def forward(self, input):
x, meta = input
identity = x
if self.downsample is not None:
identity = self.downsample(x)
if not self.sparse:
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out += identity
else:
assert meta is not None
m = self.masker(x, meta)
mask_dilate, mask = m['dilate'], m['std']
x = dynconv.conv3x3(self.conv1, x, None, mask_dilate)
x = dynconv.bn_relu(self.bn1, self.relu, x, mask_dilate)
x = dynconv.conv3x3(self.conv2, x, mask_dilate, mask)
x = dynconv.bn_relu(self.bn2, None, x, mask)
out = identity + dynconv.apply_mask(x, mask)
out = self.relu(out)
return out, meta
def forward(self, v):
x, meta = v
if not self.sparse:
out = self.conv(x)
if self.identity:
out += x
return out, meta
else:
assert self.identity and self.expand_ratio != 1
m, meta = self.masker(x, meta)
mask, mask_dilate = m['std'], m['dilate']
fast_inference = not self.training
out = x.clone(
) # clone should not be needed, but otherwise seems to be bugged
if fast_inference:
x = dynconv.gather(x, mask_dilate)
x = dynconv.conv1x1(self.conv[0],
x,
mask_dilate,
fast=fast_inference)
x = dynconv.bn_relu(self.conv[1],
self.conv[2],
x,
mask_dilate,
fast=fast_inference)
x = dynconv.conv3x3_dw(self.conv[3],
x,
mask_dilate,
mask,
fast=fast_inference)
x = dynconv.bn_relu(self.conv[4],
self.conv[5],
x,
mask,
fast=fast_inference)
x = dynconv.conv1x1(self.conv[6], x, mask, fast=fast_inference)
x = dynconv.bn_relu(self.conv[7],
None,
x,
mask,
fast=fast_inference)
if fast_inference:
out = dynconv.scatter(x, out, mask, sum_out=True)
else:
out = out + dynconv.apply_mask(x, mask)
return out, meta