def __init__(
self,
in_channels,
out_channels,
conv_args="conv",
bn_args="bn",
relu_args="relu",
# additional arguments for conv
**kwargs,
):
super().__init__()
conv_op = build_conv(
in_channels=in_channels,
out_channels=out_channels,
**hp.merge_unify_args(conv_args, kwargs),
)
# register in order
self.empty_input = build_empty_input_op(conv_op)
self.conv = conv_op
self.bn = (
build_bn(num_channels=out_channels, **hp.unify_args(bn_args))
if bn_args is not None
else None
)
self.relu = (
build_relu(num_channels=out_channels, **hp.unify_args(relu_args))
if relu_args is not None
else None
)
self.out_channels = out_channels
def __init__(self, width_ratio=1.0, bn_args="bn", width_divisor=1):
self.width_ratio = width_ratio
self.last_depth = -1
self.width_divisor = width_divisor
# basic arguments that will be provided to all primitivies, they could be
# overrided by primitive parameters
self.basic_args = {
"bn_args": hp.unify_args(bn_args),
"width_divisor": width_divisor,
}
def __init__(
self,
in_channels,
out_channels,
conv_args="conv",
bn_args="bn",
relu_args="relu",
upsample_args="default",
# additional arguments for conv
**kwargs,
):
super().__init__()
conv_full_args = hp.merge_unify_args(conv_args, kwargs)
conv_stride = conv_full_args.pop("stride", 1)
# build upsample op if stride is negative
upsample_op, conv_stride = build_upsample_neg_stride(
stride=conv_stride, **hp.unify_args(upsample_args))
# build conv
conv_op = build_conv(
in_channels=in_channels,
out_channels=out_channels,
stride=conv_stride,
**conv_full_args,
)
# register in order
self.conv = conv_op
self.bn = (build_bn(num_channels=out_channels,
**hp.unify_args(bn_args))
if bn_args is not None and len(bn_args) != 0 else None)
self.relu = (build_relu(num_channels=out_channels,
**hp.unify_args(relu_args))
if relu_args is not None else None)
self.upsample = upsample_op
self.out_channels = out_channels
def __init__(self,
width_ratio=1.0,
bn_args=None,
width_divisor=1,
basic_args=None):
self.width_ratio = width_ratio
self.last_depth = -1
# basic arguments that will be provided to all primitivies, they could be
# overrided by primitive parameters
self.basic_args = {
**({
"bn_args": hp.unify_args(bn_args)
} if bn_args else {}),
"width_divisor":
width_divisor,
}
if basic_args is not None:
assert isinstance(basic_args, dict)
self.add_basic_args(**basic_args)
def __init__(
self,
in_channels,
out_channels,
expansion=6,
kernel_size=-1,
stride=1,
bias=False,
pw_args="conv",
pwl_args="conv",
pool_args=None,
bn_args="bn",
relu_args="relu",
se_args=None, # se after pool
pw_se_args=None, # se after pw conv
res_conn_args="default",
width_divisor=8,
always_pw=False,
less_se_channels=False,
):
super().__init__()
mid_channels = hp.get_divisible_by(in_channels * expansion, width_divisor)
self.pw = None
if in_channels != mid_channels or always_pw:
self.pw = bb.ConvBNRelu(
in_channels=in_channels,
out_channels=mid_channels,
conv_args={
"kernel_size": 1,
"stride": 1,
"padding": 0,
"bias": bias,
**hp.unify_args(pw_args),
},
bn_args=bn_args,
relu_args=relu_args,
)
pw_se_ratio = 0.25
if less_se_channels:
pw_se_ratio /= expansion
self.pw_se = bb.build_se(
in_channels=mid_channels,
mid_channels=(mid_channels * pw_se_ratio),
width_divisor=width_divisor,
**hp.merge_unify_args({"relu_args": relu_args}, pw_se_args)
)
if kernel_size == -1:
self.dw = nn.AdaptiveAvgPool2d(1)
else:
self.dw = nn.AvgPool2d(
kernel_size, stride=stride, **hp.unify_args(pool_args)
)
se_ratio = 0.25
if less_se_channels:
se_ratio /= expansion
self.se = bb.build_se(
in_channels=mid_channels,
mid_channels=(mid_channels * se_ratio),
width_divisor=width_divisor,
**hp.merge_unify_args({"relu_args": relu_args}, se_args)
)
self.pwl = bb.ConvBNRelu(
in_channels=mid_channels,
out_channels=out_channels,
conv_args={
"kernel_size": 1,
"stride": 1,
"padding": 0,
"bias": bias,
**hp.merge_unify_args(pwl_args),
},
# no bn
bn_args=None,
# has relu
relu_args=relu_args,
)
self.res_conn = bb.build_residual_connect(
in_channels=in_channels,
out_channels=out_channels,
stride=stride,
**hp.unify_args(res_conn_args)
)
self.out_channels = out_channels
def __init__(
self,
in_channels,
out_channels,
expansion=6,
kernel_size=3,
stride=1,
bias=False,
conv_args="conv",
bn_args="bn",
relu_args="relu",
se_args=None,
round_se_channels=False,
res_conn_args="default",
upsample_args="default",
width_divisor=8,
pw_args=None,
pw_bn_args=None,
dw_args=None,
dw_bn_args=None,
pwl_args=None,
pwl_bn_args=None,
dw_skip_bnrelu=False,
skip_pwl_bn=False,
pw_groups=1,
dw_group_ratio=1, # dw_group == mid_channels // dw_group_ratio
pwl_groups=1,
always_pw=False,
less_se_channels=False,
zero_last_bn_gamma=False,
drop_connect_rate=None,
):
super().__init__()
conv_args = hp.unify_args(conv_args)
bn_args = hp.unify_args(bn_args)
relu_args = hp.unify_args(relu_args)
mid_channels = hp.get_divisible_by(in_channels * expansion, width_divisor)
res_conn = bb.build_residual_connect(
in_channels=in_channels,
out_channels=out_channels,
stride=stride,
drop_connect_rate=drop_connect_rate,
**hp.unify_args(res_conn_args)
)
self.pw = None
self.shuffle = None
if in_channels != mid_channels or always_pw:
self.pw = bb.ConvBNRelu(
in_channels=in_channels,
out_channels=mid_channels,
conv_args={
"kernel_size": 1,
"stride": 1,
"padding": 0,
"bias": bias,
"groups": pw_groups,
**hp.merge_unify_args(conv_args, pw_args),
},
bn_args=hp.merge_unify_args(bn_args, pw_bn_args),
relu_args=relu_args,
)
if pw_groups > 1:
self.shuffle = bb.ChannelShuffle(pw_groups)
# use negative stride for upsampling
self.upsample, dw_stride = bb.build_upsample_neg_stride(
stride=stride, **hp.unify_args(upsample_args)
)
dw_padding = (
kernel_size // 2
if not (dw_args and "padding" in dw_args)
else dw_args.pop("padding")
)
self.dw = bb.ConvBNRelu(
in_channels=mid_channels,
out_channels=mid_channels,
conv_args={
"kernel_size": kernel_size,
"stride": dw_stride,
"padding": dw_padding,
"groups": mid_channels // dw_group_ratio,
"bias": bias,
**hp.merge_unify_args(conv_args, dw_args),
},
bn_args=hp.merge_unify_args(bn_args, dw_bn_args)
if not dw_skip_bnrelu
else None,
relu_args=relu_args if not dw_skip_bnrelu else None,
)
se_ratio = 0.25
if less_se_channels:
se_ratio /= expansion
self.se = bb.build_se(
in_channels=mid_channels,
mid_channels=(
int(mid_channels * se_ratio)
if not round_se_channels
else round(mid_channels * se_ratio)
),
width_divisor=width_divisor,
**hp.merge(relu_args=relu_args, kwargs=hp.unify_args(se_args))
)
self.pwl = bb.ConvBNRelu(
in_channels=mid_channels,
out_channels=out_channels,
conv_args={
"kernel_size": 1,
"stride": 1,
"padding": 0,
"bias": bias,
"groups": pwl_groups,
**hp.merge_unify_args(conv_args, pwl_args),
},
bn_args={
**hp.merge_unify_args(bn_args, pwl_bn_args),
**{
"zero_gamma": (
zero_last_bn_gamma if res_conn is not None else False
)
},
}
if not skip_pwl_bn
else None,
relu_args=None,
)
self.res_conn = res_conn
self.out_channels = out_channels