def add_channelwise_conv(self,
in_filters,
out_filters,
kernels,
strides=[1, 1, 1] * 1,
pads=[0, 0, 0] * 2,
suffix=''):
self.comp_idx += 1
assert in_filters == out_filters
self.prev_blob = brew.conv_nd(
self.model,
self.prev_blob,
'%scomp_%d_conv_%d%s' %
(self.prefix, self.comp_count, self.comp_idx, suffix),
in_filters,
out_filters,
weight_init=("MSRAFill", {}),
kernel=kernels,
strides=strides,
pads=pads,
group=in_filters,
no_bias=self.no_bias,
use_cudnn=False,
engine="CHANNELWISE_3D",
)
return self.prev_blob
def ConvNd(self, *args, **kwargs):
return brew.conv_nd(
self,
*args,
use_cudnn=self.use_cudnn,
order=self.order,
cudnn_exhaustive_search=self.cudnn_exhaustive_search,
ws_nbytes_limit=self.ws_nbytes_limit,
**kwargs)
def ConvNd(self, *args, **kwargs):
return brew.conv_nd(
self,
*args,
use_cudnn=self.use_cudnn,
order=self.order,
cudnn_exhaustive_search=self.cudnn_exhaustive_search,
ws_nbytes_limit=self.ws_nbytes_limit,
**kwargs
)
def add_conv(
self,
in_filters,
out_filters,
kernels,
strides=[1, 1, 1] * 1,
pads=[0, 0, 0] * 2,
block_type='3d', # set this to be '3d', '2.5d', or 'track',
group=1,
):
self.comp_idx += 1
if group > 1:
assert block_type == '3d-group'
log.info('in: %d out: %d' % (in_filters, out_filters))
if block_type == '2.5d':
i = 3 * in_filters * out_filters * kernels[1] * kernels[2]
i /= in_filters * kernels[1] * kernels[2] + 3 * out_filters
middle_filters = int(i)
self.prev_blob = self.model.ConvNd(
self.prev_blob,
'%scomp_%d_conv_%d_middle' %
(self.prefix, self.comp_count, self.comp_idx),
in_filters,
middle_filters,
[1, kernels[1], kernels[2]],
weight_init=("MSRAFill", {}),
strides=[1, strides[1], strides[2]] * 1,
pads=[0, pads[1], pads[2]] * 2,
no_bias=self.no_bias,
)
self.add_spatial_bn(middle_filters, suffix='_middle')
self.add_relu()
self.prev_blob = self.model.ConvNd(
self.prev_blob,
'%scomp_%d_conv_%d' %
(self.prefix, self.comp_count, self.comp_idx),
middle_filters,
out_filters,
[kernels[0], 1, 1],
weight_init=("MSRAFill", {}),
strides=[strides[0], 1, 1] * 1,
pads=[pads[0], 0, 0] * 2,
no_bias=self.no_bias,
)
elif block_type == '0.3d' or block_type == '0.3d+relu':
self.prev_blob = self.model.ConvNd(
self.prev_blob,
'%scomp_%d_conv_%d_middle' %
(self.prefix, self.comp_count, self.comp_idx),
in_filters,
out_filters,
[1, 1, 1],
weight_init=("MSRAFill", {}),
strides=[1, 1, 1] * 1,
pads=[0, 0, 0] * 2,
no_bias=self.no_bias,
)
self.add_spatial_bn(out_filters, suffix='_middle')
if block_type == '0.3d+relu':
self.add_relu()
self.prev_blob = brew.conv_nd(
self.model,
self.prev_blob,
'%scomp_%d_conv_%d' %
(self.prefix, self.comp_count, self.comp_idx),
out_filters,
out_filters,
weight_init=("MSRAFill", {}),
kernel=kernels,
strides=strides,
pads=pads,
group=out_filters,
no_bias=self.no_bias,
use_cudnn=False,
engine="CHANNELWISE_3D",
)
elif block_type == '3d':
self.prev_blob = self.model.ConvNd(
self.prev_blob,
'%scomp_%d_conv_%d' %
(self.prefix, self.comp_count, self.comp_idx),
in_filters,
out_filters,
kernels,
weight_init=("MSRAFill", {}),
strides=strides,
pads=pads,
no_bias=self.no_bias,
)
elif block_type == '3d-sep': # channel_wise 3d conv block
self.add_channelwise_conv(
in_filters,
out_filters,
kernels,
strides=strides,
pads=pads,
)
elif block_type == '2.5d-sep':
assert in_filters == out_filters
middle_filters = out_filters
self.add_channelwise_conv(in_filters,
middle_filters,
[1, kernels[1], kernels[2]],
strides=[1, strides[1], strides[2]] * 1,
pads=[0, pads[1], pads[2]] * 2,
suffix='_middle')
self.add_spatial_bn(middle_filters, suffix='_middle')
self.add_relu()
self.prev_blob = self.model.ConvNd(
self.prev_blob,
'%scomp_%d_conv_%d_1x' %
(self.prefix, self.comp_count, self.comp_idx),
middle_filters,
middle_filters,
[1, 1, 1],
weight_init=("MSRAFill", {}),
strides=[1, 1, 1] * 1,
pads=[0, 0, 0] * 2,
no_bias=self.no_bias,
)
self.add_spatial_bn(middle_filters, suffix='_1x')
self.add_relu()
self.add_channelwise_conv(middle_filters,
out_filters, [kernels[0], 1, 1],
strides=[strides[0], 1, 1] * 1,
pads=[pads[0], 0, 0] * 2)
else:
log.info('Unknown block type!')
return self.prev_blob