def __init__(self, pRpn):
super().__init__(pRpn)
# init bias for cls
prior_prob = 0.01
pi = -math.log((1-prior_prob) / prior_prob)
# shared classification weight and bias
self.cls_conv1_weight = X.var("cls_conv1_weight", init=X.gauss(std=0.01))
self.cls_conv1_bias = X.var("cls_conv1_bias", init=X.zero_init())
self.cls_conv2_weight = X.var("cls_conv2_weight", init=X.gauss(std=0.01))
self.cls_conv2_bias = X.var("cls_conv2_bias", init=X.zero_init())
self.cls_conv3_weight = X.var("cls_conv3_weight", init=X.gauss(std=0.01))
self.cls_conv3_bias = X.var("cls_conv3_bias", init=X.zero_init())
self.cls_conv4_weight = X.var("cls_conv4_weight", init=X.gauss(std=0.01))
self.cls_conv4_bias = X.var("cls_conv4_bias", init=X.zero_init())
self.cls_pred_weight = X.var("cls_pred_weight", init=X.gauss(std=0.01))
self.cls_pred_bias = X.var("cls_pred_bias", init=X.constant(pi))
# shared regression weight and bias
self.bbox_conv1_weight = X.var("bbox_conv1_weight", init=X.gauss(std=0.01))
self.bbox_conv1_bias = X.var("bbox_conv1_bias", init=X.zero_init())
self.bbox_conv2_weight = X.var("bbox_conv2_weight", init=X.gauss(std=0.01))
self.bbox_conv2_bias = X.var("bbox_conv2_bias", init=X.zero_init())
self.bbox_conv3_weight = X.var("bbox_conv3_weight", init=X.gauss(std=0.01))
self.bbox_conv3_bias = X.var("bbox_conv3_bias", init=X.zero_init())
self.bbox_conv4_weight = X.var("bbox_conv4_weight", init=X.gauss(std=0.01))
self.bbox_conv4_bias = X.var("bbox_conv4_bias", init=X.zero_init())
self.bbox_pred_weight = X.var("bbox_pred_weight", init=X.gauss(std=0.01))
self.bbox_pred_bias = X.var("bbox_pred_bias", init=X.zero_init())
self._cls_logit_dict = None
self._bbox_delta_dict = None
def __init__(self, pBbox):
super().__init__(pBbox)
self.stage = pBbox.stage
self._cls_logit = None
self._bbox_delta = None
self._proposal = None
# declare weight and bias
stage = self.stage
xavier_init = mx.init.Xavier(factor_type="in",
rnd_type="uniform",
magnitude=3)
self.fc1_weight = X.var("bbox_fc1_%s_weight" % stage, init=xavier_init)
self.fc2_weight = X.var("bbox_fc2_%s_weight" % stage, init=xavier_init)
self.cls_logit_weight = X.var("bbox_cls_logit_%s_weight" % stage,
init=X.gauss(0.01))
self.bbox_delta_weight = X.var("bbox_reg_delta_%s_weight" % stage,
init=X.gauss(0.001))
self.fc1_bias = X.var("bbox_fc1_%s_bias" % stage)
self.fc2_bias = X.var("bbox_fc2_%s_bias" % stage)
self.cls_logit_bias = X.var("bbox_cls_logit_%s_bias" % stage)
self.bbox_delta_bias = X.var("bbox_reg_delta_%s_bias" % stage)
def get_output(self, conv_feat):
p = self.p
num_class = p.num_class
num_reg_class = 2 if p.regress_target.class_agnostic else num_class
head_feat = self._get_bbox_head_logit(conv_feat)
if not isinstance(head_feat, dict):
head_feat = dict(classification=head_feat, regression=head_feat)
if p.fp16:
head_feat["classification"] = X.to_fp32(head_feat["classification"], name="bbox_cls_head_to_fp32")
head_feat["regression"] = X.to_fp32(head_feat["regression"], name="bbox_reg_head_to_fp32")
cls_logit = X.fc(
head_feat["classification"],
filter=num_class,
name='bbox_cls_logit',
init=X.gauss(0.01)
)
bbox_delta = X.fc(
head_feat["regression"],
filter=4 * num_reg_class,
name='bbox_reg_delta',
init=X.gauss(0.001)
)
return cls_logit, bbox_delta
def get_output(self, conv_feat):
if self._cls_logit is not None and self._bbox_delta is not None:
return self._cls_logit, self._bbox_delta
p = self.p
num_base_anchor = len(p.anchor_generate.ratio) * len(
p.anchor_generate.scale)
conv_channel = p.head.conv_channel
conv = X.convrelu(conv_feat,
kernel=3,
filter=conv_channel,
name="rpn_conv_3x3",
no_bias=False,
init=X.gauss(0.01))
if p.fp16:
conv = X.to_fp32(conv, name="rpn_conv_3x3_fp32")
cls_logit = X.conv(conv,
filter=2 * num_base_anchor,
name="rpn_cls_logit",
no_bias=False,
init=X.gauss(0.01))
bbox_delta = X.conv(conv,
filter=4 * num_base_anchor,
name="rpn_bbox_delta",
no_bias=False,
init=X.gauss(0.01))
self._cls_logit = cls_logit
self._bbox_delta = bbox_delta
return self._cls_logit, self._bbox_delta
def get_output(self, conv_fpn_feat):
if self.cls_logit_dict is not None and self.bbox_delta_dict is not None:
return self.cls_logit_dict, self.bbox_delta_dict
p = self.p
num_base_anchor = len(p.anchor_generate.ratio) * len(p.anchor_generate.scale)
conv_channel = p.head.conv_channel
# FPN RPN share weight
rpn_conv_weight = X.var('rpn_conv_weight', init=X.gauss(0.01))
rpn_conv_bias = X.var('rpn_conv_bias', init=X.zero_init())
rpn_conv_cls_weight = X.var('rpn_conv_cls_weight', init=X.gauss(0.01))
rpn_conv_cls_bias = X.var('rpn_conv_cls_bias', init=X.zero_init())
rpn_conv_bbox_weight = X.var('rpn_conv_bbox_weight', init=X.gauss(0.01))
rpn_conv_bbox_bias = X.var('rpn_conv_bbox_bias', init=X.zero_init())
cls_logit_dict = {}
bbox_delta_dict = {}
for stride in p.anchor_generate.stride:
rpn_conv = X.conv(
conv_fpn_feat['stride%s' % stride],
kernel=3,
filter=conv_channel,
name="rpn_conv_3x3_%s" % stride,
no_bias=False,
weight=rpn_conv_weight,
bias=rpn_conv_bias
)
rpn_relu = X.relu(rpn_conv, name='rpn_relu_%s' % stride)
if p.fp16:
rpn_relu = X.to_fp32(rpn_relu, name="rpn_relu_%s_fp32" % stride)
cls_logit = X.conv(
rpn_relu,
filter=2 * num_base_anchor,
name="rpn_cls_score_stride%s" % stride,
no_bias=False,
weight=rpn_conv_cls_weight,
bias=rpn_conv_cls_bias
)
bbox_delta = X.conv(
rpn_relu,
filter=4 * num_base_anchor,
name="rpn_bbox_pred_stride%s" % stride,
no_bias=False,
weight=rpn_conv_bbox_weight,
bias=rpn_conv_bbox_bias
)
cls_logit_dict[stride] = cls_logit
bbox_delta_dict[stride] = bbox_delta
self.cls_logit_dict = cls_logit_dict
self.bbox_delta_dict = bbox_delta_dict
return self.cls_logit_dict, self.bbox_delta_dict
def PConvModule(x, out_channels=256, kernel_size=[3, 3, 3], dilation=[1, 1, 1], groups=[1, 1, 1], ibn=None,
part_deform=False, PConv_idx=-1, start_level=1, norm=None, bilinear_upsample=None, feat_sizes=None):
assert PConv_idx > -1 and feat_sizes is not None
name_pref = 'PConv{}_sepc'.format(PConv_idx)
sepc0_weight, sepc0_bias = X.var(name=name_pref+'0_weight', init=X.gauss(std=0.01)), X.var(name=name_pref+'0_bias', init=X.zero_init())
sepc1_weight, sepc1_bias = X.var(name=name_pref+'1_weight', init=X.gauss(std=0.01)), X.var(name=name_pref+'1_bias', init=X.zero_init())
sepc2_weight, sepc2_bias = X.var(name=name_pref+'2_weight', init=X.gauss(std=0.01)), X.var(name=name_pref+'2_bias', init=X.zero_init())
sepc0_offset_weight, sepc0_offset_bias = None, None
sepc1_offset_weight, sepc1_offset_bias = None, None
sepc2_offset_weight, sepc2_offset_bias = None, None
if part_deform:
# NOTE zero_init for offset's weight and bias
sepc0_offset_weight, sepc0_offset_bias = X.var(name=name_pref+'0_offset_weight', init=X.zero_init()), X.var(name=name_pref+'0_offset_bias', init=X.zero_init())
sepc1_offset_weight, sepc1_offset_bias = X.var(name=name_pref+'1_offset_weight', init=X.zero_init()), X.var(name=name_pref+'1_offset_bias', init=X.zero_init())
sepc2_offset_weight, sepc2_offset_bias = X.var(name=name_pref+'2_offset_weight', init=X.zero_init()), X.var(name=name_pref+'2_offset_bias', init=X.zero_init())
norm_func = []
if ibn:
assert norm is not None
norm_func = partial(norm, name=name_pref+'_ibn')
sepc_conv0_func = partial(
sepc_conv, name='PConv{}_sepc0_'.format(PConv_idx), out_channels=out_channels,
kernel_size=kernel_size[0], stride=1, padding=(kernel_size[0]+(dilation[0]-1)*2)//2,
dilation=dilation[0], groups=groups[0], deformable_groups=1, part_deform=part_deform, start_level=start_level,
weight=sepc0_weight, bias=sepc0_bias, weight_offset=sepc0_offset_weight, bias_offset=sepc0_offset_bias)
sepc_conv1_func = partial(
sepc_conv, name='PConv{}_sepc1_'.format(PConv_idx), out_channels=out_channels,
kernel_size=kernel_size[1], stride=1, padding=(kernel_size[1]+(dilation[1]-1)*2)//2,
dilation=dilation[1], groups=groups[1], deformable_groups=1, part_deform=part_deform, start_level=start_level,
weight=sepc1_weight, bias=sepc1_bias, weight_offset=sepc1_offset_weight, bias_offset=sepc1_offset_bias)
sepc_conv2_func = partial(
sepc_conv, name='PConv{}_sepc2_'.format(PConv_idx), out_channels=out_channels,
kernel_size=kernel_size[2], stride=2, padding=(kernel_size[2]+(dilation[2]-1)*2)//2,
dilation=dilation[2], groups=groups[2], deformable_groups=1, part_deform=part_deform, start_level=start_level,
weight=sepc2_weight, bias=sepc2_bias, weight_offset=sepc2_offset_weight, bias_offset=sepc2_offset_bias)
next_x = []
for level, feature in enumerate(x):
temp_fea = sepc_conv1_func(i=level, x=feature)
if level > 0:
tmp = sepc_conv2_func(i=level, x=x[level - 1])
temp_fea = temp_fea + tmp
if level < len(x) - 1:
tmp_x = sepc_conv0_func(i=level,x=x[level+1])
if bilinear_upsample:
tmp_x = mx.contrib.symbol.BilinearResize2D(tmp_x, scale_height=2, scale_width=2,
name='PConv{}_upsampling_level{}'.format(PConv_idx,level))
else:
tmp_x = mx.sym.UpSampling(tmp_x, scale=2, sample_type='nearest', num_args=1,
name='PConv{}_upsampling_level{}'.format(PConv_idx,level))
tmp_x = mx.sym.slice_like(tmp_x, temp_fea)
temp_fea = temp_fea + tmp_x
next_x.append(temp_fea)
if ibn:
next_x = ibn_func(next_x, norm_func, feat_sizes)
next_x = [relu(item, name='PConv{}_level{}_relu'.format(PConv_idx, level)) for level,item in enumerate(next_x)]
return next_x
def get_output(self, conv_feat):
if self._cls_logit is not None and self._bbox_delta is not None:
return self._cls_logit, self._bbox_delta
p = self.p
num_base_anchor = len(p.anchor_generate.ratio) * len(p.anchor_generate.scale)
conv_channel = p.head.conv_channel
if p.normalizer.__name__ == "fix_bn":
conv = X.convrelu(
conv_feat,
kernel=3,
filter=conv_channel,
name="rpn_conv_3x3",
no_bias=False,
init=X.gauss(0.01)
)
elif p.normalizer.__name__ in ["sync_bn", "gn"]:
conv = X.convnormrelu(
p.normalizer,
conv_feat,
kernel=3,
filter=conv_channel,
name="rpn_conv_3x3",
no_bias=False,
init=X.gauss(0.01)
)
else:
raise NotImplementedError("Unsupported normalizer: {}".format(p.normalizer.__name__))
if p.fp16:
conv = X.to_fp32(conv, name="rpn_conv_3x3_fp32")
cls_logit = X.conv(
conv,
filter=2 * num_base_anchor,
name="rpn_cls_logit",
no_bias=False,
init=X.gauss(0.01)
)
bbox_delta = X.conv(
conv,
filter=4 * num_base_anchor,
name="rpn_bbox_delta",
no_bias=False,
init=X.gauss(0.01)
)
self._cls_logit = cls_logit
self._bbox_delta = bbox_delta
return self._cls_logit, self._bbox_delta
def get_rcnn_feature(self, rcnn_feat):
p = self.pNeck
conv_channel = p.conv_channel
conv_neck = X.convrelu(rcnn_feat,
filter=conv_channel,
no_bias=False,
init=X.gauss(0.01),
name="conv_neck")
return conv_neck
def SEPCFPN(inputs, out_channels=256, pconv_deform=False, lcconv_deform=None, ibn=None, Pconv_num=4,
start_level=1, norm=None, bilinear_upsample=None, feat_sizes=None):
assert feat_sizes is not None
Pconvs_list = []
for i in range(Pconv_num):
Pconvs_list.append(partial(
PConvModule, out_channels=out_channels, ibn=ibn, part_deform=pconv_deform,
PConv_idx=i, start_level=start_level, norm=norm, bilinear_upsample=bilinear_upsample, feat_sizes=feat_sizes))
if lcconv_deform is not None:
assert lcconv_deform in [False, True]
lconv_weight, lconv_bias = X.var(name='LConv_weight', init=X.gauss(std=0.01)), X.var(name='LConv_bias',init=X.zero_init())
cconv_weight, cconv_bias = X.var(name='CConv_weight', init=X.gauss(std=0.01)), X.var(name='CConv_bias',init=X.zero_init())
lconv_offset_weight, lconv_offset_bias = None, None
cconv_offset_weight, cconv_offset_bias = None, None
if lcconv_deform:
lconv_offset_weight, lconv_offset_bias=X.var(name='LConv_offset_weight', init=X.zero_init()), X.var(name='LConv_offset_bias', init=X.zero_init())
cconv_offset_weight, cconv_offset_bias=X.var(name='CConv_offset_weight', init=X.zero_init()), X.var(name='CConv_offset_bias', init=X.zero_init())
lconv_func = partial(sepc_conv, name='LConv{}_',out_channels=out_channels, kernel_size=3, stride=1, padding=1,
dilation=1, groups=1, deformable_groups=1, part_deform=lcconv_deform, start_level=start_level,
weight=lconv_weight, bias=lconv_bias, weight_offset=lconv_offset_weight, bias_offset=lconv_offset_bias)
cconv_func = partial(sepc_conv, name='CConv{}_', out_channels=out_channels, kernel_size=3, stride=1, padding=1,
dilation=1, groups=1, deformable_groups=1, part_deform=lcconv_deform, start_level=start_level,
weight=cconv_weight, bias=cconv_bias, weight_offset=cconv_offset_weight, bias_offset=cconv_offset_bias)
if ibn:
assert norm is not None
lbn = partial(norm, name='lconv_ibn')
cbn = partial(norm, name='cconv_ibn')
x = inputs
for pconv in Pconvs_list:
x = pconv(x)
if lcconv_deform is None:
return x
cls_outs = [cconv_func(i=level, x=item) for level, item in enumerate(x)]
loc_outs = [lconv_func(i=level, x=item) for level, item in enumerate(x)]
if ibn:
cls_outs = ibn_func(cls_outs, cbn, feat_sizes)
loc_outs = ibn_func(loc_outs, lbn, feat_sizes)
outs = [mx.sym.Concat(*[relu(s), relu(l)], num_args=2, dim=1) for s, l in zip(cls_outs, loc_outs)]
return outs
def get_output(self, conv_feat):
p = self.p
num_class = p.num_class
num_reg_class = 2 if p.regress_target.class_agnostic else num_class
head_feat = self._get_bbox_head_logit(conv_feat)
if p.fp16:
head_feat = X.to_fp32(head_feat, name="bbox_head_to_fp32")
cls_logit = X.fc(head_feat,
filter=num_class,
name='bbox_cls_logit',
init=X.gauss(0.01))
bbox_delta = X.fc(head_feat,
filter=4 * num_reg_class,
name='bbox_reg_delta',
init=X.gauss(0.001))
return cls_logit, bbox_delta
def __init__(self, pBbox):
super(CascadeBbox2fcHead, self).__init__(pBbox)
self.stage = pBbox.stage
self._cls_logit = None
self._bbox_delta = None
self._proposal = None
# for stage '1st_3rd', using weight from 1st stage
weight_stage = self.stage.split('_')[0]
self.fc1_weight = X.var("bbox_fc1_" + weight_stage + "_weight")
self.fc2_weight = X.var("bbox_fc2_" + weight_stage + "_weight")
self.cls_logit_weight = X.var("bbox_cls_logit_" + weight_stage +
"_weight",
init=X.gauss(0.01))
self.cls_logit_bias = X.var("bbox_cls_logit_" + weight_stage + "_bias")
self.bbox_delta_weight = X.var("bbox_reg_delta_" + weight_stage +
"_weight",
init=X.gauss(0.001))
self.bbox_delta_bias = X.var("bbox_reg_delta_" + weight_stage +
"_bias")
def _reg_head(self, conv_feat):
num_block = self.p.num_block or 4
for i in range(num_block):
conv_feat = X.conv(conv_feat,
kernel=3,
filter=256,
init=X.gauss(0.01),
name="bbox_reg_block%s" % (i + 1))
conv_feat = self.add_norm(conv_feat)
conv_feat = X.relu(conv_feat)
return conv_feat
def get_output(self, fpn_conv_feats, roi_feat, rois, is_train):
'''
Args:
fpn_conv_feats: dict of FPN features, each [batch_image, in_channels, fh, fw]
roi_feat: [batch_image * image_roi, 256, roi_size, roi_size]
rois: [batch_image, image_roi, 4]
is_train: boolean
Returns:
cls_logit: [batch_image * image_roi, num_class]
bbox_delta: [batch_image * image_roi, num_class * 4]
tsd_cls_logit: [batch_image * image_roi, num_class]
tsd_bbox_delta: [batch_image * image_roi, num_class * 4]
delta_c: [batch_image * image_roi, 2*roi_size*roi_size, 1, 1]
delta_r: [batch_image * image_roi, 2, 1, 1]
'''
xavier_init = mx.init.Xavier(factor_type="in",
rnd_type="uniform",
magnitude=3)
# roi_feat: [batch_roi, 256, 7, 7]
flatten = X.reshape(
roi_feat, shape=(0, -1, 1, 1),
name="bbox_feat_reshape") # [batch_roi, 256*7*7, 1, 1]
x1 = flatten
x2 = X.relu(X.conv(data=x1,
kernel=1,
filter=256,
name="delta_shared_fc1",
no_bias=False),
name="delta_shared_fc1_relu") # [batch_roi, 256, 1, 1]
delta_c = X.relu(X.conv(x2,
filter=256,
name="delta_c_fc1",
init=X.gauss(0.01)),
name="delta_c_fc1_relu") # [batch_roi, 256, 1, 1]
delta_c = X.conv(delta_c,
filter=2 * self.p.roi_size**2,
name="delta_c_fc2",
init=X.gauss(0.01)) # [batch_roi, 2*7*7, 1, 1]
delta_r = X.relu(X.conv(x2,
filter=256,
name="delta_r_fc1",
init=X.gauss(0.01)),
name="delta_r_fc1_relu") # [batch_roi, 256, 1, 1]
delta_r = X.conv(delta_r,
filter=2,
name="delta_r_fc2",
init=X.gauss(0.01)) # [batch_roi, 2, 1, 1]
image_roi = self.p.image_roi if is_train else 1000
batch_image = self.p.batch_image
TSD_cls_feats = self.delta_c_pool.get_roi_feature(
fpn_conv_feats,
rois,
delta_c,
image_rois=image_roi,
batch_image=batch_image) # [batch_roi, 256, 7, 7]
TSD_loc_feats = self.delta_r_pool.get_roi_feature(
fpn_conv_feats,
rois,
delta_r,
image_rois=image_roi,
batch_image=batch_image) # [batch_roi, 256, 7, 7]
TSD_x_cls = self._convs_and_fcs(
TSD_cls_feats,
self.p.TSD.num_shared_convs,
self.p.TSD.num_shared_fcs,
name='TSD_pc',
conv_init=xavier_init,
fc_init=X.gauss(0.01)) # [batch_roi, batch_roi, 1, 1]
TSD_x_reg = self._convs_and_fcs(
TSD_loc_feats,
self.p.TSD.num_shared_convs,
self.p.TSD.num_shared_fcs,
name='TSD_pr',
conv_init=xavier_init,
fc_init=X.gauss(0.01)) # [batch_roi, batch_roi, 1, 1]
TSD_x_cls = self._convs_and_fcs(
TSD_x_cls,
0,
self.p.TSD.num_cls_fcs,
name='TSD_cls',
conv_init=xavier_init,
fc_init=X.gauss(0.01)) # [batch_roi, batch_roi, 1, 1]
TSD_x_reg = self._convs_and_fcs(
TSD_x_reg,
0,
self.p.TSD.num_reg_fcs,
name='TSD_reg',
conv_init=xavier_init,
fc_init=X.gauss(0.01)) # [batch_roi, batch_roi, 1, 1]
num_class = self.p.num_class
num_reg_class = 2 if self.p.regress_target.class_agnostic else num_class
tsd_cls_logit = X.fc(TSD_x_cls,
filter=num_class,
name='tsd_cls_logit',
init=X.gauss(0.01))
tsd_bbox_delta = X.fc(TSD_x_reg,
filter=4 * num_reg_class,
name='tsd_reg_delta',
init=X.gauss(0.01))
x = self._convs_and_fcs(roi_feat,
self.p.TSD.num_shared_convs,
self.p.TSD.num_shared_fcs,
name='shared_fc',
conv_init=xavier_init,
fc_init=X.gauss(0.01))
x_cls = x
x_reg = x
x_cls = self._convs_and_fcs(x_cls,
0,
self.p.TSD.num_cls_fcs,
name='cls',
conv_init=xavier_init,
fc_init=X.gauss(0.01))
x_reg = self._convs_and_fcs(x_reg,
0,
self.p.TSD.num_reg_fcs,
name='reg',
conv_init=xavier_init,
fc_init=X.gauss(0.01))
cls_logit = X.fc(x_cls,
filter=num_class,
name='bbox_cls_logit',
init=X.gauss(0.01))
bbox_delta = X.fc(x_reg,
filter=4 * num_reg_class,
name='bbox_reg_delta',
init=X.gauss(0.01))
if self.p.fp16:
cls_logit = X.to_fp32(cls_logit, name="cls_logits_fp32")
bbox_delta = X.to_fp32(bbox_delta, name="bbox_delta_fp32")
tsd_cls_logit = X.to_fp32(tsd_cls_logit, name="tsd_cls_logit_fp32")
tsd_bbox_delta = X.to_fp32(tsd_bbox_delta,
name="tsd_bbox_delta_fp32")
delta_c = X.to_fp32(delta_c, name="delta_c_fp32")
delta_r = X.to_fp32(delta_r, name="delta_r_fp32")
return cls_logit, bbox_delta, tsd_cls_logit, tsd_bbox_delta, delta_c, delta_r
def get_output(self, conv_fpn_feat):
p = self.p
centerness_logit_dict = {}
cls_logit_dict = {}
offset_logit_dict = {}
# heads are shared across stages
shared_conv1_w = X.var(name="shared_conv1_3x3_weight", init=X.gauss(0.01))
shared_conv1_b = X.var(name="shared_conv1_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
shared_conv2_w = X.var(name="shared_conv2_3x3_weight", init=X.gauss(0.01))
shared_conv2_b = X.var(name="shared_conv2_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
shared_conv3_w = X.var(name="shared_conv3_3x3_weight", init=X.gauss(0.01))
shared_conv3_b = X.var(name="shared_conv3_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
shared_conv4_w = X.var(name="shared_conv4_3x3_weight", init=X.gauss(0.01))
shared_conv4_b = X.var(name="shared_conv4_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
centerness_conv_w = X.var(name="centerness_conv_3x3_weight", init=X.gauss(0.01))
centerness_conv_b = X.var(name="centerness_conv_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
cls_conv_w = X.var(name="cls_conv_3x3_weight", init=X.gauss(0.01))
cls_conv_b = X.var(name="cls_conv_3x3_bias", init=X.constant(-math.log(99)), lr_mult=2, wd_mult=0) # init with -log((1-0.01)/0.01)
offset_conv1_w = X.var(name="offset_conv1_3x3_weight", init=X.gauss(0.01))
offset_conv1_b = X.var(name="offset_conv1_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
offset_conv2_w = X.var(name="offset_conv2_3x3_weight", init=X.gauss(0.01))
offset_conv2_b = X.var(name="offset_conv2_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
offset_conv3_w = X.var(name="offset_conv3_3x3_weight", init=X.gauss(0.01))
offset_conv3_b = X.var(name="offset_conv3_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
offset_conv4_w = X.var(name="offset_conv4_3x3_weight", init=X.gauss(0.01))
offset_conv4_b = X.var(name="offset_conv4_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
offset_conv5_w = X.var(name="offset_conv5_3x3_weight", init=X.gauss(0.01))
offset_conv5_b = X.var(name="offset_conv5_3x3_bias", init=X.zero_init(), lr_mult=2, wd_mult=0)
for stride in p.FCOSParam.stride:
# centerness & cls shared layer
shared_conv1 = X.conv(
conv_fpn_feat['stride%s' % stride],
kernel=3, filter=256, no_bias=False,
name="shared_conv1_3x3_%s" % stride,
weight=shared_conv1_w,
bias=shared_conv1_b,
)
shared_gn1 = X.gn(shared_conv1, name='shared_gn1_3x3_%s' % stride, num_group=32)
shared_relu1 = X.relu(shared_gn1, name='shared_relu1_3x3_%s' % stride)
shared_conv2 = X.conv(
shared_relu1,
kernel=3, filter=256, no_bias=False,
name="shared_conv2_3x3_%s" % stride,
weight=shared_conv2_w,
bias=shared_conv2_b,
)
shared_gn2 = X.gn(shared_conv2, name='shared_gn2_3x3_%s' % stride, num_group=32)
shared_relu2 = X.relu(shared_gn2, name='shared_relu2_3x3_%s' % stride)
shared_conv3 = X.conv(
shared_relu2,
kernel=3, filter=256, no_bias=False,
name="shared_conv3_3x3_%s" % stride,
weight=shared_conv3_w,
bias=shared_conv3_b,
)
shared_gn3 = X.gn(shared_conv3, name='shared_gn3_3x3_%s' % stride, num_group=32)
shared_relu3 = X.relu(shared_gn3, name='shared_relu3_3x3_%s' % stride)
shared_conv4 = X.conv(
shared_relu3,
kernel=3, filter=256, no_bias=False,
name="shared_conv4_3x3_%s" % stride,
weight=shared_conv4_w,
bias=shared_conv4_b,
)
shared_gn4 = X.gn(shared_conv4, name='shared_gn4_3x3_%s' % stride, num_group=32)
shared_relu4 = X.relu(shared_gn4, name='shared_relu4_3x3_%s' % stride)
# centerness head
center_logit = X.conv(
shared_relu4,
kernel=3,
filter=1,
name="center_conv_3x3_%s" % stride,
no_bias=False,
weight=centerness_conv_w,
bias=centerness_conv_b,
)
# cls head
cls_logit = X.conv(
shared_relu4,
kernel=3,
filter=p.FCOSParam.num_classifier, # remove bg channel
name="cls_conv_3x3_%s" % stride,
no_bias=False,
weight=cls_conv_w,
bias=cls_conv_b,
)
# offset head
offset_conv1 = X.conv(
conv_fpn_feat['stride%s' % stride],
kernel=3,
filter=256,
name="offset_conv1_3x3_%s" % stride,
no_bias=False,
weight=offset_conv1_w,
bias=offset_conv1_b,
)
offset_gn1 = X.gn(offset_conv1, name='offset_gn1_3x3_%s' % stride, num_group=32)
offset_relu1 = X.relu(offset_gn1, name='offset_relu1_3x3_%s' % stride)
offset_conv2 = X.conv(
offset_relu1,
kernel=3,
filter=256,
name="offset_conv2_3x3_%s" % stride,
no_bias=False,
weight=offset_conv2_w,
bias=offset_conv2_b,
)
offset_gn2 = X.gn(offset_conv2, name='offset_gn2_3x3_%s' % stride, num_group=32)
offset_relu2 = X.relu(offset_gn2, name='offset_relu2_3x3_%s' % stride)
offset_conv3 = X.conv(
offset_relu2,
kernel=3,
filter=256,
name="offset_conv3_3x3_%s" % stride,
no_bias=False,
weight=offset_conv3_w,
bias=offset_conv3_b,
)
offset_gn3 = X.gn(offset_conv3, name='offset_gn3_3x3_%s' % stride, num_group=32)
offset_relu3 = X.relu(offset_gn3, name='offset_relu3_3x3_%s' % stride)
offset_conv4 = X.conv(
offset_relu3,
kernel=3,
filter=256,
name="offset_conv1_3x3_%s" % stride,
no_bias=False,
weight=offset_conv4_w,
bias=offset_conv4_b,
)
offset_gn4 = X.gn(offset_conv4, name='offset_gn4_3x3_%s' % stride, num_group=32)
offset_relu4 = X.relu(offset_gn4, name='offset_relu4_3x3_%s' % stride)
offset_logit = X.conv(
offset_relu4,
kernel=3,
filter=4,
name="offset_conv5_3x3_%s" % stride,
no_bias=False,
weight=offset_conv5_w,
bias=offset_conv5_b,
)
offset_logit = mx.sym.broadcast_mul(lhs=offset_logit, rhs=X.var(name="offset_scale_%s_w" % stride, init=X.constant(1), shape=(1,1,1,1)))
offset_logit = mx.sym.exp(offset_logit)
centerness_logit_dict[stride] = center_logit
cls_logit_dict[stride] = cls_logit
offset_logit_dict[stride] = offset_logit
self.centerness_logit_dict = centerness_logit_dict
self.cls_logit_dict = cls_logit_dict
self.offset_logit_dict = offset_logit_dict
return self.centerness_logit_dict, self.cls_logit_dict, self.offset_logit_dict
def get_output(self, conv_fpn_feat):
if self.cls_logit_dict is not None and self.bbox_delta_dict is not None:
return self.cls_logit_dict, self.bbox_delta_dict
p = self.p
num_base_anchor = len(p.anchor_generate.ratio) * len(p.anchor_generate.scale)
conv_channel = p.head.conv_channel
# FPN RPN share weight
rpn_conv_weight = X.var('rpn_conv_weight', init=X.gauss(0.01))
rpn_conv_bias = X.var('rpn_conv_bias', init=X.zero_init())
rpn_conv_gamma = X.var('rpn_conv_gamma')
rpn_conv_beta = X.var('rpn_conv_beta')
rpn_conv_mmean = X.var('rpn_conv_moving_mean')
rpn_conv_mvar = X.var('rpn_conv_moving_var')
rpn_conv_cls_weight = X.var('rpn_conv_cls_weight', init=X.gauss(0.01))
rpn_conv_cls_bias = X.var('rpn_conv_cls_bias', init=X.zero_init())
rpn_conv_bbox_weight = X.var('rpn_conv_bbox_weight', init=X.gauss(0.01))
rpn_conv_bbox_bias = X.var('rpn_conv_bbox_bias', init=X.zero_init())
cls_logit_dict = {}
bbox_delta_dict = {}
for stride in p.anchor_generate.stride:
rpn_conv = X.conv(
conv_fpn_feat['stride%s' % stride],
kernel=3,
filter=conv_channel,
name="rpn_conv_3x3_%s" % stride,
no_bias=False,
weight=rpn_conv_weight,
bias=rpn_conv_bias
)
if p.normalizer.__name__ == "fix_bn":
pass
elif p.normalizer.__name__ == "sync_bn":
rpn_conv = p.normalizer(
rpn_conv,
gamma=rpn_conv_gamma,
beta=rpn_conv_beta,
moving_mean=rpn_conv_mmean,
moving_var=rpn_conv_mvar,
name="rpn_conv_3x3_bn_%s" % stride
)
elif p.normalizer.__name__ == "gn":
rpn_conv = p.normalizer(
rpn_conv,
gamma=rpn_conv_gamma,
beta=rpn_conv_beta,
name="rpn_conv_3x3_gn_%s" % stride
)
else:
raise NotImplementedError("Unsupported normalizer {}".format(p.normalizer.__name__))
rpn_relu = X.relu(rpn_conv, name='rpn_relu_%s' % stride)
if p.fp16:
rpn_relu = X.to_fp32(rpn_relu, name="rpn_relu_%s_fp32" % stride)
cls_logit = X.conv(
rpn_relu,
filter=2 * num_base_anchor,
name="rpn_cls_score_stride%s" % stride,
no_bias=False,
weight=rpn_conv_cls_weight,
bias=rpn_conv_cls_bias
)
bbox_delta = X.conv(
rpn_relu,
filter=4 * num_base_anchor,
name="rpn_bbox_pred_stride%s" % stride,
no_bias=False,
weight=rpn_conv_bbox_weight,
bias=rpn_conv_bbox_bias
)
cls_logit_dict[stride] = cls_logit
bbox_delta_dict[stride] = bbox_delta
self.cls_logit_dict = cls_logit_dict
self.bbox_delta_dict = bbox_delta_dict
return self.cls_logit_dict, self.bbox_delta_dict
def __init__(self, pHead):
self.p = patch_config_as_nothrow(pHead)
num_points = self.p.point_generate.num_points
self.dcn_kernel = int(math.sqrt(num_points))
self.dcn_pad = int((self.dcn_kernel - 1) / 2)
assert self.dcn_kernel * self.dcn_kernel == num_points, \
"The points number should be square."
assert self.dcn_kernel % 2 == 1, "The dcn kernel size should be odd."
# init moment method
dtype = "float16" if self.p.fp16 else "float32"
self.moment_transfer = X.var(name="moment_transfer",
shape=(2, ),
init=X.zero_init(),
lr_mult=0.01,
dtype=dtype)
# init bias for cls
prior_prob = 0.01
pi = -math.log((1 - prior_prob) / prior_prob)
# shared classification weight and bias
self.cls_conv1_weight = X.var("cls_conv1_weight",
init=X.gauss(std=0.01))
self.cls_conv1_bias = X.var("cls_conv1_bias", init=X.zero_init())
self.cls_conv2_weight = X.var("cls_conv2_weight",
init=X.gauss(std=0.01))
self.cls_conv2_bias = X.var("cls_conv2_bias", init=X.zero_init())
self.cls_conv3_weight = X.var("cls_conv3_weight",
init=X.gauss(std=0.01))
self.cls_conv3_bias = X.var("cls_conv3_bias", init=X.zero_init())
self.cls_conv_weight = X.var("cls_conv_weight", init=X.gauss(std=0.01))
self.cls_conv_bias = X.var("cls_conv_bias", init=X.zero_init())
self.cls_out_weight = X.var("cls_out_weight", init=X.gauss(std=0.01))
self.cls_out_bias = X.var("cls_out_bias", init=X.constant(pi))
# shared regression weight and bias
self.reg_conv1_weight = X.var("reg_conv1_weight",
init=X.gauss(std=0.01))
self.reg_conv1_bias = X.var("reg_conv1_bias", init=X.zero_init())
self.reg_conv2_weight = X.var("reg_conv2_weight",
init=X.gauss(std=0.01))
self.reg_conv2_bias = X.var("reg_conv2_bias", init=X.zero_init())
self.reg_conv3_weight = X.var("reg_conv3_weight",
init=X.gauss(std=0.01))
self.reg_conv3_bias = X.var("reg_conv3_bias", init=X.zero_init())
self.pts_init_conv_weight = X.var("pts_init_conv_weight",
init=X.gauss(std=0.01))
self.pts_init_conv_bias = X.var("pts_init_conv_bias",
init=X.zero_init())
self.pts_init_out_weight = X.var("pts_init_out_weight",
init=X.gauss(std=0.01))
self.pts_init_out_bias = X.var("pts_init_out_bias", init=X.zero_init())
self.pts_refine_conv_weight = X.var("pts_refine_conv_weight",
init=X.gauss(std=0.01))
self.pts_refine_conv_bias = X.var("pts_refine_conv_bias",
init=X.zero_init())
self.pts_refine_out_weight = X.var("pts_refine_out_weight",
init=X.gauss(std=0.01))
self.pts_refine_out_bias = X.var("pts_refine_out_bias",
init=X.zero_init())
self._pts_out_inits = None
self._pts_out_refines = None
self._cls_outs = None
def get_refine_output(self, roi_feature, cls_logit, bbox_delta,
cls_sec_logit, bbox_sec_delta):
p = self.p
num_class = p.num_class
repeat_time = p.repeat_time
class_agnostic = p.regress_target.class_agnostic
num_reg_class = 2 if class_agnostic else num_class
cls_logit = mx.sym.slice_axis(mx.sym.softmax(cls_logit),
axis=1,
begin=1,
end=num_class)
cls_sec_logit = mx.sym.slice_axis(mx.sym.softmax(cls_sec_logit),
axis=1,
begin=1,
end=num_class)
bbox_delta = mx.sym.slice_axis(bbox_delta,
axis=1,
begin=4,
end=num_reg_class * 4)
bbox_sec_delta = mx.sym.slice_axis(bbox_sec_delta,
axis=1,
begin=4,
end=num_reg_class * 4)
pred_feat1 = mx.sym.tile(mx.sym.concat(*[bbox_delta, cls_logit],
dim=1),
reps=(1, repeat_time))
pred_feat2 = mx.sym.tile(mx.sym.concat(
*[bbox_sec_delta, cls_sec_logit], dim=1),
reps=(1, repeat_time))
refine_feat1 = mx.sym.concat(*[roi_feature, pred_feat1], dim=1)
refine_feat2 = mx.sym.concat(*[roi_feature, pred_feat2], dim=1)
head_feat1 = X.fc(refine_feat1,
filter=1024,
weight=self.fc3_weight,
bias=self.fc3_bias,
name='fc3_conv_refine1')
head_feat1 = X.relu(head_feat1)
head_feat2 = X.fc(refine_feat2,
filter=1024,
weight=self.fc3_weight,
bias=self.fc3_bias,
name='fc3_conv_refine2')
head_feat2 = X.relu(head_feat2)
refine_cls_logit = X.fc(head_feat1,
filter=num_class,
name='refine_bbox_cls_logit1',
init=X.gauss(0.01))
refine_cls_sec_logit = X.fc(head_feat2,
filter=num_class,
name='refine_bbox_cls_logit2',
init=X.gauss(0.01))
refine_bbox_delta = X.fc(head_feat1,
filter=4 * num_reg_class,
name='refine_bbox_reg_delta1',
init=X.gauss(0.001))
refine_bbox_sec_delta = X.fc(head_feat2,
filter=4 * num_reg_class,
name='refine_bbox_reg_delta2',
init=X.gauss(0.001))
return refine_cls_logit, refine_bbox_delta, refine_cls_sec_logit, refine_bbox_sec_delta
def get_output(self, conv_feat):
if self._cls_logit_list is not None and self._bbox_delta_list is not None:
return self._cls_logit_list, self._bbox_delta_list
p = self.p
stride = p.anchor_generate.stride
if not isinstance(stride, tuple):
stride = (stride)
conv_channel = p.head.conv_channel
num_base_anchor = len(p.anchor_generate.ratio) * len(
p.anchor_generate.scale)
num_class = p.num_class
prior_prob = 0.01
pi = -math.log((1 - prior_prob) / prior_prob)
# shared classification weight and bias
self.cls_conv1_weight = X.var("cls_conv1_weight",
init=X.gauss(std=0.01))
self.cls_conv1_bias = X.var("cls_conv1_bias", init=X.zero_init())
self.cls_conv2_weight = X.var("cls_conv2_weight",
init=X.gauss(std=0.01))
self.cls_conv2_bias = X.var("cls_conv2_bias", init=X.zero_init())
self.cls_conv3_weight = X.var("cls_conv3_weight",
init=X.gauss(std=0.01))
self.cls_conv3_bias = X.var("cls_conv3_bias", init=X.zero_init())
self.cls_conv4_weight = X.var("cls_conv4_weight",
init=X.gauss(std=0.01))
self.cls_conv4_bias = X.var("cls_conv4_bias", init=X.zero_init())
self.cls_pred_weight = X.var("cls_pred_weight", init=X.gauss(std=0.01))
self.cls_pred_bias = X.var("cls_pred_bias", init=X.constant(pi))
# shared regression weight and bias
self.bbox_conv1_weight = X.var("bbox_conv1_weight",
init=X.gauss(std=0.01))
self.bbox_conv1_bias = X.var("bbox_conv1_bias", init=X.zero_init())
self.bbox_conv2_weight = X.var("bbox_conv2_weight",
init=X.gauss(std=0.01))
self.bbox_conv2_bias = X.var("bbox_conv2_bias", init=X.zero_init())
self.bbox_conv3_weight = X.var("bbox_conv3_weight",
init=X.gauss(std=0.01))
self.bbox_conv3_bias = X.var("bbox_conv3_bias", init=X.zero_init())
self.bbox_conv4_weight = X.var("bbox_conv4_weight",
init=X.gauss(std=0.01))
self.bbox_conv4_bias = X.var("bbox_conv4_bias", init=X.zero_init())
self.bbox_pred_weight = X.var("bbox_pred_weight",
init=X.gauss(std=0.01))
self.bbox_pred_bias = X.var("bbox_pred_bias", init=X.zero_init())
cls_logit_list = []
bbox_delta_list = []
for i, s in enumerate(stride):
cls_logit = self._cls_subnet(conv_feat=conv_feat[i],
conv_channel=conv_channel,
num_base_anchor=num_base_anchor,
num_class=num_class)
bbox_delta = self._bbox_subnet(conv_feat=conv_feat[i],
conv_channel=conv_channel,
num_base_anchor=num_base_anchor,
num_class=num_class)
cls_logit_list.append(cls_logit)
bbox_delta_list.append(bbox_delta)
self._cls_logit_list = cls_logit_list
self._bbox_delta_list = bbox_delta_list
return self._cls_logit_list, self._bbox_delta_list