def _im2col_gpu(img, kernel_size, stride, pad):
"""im2col function for GPU.
This code is ported from Chainer:
https://github.com/chainer/chainer/blob/v6.4.0/chainer/utils/conv.py
"""
n, c, h, w = img.shape
kh, kw = pair(kernel_size)
sy, sx = pair(stride)
ph, pw = pair(pad)
out_h = get_conv_outsize(h, kh, sy, ph)
out_w = get_conv_outsize(w, kw, sx, pw)
dy, dx = 1, 1
col = cuda.cupy.empty((n, c, kh, kw, out_h, out_w), dtype=img.dtype)
cuda.cupy.ElementwiseKernel(
'raw T img, int32 h, int32 w, int32 out_h, int32 out_w,'
'int32 kh, int32 kw, int32 sy, int32 sx, int32 ph, int32 pw,'
'int32 dy, int32 dx', 'T col', '''
int c0 = i / (kh * kw * out_h * out_w);
int ky = i / (kw * out_h * out_w) % kh;
int kx = i / (out_h * out_w) % kw;
int out_y = i / out_w % out_h;
int out_x = i % out_w;
int in_y = ky * dy + out_y * sy - ph;
int in_x = kx * dx + out_x * sx - pw;
if (in_y >= 0 && in_y < h && in_x >= 0 && in_x < w) {
col = img[in_x + w * (in_y + h * c0)];
} else {
col = 0;
}
''', 'im2col')(img.reduced_view(), h, w, out_h, out_w, kh, kw, sy, sx,
ph, pw, dy, dx, col)
return col
def im2col_array(img, kernel_size, stride, pad, to_matrix=True):
N, C, H, W = img.shape
KH, KW = pair(kernel_size)
SH, SW = pair(stride)
PH, PW = pair(pad)
OH = get_conv_outsize(H, KH, SH, PH)
OW = get_conv_outsize(W, KW, SW, PW)
img = np.pad(img, ((0, 0), (0, 0), (PH, PH + SH - 1), (PW, PW + SW - 1)),
mode='constant',
constant_values=(0, ))
col = np.ndarray((N, C, KH, KW, OH, OW), dtype=img.dtype)
# fig 57-1 mini-batch verion
for j in range(KH):
j_lim = j + SH * OH
for i in range(KW):
i_lim = i + SW * OW
# assign data applied by kernel at j, i.
# https://qiita.com/jun40vn/items/d2e8711cabc9cfb1e0d5
# added batch dim and channel dim
col[:, :, j, i, :, :] = img[:, :, j:j_lim:SH, i:i_lim:SW]
# reshape method
if to_matrix:
# fig 57-1
col = col.transpose((0, 4, 5, 1, 2, 3)).reshape((N * OH * OW, -1))
return col
def im2col_array(img, kernel_size, stride, pad, to_matrix=True):
N, C, H, W = img.shape
KH, KW = pair(kernel_size)
SH, SW = pair(stride)
PH, PW = pair(pad)
OH = get_conv_outsize(H, KH, SH, PH)
OW = get_conv_outsize(W, KW, SW, PW)
xp = cuda.get_array_module(img)
if xp != np:
col = _im2col_gpu(img, kernel_size, stride, pad)
else:
img = np.pad(img,
((0, 0), (0, 0), (PH, PH + SH - 1), (PW, PW + SW - 1)),
mode='constant',
constant_values=(0, ))
col = np.ndarray((N, C, KH, KW, OH, OW), dtype=img.dtype)
for j in range(KH):
j_lim = j + SH * OH
for i in range(KW):
i_lim = i + SW * OW
col[:, :, j, i, :, :] = img[:, :, j:j_lim:SH, i:i_lim:SW]
if to_matrix:
col = col.transpose((0, 4, 5, 1, 2, 3)).reshape((N * OH * OW, -1))
return col
def pooling_simple(x, kernel_size, stride=1, pad=0):
x = as_variable(x)
N, C, H, W = x.shape
KH, KW = pair(kernel_size)
PH, PW = pair(pad)
SH, SW = pair(stride)
OH = get_conv_outsize(H, KH, SH, PH)
OW = get_conv_outsize(W, KW, SW, PW)
col = im2col(x, kernel_size, stride, pad, to_matrix=True)
col = col.reshape(-1, KH * KW)
y = col.max(axis=1)
y = y.reshape(N, OH, OW, C).transpose(0, 3, 1, 2)
return y
def conv2d_simple(x, W, b=None, stride=1, pad=0):
x, W = as_variable(x), as_variable(W)
Weight = W
N, C, H, W = x.shape
OC, C, KH, KW = Weight.shape
SH, SW = pair(stride)
PH, PW = pair(pad)
OH = get_conv_outsize(H, KH, SH, PH)
OW = get_conv_outsize(W, KW, SW, PW)
col = im2col(x, (KH, KW), stride, pad, to_matrix=True)
Weight = Weight.reshape(OC, -1).transpose()
t = linear(col, Weight, b)
y = t.reshape(N, OH, OW, OC).transpose(0, 3, 1, 2)
return y
def conv2d_simple(x, W, b=None, stride=1, pad=0):
x, W = as_variable(x), as_variable(W)
n, c, h, w = x.shape
out_c, c, kh, kw = W.shape
sh, sw = _pair(stride)
ph, pw = _pair(pad)
out_h = utils.get_conv_outsize(h, kh, sh, ph)
out_w = utils.get_conv_outsize(w, kw, sw, pw)
col = im2col(x, (kh, kw), stride, pad)
col = col.transpose((0, 4, 5, 1, 2, 3)).reshape((n * out_h * out_w, -1))
W = W.reshape((out_c, -1)).transpose()
t = linear(col, W, b)
y = t.reshape((n, out_h, out_w, -1)).transpose((0, 3, 1, 2))
return y
def col2im_array(col, img_shape, kernel_size, stride, pad, to_matrix=True):
N, C, H, W = img_shape
KH, KW = pair(kernel_size)
SH, SW = pair(stride)
PH, PW = pair(pad)
OH = get_conv_outsize(H, KH, SH, PH)
OW = get_conv_outsize(W, KW, SW, PW)
if to_matrix:
col = col.reshape(N, OH, OW, C, KH, KW).transpose(0, 3, 4, 5, 1, 2)
img = np.zeros((N, C, H + 2 * PH + SH - 1, W + 2 * PW + SW - 1),
dtype=col.dtype)
for j in range(KH):
j_lim = j + SH * OH
for i in range(KW):
i_lim = i + SW * OW
img[:, :, j:j_lim:SH, i:i_lim:SW] += col[:, :, j, i, :, :]
return img[:, :, PH:H + PH, PW:W + PW]
if '__file__' in globals():
import os, sys
sys.path.append(os.path.join(os.path.dirname(__file__), '..'))
from dezero.utils import get_conv_outsize
H, W = 4, 4 # input shape
KH, KW = 3, 3 # kernel shape
SH, SW = 1, 1 # stride
PH, PW = 1, 1 # padding
OH = get_conv_outsize(H, KH, SH, PH)
OW = get_conv_outsize(W, KW, SW, PW)
print(OH, OW)
