshape(shape), x.shape)
if self.normalize:
coeff = cupy.maximum(1, (t != self.ignore_label).sum())
else:
coeff = max(1, len(t))
self._coeff = cupy.divide(1.0, coeff, dtype=x.dtype)
log_y = cupy.rollaxis(log_y, 1, log_y.ndim)
if self.reduce == 'mean':
ret = cuda.reduce(
'S t, raw T log_y, int32 n_channel, raw T coeff, '
'S ignore_label',
'T out',
't == ignore_label ? T(0) : log_y[_j * n_channel + t]',
'a + b', 'out = a * -coeff[0]', '0', 'crossent_fwd'
)(t, log_y.reduced_view(), log_y.shape[-1],
self._coeff, self.ignore_label)
else:
ret = cuda.elementwise(
'S t, raw T log_y, int32 n_channel, T ignore', 'T out',
'''
if (t == ignore) {
out = 0;
} else {
out = -log_y[i * n_channel + t];
}
''',
'softmax_crossent_no_reduce_fwd'
)(t, log_y.reduced_view(), log_y.shape[-1], self.ignore_label)
ret = ret.reshape(t.shape)
return ret,
def backward_cpu(self, inputs, grad_outputs):
x, t = inputs
gloss = grad_outputs[0]
if hasattr(self, 'y'):
y = self.y.copy()
else:
y = log_softmax._log_softmax(x)
numpy.exp(y, out=y)
if y.ndim == 2:
gx = y
gx[numpy.arange(len(t)), numpy.maximum(t, 0)] -= 1
if self.class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
c = _broadcast_to(self.class_weight.reshape(shape), x.shape)
c = c[numpy.arange(len(t)), numpy.maximum(t, 0)]
gx *= _broadcast_to(numpy.expand_dims(c, 1), gx.shape)
gx *= (t != self.ignore_label).reshape((len(t), 1))
else:
# in the case where y.ndim is higher than 2,
# we think that a current implementation is inefficient
# because it yields two provisional arrays for indexing.
n_unit = t.size // len(t)
gx = y.reshape(y.shape[0], y.shape[1], -1)
fst_index = numpy.arange(t.size) // n_unit
trd_index = numpy.arange(t.size) % n_unit
gx[fst_index, numpy.maximum(t.ravel(), 0), trd_index] -= 1
if self.class_weight is not None:
shape = [1 if d != 1 else -1 for d in six.moves.range(x.ndim)]
c = _broadcast_to(self.class_weight.reshape(shape), x.shape)
c = c.reshape(gx.shape)
c = c[fst_index, numpy.maximum(t.ravel(), 0), trd_index]
c = c.reshape(y.shape[0], 1, -1)
gx *= _broadcast_to(c, gx.shape)
gx *= (t != self.ignore_label).reshape((len(t), 1, -1))
gx = gx.reshape(y.shape)
if self.reduce == 'mean':
gx *= gloss * self._coeff
else:
gx *= gloss[:, None]
return gx, None
def backward_gpu(self, inputs, grad_outputs):
cupy = cuda.cupy
x, t = inputs
if hasattr(self, 'y'):
y = self.y
else:
y = log_softmax._log_softmax(x)
cupy.exp(y, out=y)
gloss = grad_outputs[0]
n_unit = t.size // len(t)
if self.reduce == 'mean':
coeff = gloss * self._coeff
else:
coeff = gloss[:, None, ...]
if self.class_weight is None:
gx = cuda.elementwise(
'T y, S t, T coeff, S n_channel, S n_unit, S ignore_label',
'T gx',
'''
const int c = (i / n_unit % n_channel);
gx = t == ignore_label ? 0 : coeff * (y - (c == t));
''',
'softmax_crossent_bwd')(
y, cupy.expand_dims(t, 1), coeff, x.shape[1],
n_unit, self.ignore_label)
else:
gx = cuda.elementwise(
'T y, raw T w, S t, T coeff, S n_channel, S n_unit, '
'S ignore_label',
'T gx',
'''
const int c = (i / n_unit % n_channel);
gx = t == ignore_label ? 0 : coeff * (y - (c == t)) * w[t];
''',
'softmax_crossent_weight_bwd')(
y, self.class_weight, cupy.expand_dims(t, 1), coeff,
x.shape[1], n_unit, self.ignore_label)
return gx, None
def softmax_cross_entropy(
x, t, normalize=True, cache_score=True, class_weight=None,
ignore_label=-1, reduce='mean'):
"""Computes cross entropy loss for pre-softmax activations.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
Variable holding a multidimensional array whose element indicates
unnormalized log probability: the first axis of the variable
represents the number of samples, and the second axis represents
the number of classes. While this function computes a usual softmax
cross entropy if the number of dimensions is equal to 2, it
computes a cross entropy of the replicated softmax if the number of
dimensions is greater than 2.
t (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
Variable holding an :class:`numpy.int32` vector of ground truth
labels. If ``t[i] == ignore_label``, corresponding ``x[i]`` is
ignored.
normalize (bool): If ``True``, this function normalizes the cross
entropy loss across all instances. If ``False``, it only
normalizes along a batch size.
cache_score (bool): When it is ``True``, the function stores result
of forward computation to use it on backward computation. It
reduces computational cost though consumes more memory.
class_weight (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
An array that contains constant weights that will be multiplied
with the loss values along with the second dimension. The shape of
this array should be ``(x.shape[1],)``. If this is not ``None``,
each class weight ``class_weight[i]`` is actually multiplied to
``y[:, i]`` that is the corresponding log-softmax output of ``x``
and has the same shape as ``x`` before calculating the actual loss
value.
ignore_label (int): Label value you want to ignore. Its default value
is ``-1``. See description of the argument `t`.
reduce (str): A string that determines whether to reduce the loss
values. If it is ``'mean'``, it computes the sum of the individual
cross entropy and normalize it according to ``normalize`` option.
If it is ``'no'``, this function computes cross entropy for each
instance and does not normalize it (``normalize`` option is
ignored). In this case, the loss value of the ignored instance,
which has ``ignore_label`` as its target value, is set to ``0``.
Returns:
~chainer.Variable: A variable holding a scalar array of the cross
entropy loss. If ``reduce`` is ``'mean'``, it is a scalar array.
If ``reduce`` is ``'no'``, the shape is same as that of ``x``.
.. note::
This function is differentiable only by ``x``.
.. admonition:: Example
>>> x = np.array([[-1, 0, 1, 2], [2, 0, 1, -1]]).astype('f')
>>> x
array([[-1., 0., 1., 2.],
[ 2., 0., 1., -1.]], dtype=float32)
>>> t = np.array([3, 0]).astype('i')
>>> t
array([3, 0], dtype=int32)
>>> y = F.softmax_cross_entropy(x, t)
>>> y
variable(0.4401897192001343)
>>> log_softmax = -F.log_softmax(x)
>>> expected_loss = np.mean([log_softmax[row, column].data \
for row, column in enumerate(t)])
>>> y.array == expected_loss
True
"""
return SoftmaxCrossEntropy(
normalize, cache_score, class_weight, ignore_label, reduce)(x, t)