def get_nasfpn_neck(self, data):
if self.neck is not None:
return self.neck
dim_reduced = self.p.dim_reduced
norm = self.p.normalizer
num_stage = self.p.num_stage
S0_kernel = self.p.S0_kernel
import mxnet as mx
xavier_init = mx.init.Xavier(factor_type="avg", rnd_type="uniform", magnitude=3)
c2, c3, c4, c5 = data
c6 = X.pool(data=c5, name="C6", kernel=2, stride=2, pad=0)
c7 = X.pool(data=c5, name="C7", kernel=4, stride=4, pad=0)
c_features = [c3, c4, c5, c6, c7]
# 0 stage
p0_names = ['S0_P3', 'S0_P4', 'S0_P5', 'S0_P6', 'S0_P7']
p_features = self.get_P0_features(c_features, p0_names, dim_reduced, xavier_init, norm, S0_kernel)
# stack stage
for i in range(num_stage):
p_features = self.get_fused_P_feature(p_features, i + 1, dim_reduced, xavier_init, norm)
self.neck = dict(
stride8=p_features['S{}_P3'.format(num_stage)],
stride16=p_features['S{}_P4'.format(num_stage)],
stride32=p_features['S{}_P5'.format(num_stage)],
stride64=p_features['S{}_P6'.format(num_stage)],
stride128=p_features['S{}_P7'.format(num_stage)]
)
return self.neck
def get_nasfpn_neck(self, data):
dim_reduced = self.dim_reduced
norm = self.norm
num_stage = self.num_stage
import mxnet as mx
xavier_init = mx.init.Xavier(factor_type="in",
rnd_type="uniform",
magnitude=3)
c2, c3, c4, c5 = data
c6 = X.pool(data=c5, name="C6", kernel=3, stride=2, pool_type="max")
c7 = X.pool(data=c5, name="C7", kernel=5, stride=4, pool_type="max")
c_features = [c3, c4, c5, c6, c7]
# the 0 stage
p0_names = ['S0_P3', 'S0_P4', 'S0_P5', 'S0_P6', 'S0_P7']
p_features = self.get_P0_features(c_features, p0_names, dim_reduced,
xavier_init, norm)
# stack stage
for i in range(num_stage):
p_features = self.get_fused_P_feature(p_features, i + 1,
dim_reduced, xavier_init,
norm)
return p_features['S{}_P3'.format(num_stage)], \
p_features['S{}_P4'.format(num_stage)], \
p_features['S{}_P5'.format(num_stage)], \
p_features['S{}_P6'.format(num_stage)], \
p_features['S{}_P7'.format(num_stage)]
def _get_bbox_head_logit(self, conv_feat):
if self._head_feat is not None:
return self._head_feat
from mxnext.backbone.resnet_v2 import Builder
unit = Builder.resnet_stage(
conv_feat,
name="stage4",
num_block=3,
filter=2048,
stride=1,
dilate=1,
norm_type=self.p.normalizer,
norm_mom=0.9,
ndev=8
)
bn1 = X.fixbn(unit, name='bn1')
relu1 = X.relu(bn1, name='relu1')
relu1 = X.to_fp32(relu1, name='c5_to_fp32')
pool1 = X.pool(relu1, global_pool=True, name='pool1')
self._head_feat = pool1
return self._head_feat
def _get_output(self, mask_pred_logits, conv_feat):
num_class = self.pBbox.num_class
msra_init = mx.init.Xavier(rnd_type="gaussian", factor_type="out", magnitude=2)
normal_init = mx.init.Normal(0.01)
kaiming_uniform = mx.init.Xavier(rnd_type='uniform', factor_type='in', magnitude=3)
mask_pred_logits = mx.sym.expand_dims(mask_pred_logits, axis=1)
iou_head_maxpool_1 = X.pool(
mask_pred_logits,
name='iou_head_maxpool_1',
kernel=2,
stride=2,
pad=0,
)
iou_head_input = X.concat([conv_feat, iou_head_maxpool_1], axis=1, name='iou_head_input')
hi = iou_head_input
for ii in range(3):
hi = X.conv(
hi,
filter=256,
kernel=3,
stride=1,
name='iou_head_conv_%d'%ii,
no_bias=False,
init=msra_init,
)
hi = X.relu(hi)
hi = X.conv(
hi,
filter=256,
kernel=3,
stride=2,
name='iou_head_conv_3',
no_bias=False,
init=msra_init
)
hi = X.relu(hi)
hi = X.flatten(data=hi)
fc1 = X.relu(X.fc(hi, filter=1024, name='iou_head_FC1', init=kaiming_uniform))
fc2 = X.relu(X.fc(fc1, filter=1024, name='iou_head_FC2', init=kaiming_uniform))
iou_pred_logits = X.fc(fc2, filter=num_class, name='iou_head_pred', init=normal_init)
return iou_pred_logits
def _get_bbox_head_logit(self, conv_feat):
if self._head_feat is not None:
return self._head_feat
from mxnext.backbone.resnext import Builder
unit = Builder.resnext_stage(conv_feat,
name="stage4",
num_block=3,
filter=2048,
stride=1,
dilate=1,
num_group=self.p.num_group,
norm_type=self.p.normalizer,
norm_mom=0.9,
ndev=8)
pool1 = X.pool(unit, global_pool=True, name='pool1')
self._head_feat = pool1
return self._head_feat
def get_fpg_neck(self, data):
if self.neck is not None:
return self.neck
dim_reduced = self.p.dim_reduced
norm = self.p.normalizer
num_stage = self.p.num_stage
S0_kernel = self.p.S0_kernel
import mxnet as mx
xavier_init = mx.init.Xavier(factor_type="avg",
rnd_type="uniform",
magnitude=3)
c2, c3, c4, c5 = data
c6 = X.pool(data=c5, name='C6', kernel=3, stride=2, pad=1) # NOTE
c_features = [c2, c3, c4, c5, c6]
# 0 stage
p0_names = ['S0_P2', 'S0_P3', 'S0_P4', 'S0_P5', 'S0_P6']
p_features = self.get_P0_features(c_features, p0_names, dim_reduced,
xavier_init, norm, S0_kernel)
self.feature_grids.append(p_features)
# stack stage
for i in range(num_stage):
self.get_fused_P_feature(i + 1, dim_reduced, xavier_init, norm)
self.neck = dict(
stride4=self.feature_grids[-1][0], #P2
stride8=self.feature_grids[-1][1], #P3
stride16=self.feature_grids[-1][2], #P4
stride32=self.feature_grids[-1][3], #P5
stride64=self.feature_grids[-1][4], #P6
)
return self.neck
def get_fused_P_feature(p_features, stage, dim_reduced, init, norm):
prefix = "S{}_".format(stage)
with mx.name.Prefix(prefix):
P3_0 = p_features['S{}_P3'.format(stage - 1)] # s8
P4_0 = p_features['S{}_P4'.format(stage - 1)] # s16
P5_0 = p_features['S{}_P5'.format(stage - 1)] # s32
P6_0 = p_features['S{}_P6'.format(stage - 1)] # s64
P7_0 = p_features['S{}_P7'.format(stage - 1)] # s128
# P4_1 = gp(P6_0, P4_0)
P6_0_to_P4 = mx.sym.UpSampling(P6_0,
scale=4,
sample_type='nearest',
name="P6_0_to_P4",
num_args=1)
P6_0_to_P4 = mx.sym.slice_like(P6_0_to_P4, P4_0)
P4_1 = merge_gp(P6_0_to_P4, P4_0, name="gp_P6_0_P4_0")
P4_1 = reluconvbn(P4_1,
dim_reduced,
init,
norm,
name="P4_1",
prefix=prefix)
# P4_2 = sum(P4_0, P4_1)
P4_2 = merge_sum(P4_0, P4_1, name="sum_P4_0_P4_1")
P4_2 = reluconvbn(P4_2,
dim_reduced,
init,
norm,
name="P4_2",
prefix=prefix)
# P3_3 = sum(P4_2, P3_0) end node
P4_2_to_P3 = mx.sym.UpSampling(P4_2,
scale=2,
sample_type='nearest',
name="P4_2_to_P3",
num_args=1)
P4_2_to_P3 = mx.sym.slice_like(P4_2_to_P3, P3_0)
P3_3 = merge_sum(P4_2_to_P3, P3_0, name="sum_P4_2_P3_0")
P3_3 = reluconvbn(P3_3,
dim_reduced,
init,
norm,
name="P3_3",
prefix=prefix)
P3 = P3_3
# P4_4 = sum(P4_2, P3_3) end node
P3_3_to_P4 = X.pool(P3_3,
name="P3_3_to_P4",
kernel=2,
stride=2,
pad=0)
P3_3_to_P4 = mx.sym.slice_like(P3_3_to_P4, P4_2)
P4_4 = merge_sum(P4_2, P3_3_to_P4, name="sum_P4_4_P3_3")
P4_4 = reluconvbn(P4_4,
dim_reduced,
init,
norm,
name="P4_4",
prefix=prefix)
P4 = P4_4
# P5_5 = sum(gp(P4_4, P3_3), P5_0) end node
P4_4_to_P5 = X.pool(P4_4,
kernel=2,
stride=2,
name="P4_4_to_P5",
pad=0)
P4_4_to_P5 = mx.sym.slice_like(P4_4_to_P5, P5_0)
P3_3_to_P5 = X.pool(P3_3,
kernel=4,
stride=4,
name="P3_3_to_P5",
pad=0)
P3_3_to_P5 = mx.sym.slice_like(P3_3_to_P5, P5_0)
gp_P4_4_P3_3 = merge_gp(P4_4_to_P5,
P3_3_to_P5,
name="gp_P4_4_P3_3")
P5_5 = merge_sum(gp_P4_4_P3_3,
P5_0,
name="sum_[gp_P4_4_P3_3]_P5_0")
P5_5 = reluconvbn(P5_5,
dim_reduced,
init,
norm,
name="P5_5",
prefix=prefix)
P5 = P5_5
# P7_6 = sum(gp(P5_5, P4_2), P7_0) end node
P4_2_to_P7 = X.pool(P4_2,
name="P4_2_to_P7",
kernel=8,
stride=8,
pad=0)
P4_2_to_P7 = mx.sym.slice_like(P4_2_to_P7, P7_0)
P5_5_to_P7 = X.pool(P5_5,
name="P5_5_to_P7",
kernel=4,
stride=4,
pad=0)
P5_5_to_P7 = mx.sym.slice_like(P5_5_to_P7, P7_0)
gp_P5_5_P4_2 = merge_gp(P5_5_to_P7,
P4_2_to_P7,
name="gp_P5_5_P4_2")
P7_6 = merge_sum(gp_P5_5_P4_2,
P7_0,
name="sum_[gp_P5_5_P4_2]_P7_0")
P7_6 = reluconvbn(P7_6,
dim_reduced,
init,
norm,
name="P7_6",
prefix=prefix)
P7 = P7_6
# P6_7 = gp(P7_6, P5_5) end node
P7_6_to_P6 = mx.sym.UpSampling(P7_6,
scale=2,
sample_type='nearest',
name="P7_6_to_P6",
num_args=1)
P7_6_to_P6 = mx.sym.slice_like(P7_6_to_P6, P6_0)
P5_5_to_P6 = X.pool(P5_5,
name="p5_5_to_P6",
kernel=2,
stride=2,
pad=0)
P5_5_to_P6 = mx.sym.slice_like(P5_5_to_P6, P6_0)
P6_7 = merge_gp(P7_6_to_P6,
P5_5_to_P6,
name="gp_P7_6_to_P6_P5_5_to_P6")
P6_7 = reluconvbn(P6_7,
dim_reduced,
init,
norm,
name="P6_7",
prefix=prefix)
P6 = P6_7
return {
'S{}_P3'.format(stage): P3,
'S{}_P4'.format(stage): P4,
'S{}_P5'.format(stage): P5,
'S{}_P6'.format(stage): P6,
'S{}_P7'.format(stage): P7
}
def get_fused_P_feature(p_features, stage, dim_reduced, init, norm):
prefix = "S{}_".format(stage)
with mx.name.Prefix(prefix):
P3_0 = p_features['S{}_P3'.format(stage - 1)] # s8
P4_0 = p_features['S{}_P4'.format(stage - 1)] # s16
P5_0 = p_features['S{}_P5'.format(stage - 1)] # s32
P6_0 = p_features['S{}_P6'.format(stage - 1)] # s64
P7_0 = p_features['S{}_P7'.format(stage - 1)] # s128
# P7_1
P7_1 = P7_0
# P6_1 = sum(P6_0, P7_1)
P7_1_to_P6 = mx.sym.UpSampling(P7_1,
scale=2,
sample_type='nearest',
name="P7_1_to_P6",
num_args=1)
P6_1 = X.merge_sum([P6_0, P7_1_to_P6], name="sum_P6_0_P7_1")
P6_1 = X.reluconvbn(P6_1,
dim_reduced,
init,
norm,
name="P6_1",
prefix=prefix)
# P5_1 = sum(P5_0, P6_1)
P6_1_to_P5 = mx.sym.UpSampling(P6_1,
scale=2,
sample_type='nearest',
name="P6_1_to_P5",
num_args=1)
P5_1 = X.merge_sum([P5_0, P6_1_to_P5], name="sum_P5_0_P6_1")
P5_1 = X.reluconvbn(P5_1,
dim_reduced,
init,
norm,
name="P5_1",
prefix=prefix)
# P4_1 = sum(P4_0, P5_1)
P5_1_to_P4 = mx.sym.UpSampling(P5_1,
scale=2,
sample_type='nearest',
name="P5_1_to_P4",
num_args=1)
P4_1 = X.merge_sum([P4_0, P5_1_to_P4], name="sum_P4_0_P5_1")
P4_1 = X.reluconvbn(P4_1,
dim_reduced,
init,
norm,
name="P4_1",
prefix=prefix)
# P3_1 = sum(P3_0, P4_1)
P4_1_to_P3 = mx.sym.UpSampling(P4_1,
scale=2,
sample_type='nearest',
name="P4_1_to_P3",
num_args=1)
P3_1 = X.merge_sum([P3_0, P4_1_to_P3], name="sum_P3_0_P4_1")
P3_1 = X.reluconvbn(P3_1,
dim_reduced,
init,
norm,
name="P3_1",
prefix=prefix)
P3_2 = P3_1
P3 = P3_2
# P4_2 = sum(P3_2, P4_1)
P3_2_to_P4 = X.pool(P3_2,
name="P3_2_to_P4",
kernel=2,
stride=2,
pad=0)
P4_2 = X.merge_sum([P4_1, P3_2_to_P4], name="sum_P4_1_P3_2")
P4_2 = X.reluconvbn(P4_2,
dim_reduced,
init,
norm,
name="P4_2",
prefix=prefix)
P4 = P4_2
# P5_2 = sum(P4_2, P5_1)
P4_2_to_P5 = X.pool(P4_2,
name="P4_2_to_P5",
kernel=2,
stride=2,
pad=0)
P5_2 = X.merge_sum([P5_1, P4_2_to_P5], name="sum_P5_1_P4_2")
P5_2 = X.reluconvbn(P5_2,
dim_reduced,
init,
norm,
name="P5_2",
prefix=prefix)
P5 = P5_2
# P6_2 = sum(P5_2, P6_1)
P5_2_to_P6 = X.pool(P5_2,
name="P5_2_to_P6",
kernel=2,
stride=2,
pad=0)
P6_2 = X.merge_sum([P6_1, P5_2_to_P6], name="sum_P6_1_P5_2")
P6_2 = X.reluconvbn(P6_2,
dim_reduced,
init,
norm,
name="P6_2",
prefix=prefix)
P6 = P6_2
# P7_2 = sum(P6_2, P7_1)
P6_2_to_P7 = X.pool(P6_2,
name="P6_2_to_P7",
kernel=2,
stride=2,
pad=0)
P7_2 = X.merge_sum([P7_1, P6_2_to_P7], name="sum_P7_1_P6_2")
P7_2 = X.reluconvbn(P7_2,
dim_reduced,
init,
norm,
name="P7_2",
prefix=prefix)
P7 = P7_2
return {
'S{}_P3'.format(stage): P3,
'S{}_P4'.format(stage): P4,
'S{}_P5'.format(stage): P5,
'S{}_P6'.format(stage): P6,
'S{}_P7'.format(stage): P7
}
def get_fused_P_feature(p_features, stage, dim_reduced, init, norm):
prefix = "S{}_".format(stage)
with mx.name.Prefix(prefix):
P2_0 = p_features['S{}_P2'.format(stage - 1)] # s4
P3_0 = p_features['S{}_P3'.format(stage - 1)] # s8
P4_0 = p_features['S{}_P4'.format(stage - 1)] # s16
P5_0 = p_features['S{}_P5'.format(stage - 1)] # s32
P6_0 = p_features['S{}_P6'.format(stage - 1)] # s64
P6_1 = P6_0
P6_1_to_P5 = mx.sym.UpSampling(P6_1,
scale=2,
sample_type='nearest',
name="P6_1_to_P5",
num_args=1)
P6_1_to_P5 = mx.sym.slice_like(P6_1_to_P5, P5_0)
P5_1 = X.merge_sum([P5_0, P6_1_to_P5], name="sum_P5_0_P6_1")
P5_1 = X.reluconvbn(P5_1,
dim_reduced,
init,
norm,
name="P5_1",
prefix=prefix)
# P4_1 = sum(P4_0, P5_1)
P5_1_to_P4 = mx.sym.UpSampling(P5_1,
scale=2,
sample_type='nearest',
name="P5_1_to_P4",
num_args=1)
P5_1_to_P4 = mx.sym.slice_like(P5_1_to_P4, P4_0)
P4_1 = X.merge_sum([P4_0, P5_1_to_P4], name="sum_P4_0_P5_1")
P4_1 = X.reluconvbn(P4_1,
dim_reduced,
init,
norm,
name="P4_1",
prefix=prefix)
P4_1_to_P3 = mx.sym.UpSampling(P4_1,
scale=2,
sample_type='nearest',
name="P4_1_to_P3",
num_args=1)
P4_1_to_P3 = mx.sym.slice_like(P4_1_to_P3, P3_0)
P3_1 = X.merge_sum([P3_0, P4_1_to_P3], name="sum_P3_0_P4_1")
P3_1 = X.reluconvbn(P3_1,
dim_reduced,
init,
norm,
name="P3_1",
prefix=prefix)
P3_1_to_P2 = mx.sym.UpSampling(P3_1,
scale=2,
sample_type='nearest',
name="P3_1_to_P2",
num_args=1)
P3_1_to_P2 = mx.sym.slice_like(P3_1_to_P2, P2_0)
P2_1 = X.merge_sum([P2_0, P3_1_to_P2], name="sum_P2_0_P3_1")
P2_1 = X.reluconvbn(P2_1,
dim_reduced,
init,
norm,
name="P2_1",
prefix=prefix)
P2_2 = P2_1
P2 = P2_2
P2_2_to_P3 = X.pool(P2_2,
name="P2_2_to_P3",
kernel=3,
stride=2,
pad=1)
P3_2 = X.merge_sum([P3_1, P2_2_to_P3], name="sum_P3_1_P2_2")
P3_2 = X.reluconvbn(P3_2,
dim_reduced,
init,
norm,
name="P3_2",
prefix=prefix)
P3 = P3_2
P3_2_to_P4 = X.pool(P3_2,
name="P3_2_to_P4",
kernel=3,
stride=2,
pad=1)
P4_2 = X.merge_sum([P4_1, P3_2_to_P4], name="sum_P4_1_P3_2")
P4_2 = X.reluconvbn(P4_2,
dim_reduced,
init,
norm,
name="P4_2",
prefix=prefix)
P4 = P4_2
P4_2_to_P5 = X.pool(P4_2,
name="P4_2_to_P5",
kernel=3,
stride=2,
pad=1)
P5_2 = X.merge_sum([P5_1, P4_2_to_P5], name="sum_P5_1_P4_2")
P5_2 = X.reluconvbn(P5_2,
dim_reduced,
init,
norm,
name="P5_2",
prefix=prefix)
P5 = P5_2
P5_2_to_P6 = X.pool(P5_2,
name="P5_2_to_P6",
kernel=3,
stride=2,
pad=1)
P6_2 = X.merge_sum([P6_1, P5_2_to_P6], name="sum_P6_1_P5_2")
P6_2 = X.reluconvbn(P6_2,
dim_reduced,
init,
norm,
name="P6_2",
prefix=prefix)
P6 = P6_2
return {
'S{}_P2'.format(stage): P2,
'S{}_P3'.format(stage): P3,
'S{}_P4'.format(stage): P4,
'S{}_P5'.format(stage): P5,
'S{}_P6'.format(stage): P6,
}
def fpn_conv_down(self, data):
if self.fpn_feat:
return self.fpn_feat
c2, c3, c4, c5 = data
if self.p.fp16:
c2 = X.to_fp32(c2, name="c2_to_fp32")
c3 = X.to_fp32(c3, name="c3_to_fp32")
c4 = X.to_fp32(c4, name="c4_to_fp32")
c5 = X.to_fp32(c5, name="c5_to_fp32")
xavier_init = mx.init.Xavier(factor_type="in",
rnd_type="uniform",
magnitude=3)
# P5
p5 = X.conv(data=c5,
filter=256,
no_bias=False,
weight=X.var(name="P5_lateral_weight", init=xavier_init),
bias=X.var(name="P5_lateral_bias", init=X.zero_init()),
name="P5_lateral")
p5_conv = X.conv(data=p5,
kernel=3,
filter=256,
no_bias=False,
weight=X.var(name="P5_conv_weight", init=xavier_init),
bias=X.var(name="P5_conv_bias", init=X.zero_init()),
name="P5_conv")
# P4
p5_up = mx.sym.UpSampling(p5,
scale=2,
sample_type="nearest",
name="P5_upsampling",
num_args=1)
p4_la = X.conv(data=c4,
filter=256,
no_bias=False,
weight=X.var(name="P4_lateral_weight",
init=xavier_init),
bias=X.var(name="P4_lateral_bias", init=X.zero_init()),
name="P4_lateral")
p5_clip = mx.sym.Crop(*[p5_up, p4_la], name="P4_clip")
p4 = mx.sym.ElementWiseSum(*[p5_clip, p4_la], name="P4_sum")
p4_conv = X.conv(data=p4,
kernel=3,
filter=256,
no_bias=False,
weight=X.var(name="P4_conv_weight", init=xavier_init),
bias=X.var(name="P4_conv_bias", init=X.zero_init()),
name="P4_conv")
# P3
p4_up = mx.sym.UpSampling(p4,
scale=2,
sample_type="nearest",
name="P4_upsampling",
num_args=1)
p3_la = X.conv(data=c3,
filter=256,
no_bias=False,
weight=X.var(name="P3_lateral_weight",
init=xavier_init),
bias=X.var(name="P3_lateral_bias", init=X.zero_init()),
name="P3_lateral")
p4_clip = mx.sym.Crop(*[p4_up, p3_la], name="P3_clip")
p3 = mx.sym.ElementWiseSum(*[p4_clip, p3_la], name="P3_sum")
p3_conv = X.conv(data=p3,
kernel=3,
filter=256,
no_bias=False,
weight=X.var(name="P3_conv_weight", init=xavier_init),
bias=X.var(name="P3_conv_bias", init=X.zero_init()),
name="P3_conv")
# P2
p3_up = mx.sym.UpSampling(p3,
scale=2,
sample_type="nearest",
name="P3_upsampling",
num_args=1)
p2_la = X.conv(data=c2,
filter=256,
no_bias=False,
weight=X.var(name="P2_lateral_weight",
init=xavier_init),
bias=X.var(name="P2_lateral_bias", init=X.zero_init()),
name="P2_lateral")
p3_clip = mx.sym.Crop(*[p3_up, p2_la], name="P2_clip")
p2 = mx.sym.ElementWiseSum(*[p3_clip, p2_la], name="P2_sum")
p2_conv = X.conv(data=p2,
kernel=3,
filter=256,
no_bias=False,
weight=X.var(name="P2_conv_weight", init=xavier_init),
bias=X.var(name="P2_conv_bias", init=X.zero_init()),
name="P2_conv")
# P6
p6 = X.pool(p5_conv,
name="P6_subsampling",
kernel=1,
stride=2,
pad=0,
pool_type='max')
if self.p.fp16:
p6 = X.to_fp16(p6, name="p6_to_fp16")
p5_conv = X.to_fp16(p5_conv, name="p5_conv_to_fp16")
p4_conv = X.to_fp16(p4_conv, name="p4_conv_to_fp16")
p3_conv = X.to_fp16(p3_conv, name="p3_conv_to_fp16")
p2_conv = X.to_fp16(p2_conv, name="p2_conv_to_fp16")
conv_fpn_feat = dict()
conv_fpn_feat.update({
"stride64": p6,
"stride32": p5_conv,
"stride16": p4_conv,
"stride8": p3_conv,
"stride4": p2_conv
})
self.fpn_feat = conv_fpn_feat
return self.fpn_feat
