/
convolution_2d.py
468 lines (393 loc) · 17.7 KB
/
convolution_2d.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
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
import numpy
import chainer
from chainer.backends import cuda
from chainer import configuration
from chainer import function_node
import chainer.functions
from chainer.utils import argument
from chainer.utils import conv
from chainer.utils import type_check
if cuda.cudnn_enabled:
cudnn = cuda.cudnn
libcudnn = cuda.cuda.cudnn
_cudnn_version = libcudnn.getVersion()
_fwd_pref = libcudnn.CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT
_bwd_filter_pref = \
libcudnn.CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT
_algorithm_fwd = {}
_algorithm_bwd_filter = {}
def _pair(x):
if hasattr(x, '__getitem__'):
return x
return x, x
def _get_algorithm_fwd(
x, W, y, conv_param, handle, x_desc, filter_desc, conv_desc, y_desc,
workspace):
key = (x.shape, W.shape, y.shape, conv_param)
if key in _algorithm_fwd:
return _algorithm_fwd[key]
ret = libcudnn.findConvolutionForwardAlgorithmEx(
handle, x_desc.value, x.data.ptr, filter_desc.value, W.data.ptr,
conv_desc.value, y_desc.value, y.data.ptr, 1, workspace.data.ptr,
workspace.size)
algo = ret[1][0]['algo']
_algorithm_fwd[key] = algo
return algo
def _get_algorithm_bwd_filter(
x, dy, dW, conv_param, handle, x_desc, dy_desc, conv_desc, filter_desc,
workspace):
key = (x.shape, dW.shape, dy.shape, conv_param)
if key in _algorithm_bwd_filter:
return _algorithm_bwd_filter[key]
ret = libcudnn.findConvolutionBackwardFilterAlgorithmEx(
handle, x_desc.value, x.data.ptr, dy_desc.value, dy.data.ptr,
conv_desc.value, filter_desc.value, dW.data.ptr, 1,
workspace.data.ptr, workspace.size)
algo = ret[1][0]['algo']
_algorithm_bwd_filter[key] = algo
return algo
class Convolution2DFunction(function_node.FunctionNode):
def __init__(self, stride=1, pad=0, cover_all=False, **kwargs):
argument.check_unexpected_kwargs(
kwargs,
deterministic="deterministic argument is not supported anymore. "
"Use chainer.using_config('cudnn_deterministic', value) context "
"where value is either `True` or `False`.",
requires_x_grad="requires_x_grad argument is not supported "
"anymore. Just remove the argument. Note that whether to compute "
"the gradient w.r.t. x is automatically decided during "
"backpropagation."
)
argument.assert_kwargs_empty(kwargs)
self.sy, self.sx = _pair(stride)
self.ph, self.pw = _pair(pad)
self.cover_all = cover_all
def check_type_forward(self, in_types):
n_in = in_types.size()
type_check.expect(2 <= n_in, n_in <= 3)
x_type = in_types[0]
w_type = in_types[1]
type_check.expect(
x_type.dtype.kind == 'f',
w_type.dtype.kind == 'f',
x_type.ndim == 4,
w_type.ndim == 4,
x_type.shape[1] == w_type.shape[1],
)
if type_check.eval(n_in) == 3:
b_type = in_types[2]
type_check.expect(
b_type.dtype == x_type.dtype,
b_type.ndim == 1,
b_type.shape[0] == w_type.shape[0],
)
def forward_cpu(self, inputs):
self.retain_inputs((0, 1)) # retain only x and W
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
if not all([isinstance(i, numpy.ndarray) for i in inputs]):
if b is not None:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'
.format(type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'
.format(type(W), type(x)))
kh, kw = W.shape[2:]
col = conv.im2col_cpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
y = numpy.tensordot(
col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
if b is not None:
y += b
return numpy.rollaxis(y, 3, 1),
def forward_gpu(self, inputs):
self.retain_inputs((0, 1)) # retain only x and W
x, W = inputs[:2]
b = inputs[2] if len(inputs) == 3 else None
if not all([isinstance(i, cuda.ndarray) for i in inputs]):
if b is not None:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}, type(b): {2}'
.format(type(W), type(x), type(b)))
else:
raise ValueError('numpy and cupy must not be used together\n'
'type(W): {0}, type(x): {1}'
.format(type(W), type(x)))
out_c, _, kh, kw = W.shape
n, c, h, w = x.shape
out_h = conv.get_conv_outsize(h, kh, self.sy, self.ph,
cover_all=self.cover_all)
assert out_h > 0, 'Height in the output should be positive.'
out_w = conv.get_conv_outsize(w, kw, self.sx, self.pw,
cover_all=self.cover_all)
assert out_w > 0, 'Width in the output should be positive.'
y = cuda.cupy.empty((n, out_c, out_h, out_w), dtype=x.dtype)
if (not self.cover_all and chainer.should_use_cudnn('>=auto') and
x.dtype == W.dtype):
x = cuda.cupy.ascontiguousarray(x)
W = cuda.cupy.ascontiguousarray(W)
if b is not None:
b = cuda.cupy.ascontiguousarray(b)
handle = cudnn.get_handle()
x_desc = cudnn.create_tensor_descriptor(x)
y_desc = cudnn.create_tensor_descriptor(y)
filter_desc = cudnn.create_filter_descriptor(W)
conv_param = ((self.ph, self.pw), (self.sy, self.sx), x.dtype)
conv_desc = cudnn.create_convolution_descriptor(
*conv_param)
if b is not None:
bias_desc = cudnn.create_tensor_descriptor(
b[None, :, None, None])
workspace_size = cuda.get_max_workspace_size()
workspace = cuda.cupy.empty((workspace_size,), dtype='b')
if configuration.config.autotune and _cudnn_version >= 5000:
algo = _get_algorithm_fwd(
x, W, y, conv_param, handle, x_desc,
filter_desc, conv_desc, y_desc, workspace)
else:
algo = libcudnn.getConvolutionForwardAlgorithm(
handle, x_desc.value, filter_desc.value,
conv_desc.value, y_desc.value, _fwd_pref, workspace_size)
oz_dtype = 'd' if x.dtype == 'd' else 'f'
one = numpy.array(1, dtype=oz_dtype).ctypes
zero = numpy.array(0, dtype=oz_dtype).ctypes
libcudnn.convolutionForward(
handle, one.data, x_desc.value, x.data.ptr,
filter_desc.value, W.data.ptr, conv_desc.value,
algo, workspace.data.ptr, workspace_size, zero.data,
y_desc.value, y.data.ptr)
# TODO(beam2d): Support unshared bias
if b is not None:
cudnn.add_tensor(
handle, one.data, bias_desc.value, b.data.ptr,
one.data, y_desc.value, y.data.ptr)
else:
# Implementation using im2col
col = conv.im2col_gpu(
x, kh, kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
y = cuda.cupy.tensordot(
col, W, ((1, 2, 3), (1, 2, 3))).astype(x.dtype, copy=False)
# TODO(beam2d): Support unshared bias
if b is not None:
y += b
y = cuda.cupy.rollaxis(y, 3, 1)
return y,
def backward(self, indexes, grad_outputs):
x, W = self.get_retained_inputs()
gy, = grad_outputs
ret = []
if 0 in indexes:
xh, xw = x.shape[2:]
gx = chainer.functions.deconvolution_2d(
gy, W, stride=(self.sy, self.sx), pad=(self.ph, self.pw),
outsize=(xh, xw))
ret.append(gx)
if 1 in indexes:
gW, = Convolution2DGradW(self).apply((x, gy))
ret.append(gW)
if 2 in indexes:
gb = chainer.functions.sum(gy, axis=(0, 2, 3))
ret.append(gb)
return ret
class Convolution2DGradW(function_node.FunctionNode):
def __init__(self, conv2d):
W_node = conv2d.inputs[1]
self.kh, self.kw = W_node.shape[2:]
self.sy = conv2d.sy
self.sx = conv2d.sx
self.ph = conv2d.ph
self.pw = conv2d.pw
self.cover_all = conv2d.cover_all
self.W_dtype = W_node.dtype
def forward_cpu(self, inputs):
self.retain_inputs((0, 1))
x, gy = inputs
col = conv.im2col_cpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
# NumPy raises an error when the array is not contiguous.
# See: https://github.com/chainer/chainer/issues/2744
# TODO(niboshi): Remove this code when NumPy is fixed.
if (not (gy.flags.c_contiguous or gy.flags.f_contiguous) and
1 in gy.shape):
gy = numpy.ascontiguousarray(gy)
gW = numpy.tensordot(
gy, col, ((0, 2, 3), (0, 4, 5))).astype(self.W_dtype, copy=False)
return gW,
def forward_gpu(self, inputs):
self.retain_inputs((0, 1))
x, gy = inputs
_, out_c, out_h, out_w = gy.shape
n, c, h, w = x.shape
if (self.cover_all or not chainer.should_use_cudnn('>=auto') or
x.dtype != self.W_dtype):
col = conv.im2col_gpu(
x, self.kh, self.kw, self.sy, self.sx, self.ph, self.pw,
cover_all=self.cover_all)
gW = cuda.cupy.tensordot(
gy, col, ((0, 2, 3), (0, 4, 5))).astype(self.W_dtype,
copy=False)
return gW,
gW = cuda.cupy.empty((out_c, c, self.kh, self.kw), dtype=self.W_dtype)
x = cuda.cupy.ascontiguousarray(x)
gy = cuda.cupy.ascontiguousarray(gy)
handle = cudnn.get_handle()
x_desc = cudnn.create_tensor_descriptor(x)
gy_desc = cudnn.create_tensor_descriptor(gy)
filter_desc = cudnn.create_filter_descriptor(gW)
conv_param = (self.ph, self.pw), (self.sy, self.sx), x.dtype
conv_desc = cudnn.create_convolution_descriptor(
*conv_param)
oz_dtype = 'd' if x.dtype == 'd' else 'f'
one = numpy.array(1, dtype=oz_dtype).ctypes
zero = numpy.array(0, dtype=oz_dtype).ctypes
workspace_size = cuda.get_max_workspace_size()
workspace = cuda.cupy.empty((workspace_size,), dtype='b')
if configuration.config.cudnn_deterministic:
algo = libcudnn.CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1
elif configuration.config.autotune and _cudnn_version >= 5000:
algo = _get_algorithm_bwd_filter(
x, gy, gW, conv_param, handle, x_desc, gy_desc,
conv_desc, filter_desc, workspace)
else:
algo = libcudnn.getConvolutionBackwardFilterAlgorithm(
handle, x_desc.value, gy_desc.value, conv_desc.value,
filter_desc.value, _bwd_filter_pref, workspace_size)
libcudnn.convolutionBackwardFilter_v3(
handle, one.data, x_desc.value, x.data.ptr, gy_desc.value,
gy.data.ptr, conv_desc.value, algo, workspace.data.ptr,
workspace_size, zero.data, filter_desc.value, gW.data.ptr)
return gW,
def backward(self, indexes, grad_outputs):
x, gy = self.get_retained_inputs()
ggW, = grad_outputs
ret = []
if 0 in indexes:
xh, xw = x.shape[2:]
gx = chainer.functions.deconvolution_2d(
gy, ggW, stride=(self.sy, self.sx), pad=(self.ph, self.pw),
outsize=(xh, xw))
ret.append(gx)
if 1 in indexes:
ggy = convolution_2d(
x, ggW, stride=(self.sy, self.sx), pad=(self.ph, self.pw),
cover_all=self.cover_all)
ret.append(ggy)
return ret
def convolution_2d(x, W, b=None, stride=1, pad=0, cover_all=False, **kwargs):
"""convolution_2d(x, W, b=None, stride=1, pad=0, cover_all=False)
Two-dimensional convolution function.
This is an implementation of two-dimensional convolution in ConvNets.
It takes three variables: the input image ``x``, the filter weight ``W``,
and the bias vector ``b``.
Notation: here is a notation for dimensionalities.
- :math:`n` is the batch size.
- :math:`c_I` and :math:`c_O` are the number of the input and output
channels, respectively.
- :math:`h_I` and :math:`w_I` are the height and width of the input image,
respectively.
- :math:`h_K` and :math:`w_K` are the height and width of the filters,
respectively.
- :math:`h_P` and :math:`w_P` are the height and width of the spatial
padding size, respectively.
Then the ``Convolution2D`` function computes correlations between filters
and patches of size :math:`(h_K, w_K)` in ``x``.
Note that correlation here is equivalent to the inner product between
expanded vectors.
Patches are extracted at positions shifted by multiples of ``stride`` from
the first position ``(-h_P, -w_P)`` for each spatial axis.
The right-most (or bottom-most) patches do not run over the padded spatial
size.
Let :math:`(s_Y, s_X)` be the stride of filter application. Then, the
output size :math:`(h_O, w_O)` is determined by the following equations:
.. math::
h_O &= (h_I + 2h_P - h_K) / s_Y + 1,\\\\
w_O &= (w_I + 2w_P - w_K) / s_X + 1.
If ``cover_all`` option is ``True``, the filter will cover the all
spatial locations. So, if the last stride of filter does not cover the
end of spatial locations, an addtional stride will be applied to the end
part of spatial locations. In this case, the output size :math:`(h_O, w_O)`
is determined by the following equations:
.. math::
h_O &= (h_I + 2h_P - h_K + s_Y - 1) / s_Y + 1,\\\\
w_O &= (w_I + 2w_P - w_K + s_X - 1) / s_X + 1.
If the bias vector is given, then it is added to all spatial locations of
the output of convolution.
The output of this function can be non-deterministic when it uses cuDNN.
If ``chainer.configuration.config.cudnn_deterministic`` is ``True`` and
cuDNN version is >= v3, it forces cuDNN to use a deterministic algorithm.
Convolution links can use a feature of cuDNN called autotuning, which
selects the most efficient CNN algorithm for images of fixed-size,
can provide a significant performance boost for fixed neural nets.
To enable, set `chainer.using_config('autotune', True)`
.. warning::
``deterministic`` argument is not supported anymore since v2.
Instead, use ``chainer.using_config('cudnn_deterministic', value)``
(value is either ``True`` or ``False``).
See :func:`chainer.using_config`.
Args:
x (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
Input variable of shape :math:`(n, c_I, h_I, w_I)`.
W (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`):
Weight variable of shape :math:`(c_O, c_I, h_K, w_K)`.
b (:class:`~chainer.Variable` or :class:`numpy.ndarray` or \
:class:`cupy.ndarray`): Bias variable of length :math:`c_O` (optional).
stride (:class:`int` or pair of :class:`int` s):
Stride of filter applications. ``stride=s`` and ``stride=(s, s)``
are equivalent.
pad (:class:`int` or pair of :class:`int` s):
Spatial padding width for input arrays.
``pad=p`` and ``pad=(p, p)`` are equivalent.
cover_all (bool): If ``True``, all spatial locations are convoluted
into some output pixels.
Returns:
~chainer.Variable:
Output variable of shape :math:`(n, c_O, h_O, w_O)`.
.. seealso:: :class:`~chainer.links.Convolution2D`
.. admonition:: Example
>>> n = 10
>>> c_i, c_o = 3, 1
>>> h_i, w_i = 30, 40
>>> h_k, w_k = 10, 10
>>> h_p, w_p = 5, 5
>>> x = np.random.uniform(0, 1, (n, c_i, h_i, w_i)).astype('f')
>>> x.shape
(10, 3, 30, 40)
>>> W = np.random.uniform(0, 1, (c_o, c_i, h_k, w_k)).astype('f')
>>> W.shape
(1, 3, 10, 10)
>>> b = np.random.uniform(0, 1, (c_o,)).astype('f')
>>> b.shape
(1,)
>>> s_y, s_x = 5, 7
>>> y = F.convolution_2d(x, W, b, stride=(s_y, s_x), pad=(h_p, w_p))
>>> y.shape
(10, 1, 7, 6)
>>> h_o = int((h_i + 2 * h_p - h_k) / s_y + 1)
>>> w_o = int((w_i + 2 * w_p - w_k) / s_x + 1)
>>> y.shape == (n, c_o, h_o, w_o)
True
>>> y = F.convolution_2d(x, W, b, stride=(s_y, s_x), pad=(h_p, w_p), \
cover_all=True)
>>> y.shape == (n, c_o, h_o, w_o + 1)
True
"""
argument.check_unexpected_kwargs(
kwargs, deterministic="deterministic argument is not "
"supported anymore. "
"Use chainer.using_config('cudnn_deterministic', value) "
"context where value is either `True` or `False`.")
argument.assert_kwargs_empty(kwargs)
fnode = Convolution2DFunction(stride, pad, cover_all)
if b is None:
args = x, W
else:
args = x, W, b
y, = fnode.apply(args)
return y