def augment(V, crop, grayscale=False):
# note assumes square outputs of size crop x crop
# randomly sample a crop in the input volume
if crop == V.sx: return V
dx = randi(0, V.sx - crop)
dy = randi(0, V.sy - crop)
W = Vol(crop, crop, V.depth)
for x in xrange(crop):
for y in xrange(crop):
if x + dx < 0 or x + dx >= V.sx or \
y + dy < 0 or y + dy >= V.sy: continue
for d in xrange(V.depth):
W.set(x, y, d, V.get(x + dx, y + dy, d))
if grayscale:
#flatten into depth=1 array
G = Vol(crop, crop, 1, 0.0)
for i in xrange(crop):
for j in xrange(crop):
G.set(i, j, 0, W.get(i, j, 0))
W = G
return W
def forward(self, V, is_training):
self.in_act = V
N = self.out_depth
V2 = Vol(self.out_sx, self.out_sy, self.out_depth, 0.0)
if self.out_sx == 1 and self.out_sy == 1:
for i in xrange(N):
offset = i * self.group_size
m = max(V.w[offset:])
index = V.w[offset:].index(m)
V2.w[i] = m
self.switches[i] = offset + index
else:
switch_counter = 0
for x in xrange(V.sx):
for y in xrange(V.sy):
for i in xrange(N):
ix = i * self.group_size
elem = V.get(x, y, ix)
elem_i = 0
for j in range(1, self.group_size):
elem2 = V.get(x, y, ix + j)
if elem2 > elem:
elem = elem2
elem_i = j
V2.set(x, y, i, elem)
self.switches[i] = ix + elem_i
switch_counter += 1
self.out_act = V2
return self.out_act
def augment(V, crop, grayscale=False):
# note assumes square outputs of size crop x crop
# randomly sample a crop in the input volume
if crop == V.sx: return V
dx = randi(0, V.sx - crop)
dy = randi(0, V.sy - crop)
W = Vol(crop, crop, V.depth)
for x in xrange(crop):
for y in xrange(crop):
if x + dx < 0 or x + dx >= V.sx or \
y + dy < 0 or y + dy >= V.sy:
continue
for d in xrange(V.depth):
W.set(x, y, d, V.get(x + dx, y + dy, d))
if grayscale:
#flatten into depth=1 array
G = Vol(crop, crop, 1, 0.0)
for i in xrange(crop):
for j in xrange(crop):
G.set(i, j, 0, W.get(i, j, 0))
W = G
return W
def forward(self, V, is_training):
self.in_act = V
N = self.out_depth
V2 = Vol(self.out_sx, self.out_sy, self.out_depth, 0.0)
if self.out_sx == 1 and self.out_sy == 1:
for i in xrange(N):
offset = i * self.group_size
m = max(V.w[offset:])
index = V.w[offset:].index(m)
V2.w[i] = m
self.switches[i] = offset + index
else:
switch_counter = 0
for x in xrange(V.sx):
for y in xrange(V.sy):
for i in xrange(N):
ix = i * self.group_size
elem = V.get(x, y, ix)
elem_i = 0
for j in range(1, self.group_size):
elem2 = V.get(x, y, ix + j)
if elem2 > elem:
elem = elem2
elem_i = j
V2.set(x, y, i, elem)
self.switches[i] = ix + elem_i
switch_counter += 1
self.out_act = V2
return self.out_act
def load_data(crop, gray, training=True):
filename = './data/cifar10_'
if training:
filename += 'train.npz'
else:
filename += 'test.npz'
data = numpy.load(filename)
xs = data['x']
ys = data['y']
for i in xrange(len(xs)):
V = Vol(32, 32, 3, 0.0)
for d in xrange(3):
for x in xrange(32):
for y in xrange(32):
px = xs[i][x * 32 + y, d] / 255.0 - 0.5
V.set(x, y, d, px)
if crop:
V = augment(V, 24, gray)
y = ys[i]
yield V, y
def load_data(crop, gray, training=True):
filename = './data/cifar10_'
if training:
filename += 'train.npz'
else:
filename += 'test.npz'
data = numpy.load(filename)
xs = data['x']
ys = data['y']
for i in xrange(len(xs)):
V = Vol(32, 32, 3, 0.0)
for d in xrange(3):
for x in xrange(32):
for y in xrange(32):
px = xs[i][x * 32 + y, d] / 255.0 - 0.5
V.set(x, y, d, px)
if crop:
V = augment(V, 24, gray)
y = ys[i]
yield V, y
def forward(self, V, is_training):
self.in_act = V
A = Vol(self.out_sx, self.out_sy, self.out_depth, 0.0)
switch_counter = 0
for d in xrange(self.out_depth):
x = -self.pad
y = -self.pad
for ax in xrange(self.out_sx):
y = -self.pad
for ay in xrange(self.out_sy):
# convolve centered at this particular location
max_a = -99999
win_x, win_y = -1, -1
for fx in xrange(self.sx):
for fy in xrange(self.sy):
off_x = x + fx
off_y = y + fy
if off_y >= 0 and off_y < V.sy \
and off_x >= 0 and off_x < V.sx:
v = V.get(off_x, off_y, d)
# max pool
if v > max_a:
max_a = v
win_x = off_x
win_y = off_y
self.switch_x[switch_counter] = win_x
self.switch_y[switch_counter] = win_y
switch_counter += 1
A.set(ax, ay, d, max_a)
y += self.stride
x += self.stride
self.out_act = A
return self.out_act
def forward(self, V, is_training):
self.in_act = V
A = Vol(self.out_sx, self.out_sy, self.out_depth, 0.0)
v_sx = V.sx
v_sy = V.sy
xy_stride = self.stride
for d in xrange(self.out_depth):
f = self.filters[d]
x = -self.pad
y = -self.pad
for ay in xrange(self.out_sy):
x = -self.pad
for ax in xrange(self.out_sx):
# convolve centered at this particular location
sum_a = 0.0
for fy in xrange(f.sy):
off_y = y + fy
for fx in xrange(f.sx):
# coordinates in the original input array coordinates
off_x = x + fx
if off_y >= 0 and off_y < V.sy and off_x >= 0 and off_x < V.sx:
for fd in xrange(f.depth):
sum_a += f.w[((f.sx * fy) + fx) * f.depth + fd] \
* V.w[((V.sx * off_y) + off_x) * V.depth + fd]
sum_a += self.biases.w[d]
A.set(ax, ay, d, sum_a)
x += xy_stride
y += xy_stride
self.out_act = A
return self.out_act
def forward(self, V, is_training):
self.in_act = V
A = Vol(self.out_sx, self.out_sy, self.out_depth, 0.0)
v_sx = V.sx
v_sy = V.sy
xy_stride = self.stride
for d in xrange(self.out_depth):
f = self.filters[d]
x = -self.pad
y = -self.pad
for ay in xrange(self.out_sy):
x = -self.pad
for ax in xrange(self.out_sx):
# convolve centered at this particular location
sum_a = 0.0
for fy in xrange(f.sy):
off_y = y + fy
for fx in xrange(f.sx):
# coordinates in the original input array coordinates
off_x = x + fx
if off_y >= 0 and off_y < V.sy and off_x >= 0 and off_x < V.sx:
for fd in xrange(f.depth):
sum_a += f.w[((f.sx * fy) + fx) * f.depth + fd] \
* V.w[((V.sx * off_y) + off_x) * V.depth + fd]
sum_a += self.biases.w[d]
A.set(ax, ay, d, sum_a)
x += xy_stride
y += xy_stride
self.out_act = A
return self.out_act