def box(self, original, w, *args, **kargs):
w = self.w.getVal(c = w, shape = original.shape, **kargs)
sub = self.sub.getVal(c = 1, shape = original.shape, **kargs)
assert (0 <= h.dten(w)).all()
assert (h.dten(sub) <= 1).all()
assert (0 <= h.dten(sub)).all()
if h.dten(w).sum().item() <= 0.0:
inter = original
else:
inter = self.a.box(original, w = w * (1 - sub), *args, **kargs)
return self.b.box(inter, w = w * sub, *args, **kargs)
def box(self, original, w, *args, **kargs):
w = self.w.getVal(c = w, shape = original.shape[:1], **kargs)
sub = self.sub.getVal(c = 1, shape = original.shape[:1], **kargs)
assert (0 <= h.dten(w)).all()
assert (h.dten(sub) <= 1).all()
assert (0 <= h.dten(sub)).all()
if self.normal:
inter = torch.randn_like(original, device = h.device)
else:
inter = (torch.rand_like(original, device = h.device) * 2 - 1)
inter = inter * w * (1 - sub)
return self.a.box(original + inter, w = w * sub, *args, **kargs)
def printNet(self, f): # only complete if we've forwardt stride=1
print("Conv2D", file=f)
sz = list(self.prev)
print(
self.activation +
", filters={}, kernel_size={}, input_shape={}, stride={}, padding={}"
.format(self.out_channels, [self.kernel_size, self.kernel_size],
list(reversed(sz)), [self.stride, self.stride],
self.padding),
file=f)
print(h.printListsNumpy([[list(p) for p in l]
for l in self.weight.permute(2, 3, 1, 0).data
]),
file=f)
print(h.printNumpy(
self.bias if self.bias is not None else h.dten(self.out_channels)),
file=f)
def init(self, in_shape, mean, std, **kargs):
self.mean_v = mean
self.std_v = std
self.mean = h.dten(mean)
self.std = 1 / h.dten(std)
return in_shape
def box(self, original, w, *args, **kargs):
epsilon = self.epsilon.getVal(c = w, shape = original.shape[:1], **kargs)
assert (0 <= h.dten(epsilon)).all()
epsilon = torch.randn(original.size()[0:1], device = h.device)[0] * epsilon
return self.a.box(original, w = epsilon, *args, **kargs)