def get_flop(self, kernel_size, stride, in_channel, out_channel, affine,
in_h, in_w):
cal_flop = flops.count_ConvBNReLU_flop(in_h,
in_w,
in_channel,
self.hidden_dim,
kernel_size=[1, 1],
is_bias=False,
is_affine=affine)
cal_flop += flops.count_Conv_flop(in_h,
in_w,
self.hidden_dim,
self.hidden_dim,
kernel_size,
False,
stride,
groups=self.hidden_dim,
is_affine=affine)
in_h, in_w = in_h // stride, in_w // stride
# pw-linear without ReLU!
cal_flop += flops.count_Conv_flop(in_h,
in_w,
self.hidden_dim,
out_channel,
kernel_size=[1, 1],
is_bias=False)
cal_flop += flops.count_BN_flop(in_h, in_w, out_channel, affine)
return cal_flop
def get_flop(self, kernel_size, stride, in_channel, out_channel, affine,
in_h, in_w):
cal_flop = flops.count_Conv_flop(in_h,
in_w,
in_channel,
in_channel,
kernel_size,
False,
stride,
groups=in_channel)
in_h, in_w = in_h // stride, in_w // stride
cal_flop += flops.count_ConvBNReLU_flop(in_h,
in_w,
in_channel,
in_channel,
kernel_size=[1, 1],
is_bias=False,
is_affine=affine)
# stack 2 separate depthwise-conv.
cal_flop += flops.count_Conv_flop(in_h,
in_w,
in_channel,
in_channel,
kernel_size,
False,
stride=1,
groups=in_channel)
cal_flop += flops.count_Conv_flop(in_h,
in_w,
in_channel,
out_channel,
kernel_size=[1, 1],
is_bias=False)
cal_flop += flops.count_BN_flop(in_h, in_w, out_channel, affine)
return cal_flop
def get_flop(self, kernel_size, stride, in_channel, out_channel, affine,
in_h, in_w):
cal_flop = flops.count_Conv_flop(in_h, in_w, in_channel, out_channel,
kernel_size, False, stride)
in_h, in_w = in_h // stride, in_w // stride
cal_flop += flops.count_BN_flop(in_h, in_w, out_channel, affine)
return cal_flop
def get_flop(self, kernel_size, stride, in_channel, out_channel, affine,
in_h, in_w):
cal_flop = flops.count_Conv_flop(in_h,
in_w,
in_channel,
out_channel,
kernel_size=kernel_size,
is_bias=False)
cal_flop += flops.count_BN_flop(in_h, in_w, out_channel, affine)
return cal_flop
def get_flop(self, kernel_size, stride, in_channel, out_channel, affine,
in_h, in_w):
cal_flop = flops.count_Conv_flop(in_h,
in_w,
in_channel,
in_channel,
kernel_size,
False,
stride,
groups=in_channel)
in_h, in_w = in_h // stride, in_w // stride
cal_flop += flops.count_Conv_flop(in_h,
in_w,
in_channel,
out_channel,
kernel_size=[1, 1],
is_bias=False)
cal_flop += flops.count_BN_flop(in_h, in_w, out_channel, affine)
return cal_flop
