def _oper_cpu(cls, x, w, b, momentum, mov_m, mov_s, inference, mode,
epsilon):
if mode == BATCH_NORMALIZE_FEATUREMAP:
axs = (0, 2, 3)
else:
axs = (0, )
if inference:
mean = mov_m
var = mov_s
else:
mean = np.mean(to_value(x), axis=axs, keepdims=True)
var = np.var(to_value(x), axis=axs, keepdims=True)
sq_var = 1.0 / np.sqrt(var + epsilon)
xh = (to_value(x) - mean) * sq_var
z = to_value(w) * xh + to_value(b)
ret = cls._create_node(z)
ret.attrs._axs = axs
ret.attrs._x = x
ret.attrs._w = w
ret.attrs._b = b
ret.attrs._m = mean
ret.attrs._v = sq_var
if not inference:
N = np.prod([x.shape[s] for s in axs])
ret.attrs._mov_m = (1 - momentum) * mov_m + momentum * mean
ret.attrs._mov_v = (
1 - momentum) * mov_s + momentum * var * N / max(N - 1., 1.)
return ret
def _backward_cpu(self, context, dy, **kwargs):
if isinstance(self.attrs._lhs, Node):
self.attrs._lhs._update_diff(
context, np.dot(dy,
to_value(self.attrs._rhs).T), **kwargs)
if isinstance(self.attrs._rhs, Node):
self.attrs._rhs._update_diff(
context, np.dot(to_value(self.attrs._lhs).T, dy), **kwargs)
def forward(self, x):
ret = peephole_lstm(x, getattr(self, "_z", None),
getattr(self, "_state", None),
self.params)
self._z = to_value(ret)
self._state = to_value(getattr(ret, '_state', None))
if hasattr(self, "_last_node") and self._last_node is not None and ret.attrs.get_attrs():
setattr(self._last_node.attrs, "_pgated_f", to_value(ret.attrs._gated_f))
self._last_node = ret
return ret
def _oper_cpu(cls, lhs, rhs):
N = len(lhs)
z = 1. / (1. + np.exp(to_value(-lhs)))
loss = -np.sum(to_value(rhs) * np.log(z + 1e-8) +
to_value(1 - rhs) * np.log(1 - z + 1e-8)) / N
ret = cls._create_node(loss)
ret.attrs._z = z
ret.attrs._lhs = lhs
ret.attrs._rhs = rhs
return ret
def _backward_cpu(self, context, dy):
p = self.attrs._p
s = self.attrs._state
ps = self.attrs._pstate
u = self.attrs._u
go = self.attrs._gated_o
gf = self.attrs._gated_f
gi = self.attrs._gated_i
pgf = np.zeros_like(gf) if self.attrs._pgated_f is None else self.attrs._pgated_f
drt, dit, dft, dot, dct = (context.restore(dt, np.zeros_like(dy))
for dt in self.attrs._dt_d)
activated_s = tanh(s)
activated_u = tanh(u)
e = dy + np.dot(drt, p["wr"].T) + np.dot(dit, p["wir"].T) + \
np.dot(dft, p["wfr"].T) + np.dot(dot, p["wor"].T)
do = gate_diff(go) * activated_s * e
ds = go * activation_diff(activated_s) * e
dc = ds + pgf * dct + p["wfc"] * dft + p["wic"] * dit + p["woc"] * do
df = gate_diff(gf) * ps * dc if ps is not None else np.zeros_like(gf)
di = gate_diff(gi) * activated_u * dc
d = gi * activation_diff(activated_u) * dc
dx = np.dot(d, p["w"].T) \
+ np.dot(di, p["wi"].T) \
+ np.dot(do, p["wo"].T) \
+ np.dot(df, p["wf"].T)
for dt_d, dt in zip(self.attrs._dt_d, (d, di, df, do, dc)):
context.store(dt_d, dt)
if isinstance(self.attrs._x, Node):
self.attrs._x._update_diff(context, dx)
for k, diff in zip(("w", "wo", "wi", "wf"), (d, do, di, df)):
if isinstance(p[k], Node):
p[k]._update_diff(context, np.dot(to_value(self.attrs._x).T, diff))
for k, diff in zip(("wr", "wor", "wir", "wfr"), (drt, dot, dit, dft)):
if isinstance(p[k], Node):
p[k]._update_diff(context, np.dot(to_value(self).T, diff))
for k, diff in zip(("wfc", "wic", "woc"), (dft, dit, do)):
if isinstance(p[k], Node):
p[k]._update_diff(context, np.sum(diff * s, axis=0, keepdims=True))
for k, diff in zip(("b", "bf", "bi", "bo"), (d, df, di, do)):
if isinstance(p[k], Node):
p[k]._update_diff(context, np.sum(diff, axis=0, keepdims=True))
def _backward_cpu(self, context, dy, **kwargs):
a = self.attrs._axs
sq_var = self.attrs._v
meaned = self.attrs._x - self.attrs._m
N = np.prod([self.attrs._x.shape[s] for s in a])
if isinstance(self.attrs._x, Node):
dxh = dy * to_value(self.attrs._w)
ds = np.sum(dxh * meaned * -np.power(sq_var, 3) / 2,
axis=a,
keepdims=True)
du = np.sum(-dxh * sq_var, axis=a, keepdims=True)
dx = dxh * sq_var + (ds * 2 * meaned + du) / N
self.attrs._x._update_diff(context, dx, **kwargs)
if isinstance(self.attrs._w, Node):
xh = meaned * sq_var
self.attrs._w._update_diff(context,
np.sum(xh * dy, axis=a, keepdims=True),
**kwargs)
if isinstance(self.attrs._b, Node):
self.attrs._b._update_diff(context,
np.sum(dy, axis=a, keepdims=True),
**kwargs)
def _backward_cpu(self, context, dy, **kwargs):
axis = self.attrs._axis
args = np.split(to_value(dy), self.attrs._index, axis=axis)
for i in range(len(self.attrs._index) + 1):
arg = getattr(self.attrs, "_arg%d" % i)
if isinstance(arg, Node):
arg._update_diff(context, args[i], **kwargs)
def _backward_cpu(self, context, dy):
ldy, rdy = np.hsplit(to_value(dy), [self.attrs._index])
if isinstance(self.attrs._lhs, Node):
self.attrs._lhs._update_diff(context, ldy)
if isinstance(self.attrs._rhs, Node):
self.attrs._rhs._update_diff(context, rdy)
def region_cordinates(roi, spatial_scale):
idx, xmin, ymin, xmax, ymax = to_value(roi)
idx = int(idx)
xmin = int(round(xmin * spatial_scale))
ymin = int(round(ymin * spatial_scale))
xmax = int(round(xmax * spatial_scale))
ymax = int(round(ymax * spatial_scale))
return idx, xmin, ymin, xmax, ymax
def _oper_cpu(cls, x, w, b, in_shape, out_shape, kernel, stride, padding):
col = im2col(to_value(x), out_shape[1:], kernel, stride, padding)
value = np.rollaxis(
np.tensordot(col, to_value(w), ([1, 2, 3], [1, 2, 3])), 3, 1)
if b is not None:
value += b
ret = cls._create_node(value)
ret.attrs._col = col
ret.attrs._x = x
ret.attrs._w = w
ret.attrs._b = b
ret.attrs._in_shape = in_shape
ret.attrs._out_shape = out_shape
ret.attrs._kernel = kernel
ret.attrs._stride = stride
ret.attrs._padding = padding
return ret
def _oper_cpu(cls, x, w, b, in_shape, kernel, stride, padding):
col = imncol(to_value(x), w, stride, padding)
if b is not None:
col += b
ret = cls._create_node(col)
ret.attrs._x = x
ret.attrs._w = w
ret.attrs._b = b
ret.attrs._kernel = kernel
ret.attrs._stride = stride
ret.attrs._padding = padding
return ret
def _oper_cpu(cls, x, w, b, in_shape, out_shape, kernel, stride, padding,
dilation, groups):
N, in_channels, in_h, in_w = x.shape
k_h, k_w = kernel
out_channels = w.shape[0]
iCg = in_channels // groups
oCg = out_channels // groups
col = im2col(to_value(x), out_shape[1:], kernel, stride, padding,
dilation)
out_h, out_w = col.shape[-2:]
col = col.transpose(1, 2, 3, 0, 4, 5)
col = col.reshape(groups, iCg * k_h * k_w, N * out_h * out_w)
w_new = w.reshape(groups, oCg, iCg * k_h * k_w)
value = np.matmul(to_value(w_new), col)
value = value.reshape(groups * oCg, N, out_h, out_w)
value = value.transpose(1, 0, 2, 3)
if b is not None:
value += b.reshape(1, b.size, 1, 1)
ret = cls._create_node(value)
ret.attrs._col = col
ret.attrs._x = x
ret.attrs._w = w
ret.attrs._b = b
ret.attrs._in_shape = in_shape
ret.attrs._out_shape = out_shape
ret.attrs._kernel = kernel
ret.attrs._stride = stride
ret.attrs._padding = padding
ret.attrs._dilation = dilation
ret.attrs._groups = groups
ret.attrs._iCg = iCg
ret.attrs._oCg = oCg
return ret
def _oper_gpu(cls, x, w, b, momentum, mov_m, mov_s, inference, mode,
epsilon):
if mode == BATCH_NORMALIZE_FEATUREMAP:
axs = 1
else:
axs = 0
if b is None:
b = get_gpu(w).zeros_like_me()
y, mean, sq_var = (get_gpu(g).empty_like_me() for g in (x, w, w))
if inference:
inv_var = 1.0 / np.sqrt(to_value(mov_s) + epsilon)
if isinstance(inv_var, Number):
inv_var = inv_var * np.ones_like(w)
mov_m = get_gpu(mov_m)
mov_s = get_gpu(mov_s)
mv_m = mov_m if isinstance(mov_m,
GPUValue) else get_gpu(w).zeros_like_me()
mv_v = mov_s if isinstance(mov_s,
GPUValue) else get_gpu(w).zeros_like_me()
with cu.cudnn_handler() as handle:
cu.cuBatchNormalizatoinForward(handle,
get_gpu(x),
mv_m,
mv_v,
get_gpu(w),
get_gpu(b),
y,
mean,
sq_var,
momentum=momentum,
mode=axs,
inference=inference,
eps=epsilon)
ret = cls._create_node(y)
ret.attrs._axs = axs
ret.attrs._x = x
ret.attrs._w = w
ret.attrs._b = b
if inference:
ret.attrs._m = mv_m
ret.attrs._v = inv_var
else:
ret.attrs._m = mean
ret.attrs._v = sq_var
ret.attrs._mov_m = mv_m
ret.attrs._mov_v = mv_v
return ret
def _backward_cpu(self, context, dy, **kwargs):
if isinstance(self.attrs._x, Node):
dy = to_value(dy)
unit_scale = self.attrs._unit_scale
scale = self.attrs._scale
a = self.attrs._a
b = self.attrs._b
n = self.attrs._n
x = self.attrs._x
sum1 = (self * dy / unit_scale).view(np.ndarray)
sum2 = sum1.copy()
for i in range(1, n // 2 + 1):
sum2[:, i:, :, :] += sum1[:, :-i, :, :]
sum2[:, :-i, :, :] += sum1[:, i:, :, :]
self.attrs._x._update_diff(context, dy * scale - 2 * a * b * x * sum2, **kwargs)
def _backward_cpu(self, context, dy):
dy = to_value(dy)
if isinstance(self.attrs._x, Node):
dx = np.tensordot(self.attrs._w, dy, (0, 1))
dx = np.rollaxis(dx, 3)
dx = col2im(dx, self.attrs._in_shape[1:],
self.attrs._stride, self.attrs._padding)
self.attrs._x._update_diff(context, dx)
if isinstance(self.attrs._w, Node):
self.attrs._w._update_diff(context, np.tensordot(
dy, self.attrs._col, ([0, 2, 3], [0, 4, 5])))
if isinstance(self.attrs._b, Node):
self.attrs._b._update_diff(context, np.sum(dy, (0, 2, 3), keepdims=True))
def _backward_cpu(self, context, dy, **kwargs):
dy = to_value(dy)
N, in_channels, in_h, in_w = self.attrs._x.shape
groups = self.attrs._groups
oCg = self.attrs._oCg
iCg = self.attrs._iCg
out_h, out_w = self.attrs._out_shape[-2:]
k_h, k_w = self.attrs._kernel
if isinstance(self.attrs._x, Node):
dy_temp = dy.transpose(1, 0, 2, 3)
dy_temp = dy_temp.reshape(groups, oCg, N * out_h * out_w)
w_temp = self.attrs._w.reshape(groups, oCg, iCg * k_h * k_w)
w_temp = w_temp.transpose(0, 2, 1)
dx = np.matmul(w_temp, dy_temp)
dx = dx.reshape(groups * iCg, k_h, k_w, N, out_h, out_w)
dx = np.rollaxis(dx, 3)
dx = col2im(dx, self.attrs._in_shape[1:], self.attrs._stride,
self.attrs._padding, self.attrs._dilation)
self.attrs._x._update_diff(context, dx, **kwargs)
if isinstance(self.attrs._w, Node):
col_temp = self.attrs._col
col_temp = col_temp.transpose(0, 2, 1)
dy_temp = dy.transpose(1, 0, 2, 3)
dy_temp = dy_temp.reshape(groups, oCg, N * out_h * out_w)
dw = np.matmul(dy_temp, col_temp)
dw = dw.reshape(groups * oCg, iCg, k_h, k_w)
self.attrs._w._update_diff(context, dw, **kwargs)
if isinstance(self.attrs._b, Node):
self.attrs._b._update_diff(context,
np.sum(dy, (0, 2, 3), keepdims=True),
**kwargs)
def _oper_cpu(cls, arg, axis, keepdims):
array = to_value(arg)
return np.amin(array, axis, keepdims=keepdims), np.argmin(array, axis)
def _oper_cpu(cls, arg, axis, keepdims):
array = to_value(arg)
# Max is calculated twice, update?
return np.amax(array, axis, keepdims=keepdims), np.argmax(array, axis)
def close(a, b):
assert np.allclose(to_value(a), to_value(b), atol=1e-4, rtol=1e-3)
def close(GPU, CPU):
print('GPU =')
print(to_value(GPU))
print('CPU =')
print(to_value(CPU))
assert np.allclose(to_value(GPU), to_value(CPU), atol=1e-4, rtol=1e-3)