def _bbox_subnet(self, conv_feat, conv_channel, num_base_anchor, num_class):
p = self.p
# regression subnet
bbox_conv1 = X.conv(
data=conv_feat,
kernel=3,
filter=conv_channel,
weight=self.bbox_conv1_weight,
bias=self.bbox_conv1_bias,
no_bias=False,
name="bbox_conv1"
)
bbox_conv1_relu = X.relu(bbox_conv1)
bbox_conv2 = X.conv(
data=bbox_conv1_relu,
kernel=3,
filter=conv_channel,
weight=self.bbox_conv2_weight,
bias=self.bbox_conv2_bias,
no_bias=False,
name="bbox_conv2"
)
bbox_conv2_relu = X.relu(bbox_conv2)
bbox_conv3 = X.conv(
data=bbox_conv2_relu,
kernel=3,
filter=conv_channel,
weight=self.bbox_conv3_weight,
bias=self.bbox_conv3_bias,
no_bias=False,
name="bbox_conv3"
)
bbox_conv3_relu = X.relu(bbox_conv3)
bbox_conv4 = X.conv(
data=bbox_conv3_relu,
kernel=3,
filter=conv_channel,
weight=self.bbox_conv4_weight,
bias=self.bbox_conv4_bias,
no_bias=False,
name="bbox_conv4"
)
bbox_conv4_relu = X.relu(bbox_conv4)
if p.fp16:
bbox_conv4_relu = X.to_fp32(bbox_conv4_relu, name="bbox_conv4_fp32")
output_channel = num_base_anchor * 4
output = X.conv(
data=bbox_conv4_relu,
kernel=3,
filter=output_channel,
weight=self.bbox_pred_weight,
bias=self.bbox_pred_bias,
no_bias=False,
name="bbox_pred"
)
return output
def se_v2_resnet_v1b_unit(input, name, filter, stride, dilate, proj, norm,
**kwargs):
"""
diff with v1: move the SE module to 3x3 conv
"""
conv1 = conv(input, name=name + "_conv1", filter=filter // 4)
bn1 = norm(conv1, name=name + "_bn1")
relu1 = relu(bn1, name=name + "_relu1")
conv2 = conv(relu1,
name=name + "_conv2",
stride=stride,
filter=filter // 4,
kernel=3)
bn2 = norm(conv2, name=name + "_bn2")
relu2 = relu(bn2, name=name + "_relu2")
relu2 = se(relu2,
prefix=name + "_se2",
f_down=filter // 16,
f_up=filter // 4)
conv3 = conv(relu2, name=name + "_conv3", filter=filter)
bn3 = norm(conv3, name=name + "_bn3")
if proj:
shortcut = conv(input, name=name + "_sc", filter=filter, stride=stride)
shortcut = norm(shortcut, name=name + "_sc_bn")
else:
shortcut = input
eltwise = add(bn3, shortcut, name=name + "_plus")
return relu(eltwise, name=name + "_relu")
def _cls_subnet(self, conv_feat, conv_channel, num_base_anchor, num_class):
p = self.p
# classification subnet
cls_conv1 = X.conv(
data=conv_feat,
kernel=3,
filter=conv_channel,
weight=self.cls_conv1_weight,
bias=self.cls_conv1_bias,
no_bias=False,
name="cls_conv1"
)
cls_conv1_relu = X.relu(cls_conv1)
cls_conv2 = X.conv(
data=cls_conv1_relu,
kernel=3,
filter=conv_channel,
weight=self.cls_conv2_weight,
bias=self.cls_conv2_bias,
no_bias=False,
name="cls_conv2"
)
cls_conv2_relu = X.relu(cls_conv2)
cls_conv3 = X.conv(
data=cls_conv2_relu,
kernel=3,
filter=conv_channel,
weight=self.cls_conv3_weight,
bias=self.cls_conv3_bias,
no_bias=False,
name="cls_conv3"
)
cls_conv3_relu = X.relu(cls_conv3)
cls_conv4 = X.conv(
data=cls_conv3_relu,
kernel=3,
filter=conv_channel,
weight=self.cls_conv4_weight,
bias=self.cls_conv4_bias,
no_bias=False,
name="cls_conv4"
)
cls_conv4_relu = X.relu(cls_conv4)
if p.fp16:
cls_conv4_relu = X.to_fp32(cls_conv4_relu, name="cls_conv4_fp32")
output_channel = num_base_anchor * (num_class - 1)
output = X.conv(
data=cls_conv4_relu,
kernel=3,
filter=output_channel,
weight=self.cls_pred_weight,
bias=self.cls_pred_bias,
no_bias=False,
name="cls_pred"
)
return output
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 conv_shared(data, name, kernel, num_filter, branch_ids=None, no_bias=True, share_weight=True,
pad=(0, 0), stride=(1, 1), dilate=(1, 1)):
if branch_ids is None:
branch_ids = range(len(data))
weight = X.var(name + '_weight')
if no_bias:
bias = None
else:
bias = X.var(name + '_bias')
conv_layers = []
for i in range(len(data)):
data_i = data[i]
stride_i = stride[i] if type(stride) is list else stride
dilate_i = dilate[i] if type(dilate) is list else dilate
pad_i = pad[i] if type(pad) is list else pad
branch_i = branch_ids[i]
if share_weight:
conv_i = X.conv(data=data_i, kernel=kernel, filter=num_filter, stride=stride_i, dilate=dilate_i, pad=pad_i,
name=name + '_shared%d' % branch_i, no_bias=no_bias, weight=weight, bias=bias)
else:
conv_i = X.conv(data=data_i, kernel=kernel, filter=num_filter, stride=stride_i, dilate=dilate_i, pad=pad_i,
name=name + '_branch%d' % branch_i, no_bias=no_bias)
conv_layers.append(conv_i)
return conv_layers
def mbconv(input, prefix, kernel, f_in, f_out, stride, proj, bottleneck_ratio,
norm, **kwargs):
with mx.name.Prefix(prefix + "_"):
if bottleneck_ratio != 1:
conv1 = conv(input,
name="conv1",
filter=f_in * bottleneck_ratio,
no_bias=False)
bn1 = norm(conv1, name="bn1")
relu1 = relu6(bn1, name="relu1")
else:
relu1 = input
conv2 = dwconv(relu1,
name="conv2",
filter=f_in * bottleneck_ratio,
kernel=kernel,
stride=stride,
no_bias=False)
bn2 = norm(conv2, name="bn2")
relu2 = relu6(bn2, name="relu2")
relu2 = se(relu2,
prefix=prefix + "_se2",
f_down=f_in // 4,
f_up=f_in * bottleneck_ratio)
conv3 = conv(relu2, name="conv3", filter=f_out, no_bias=False)
bn3 = norm(conv3, name="bn3")
if proj:
return bn3
else:
return bn3 + input
def dcn_resnet_unit(input, name, filter, stride, dilate, proj, norm, **kwargs):
conv1 = conv(input, name=name + "_conv1", filter=filter // 4)
bn1 = norm(conv1, name=name + "_bn1")
relu1 = relu(bn1, name=name + "_relu1")
# conv2 filter router
conv2_offset = conv(relu1, name=name + "_conv2_offset", filter=72, kernel=3, stride=stride, dilate=dilate)
conv2 = mx.sym.contrib.DeformableConvolution(relu1, conv2_offset, kernel=(3, 3),
stride=(stride, stride), dilate=(dilate, dilate), pad=(1, 1), num_filter=filter // 4,
num_deformable_group=4, no_bias=True, name=name + "_conv2")
bn2 = norm(conv2, name=name + "_bn2")
relu2 = relu(bn2, name=name + "_relu2")
conv3 = conv(relu2, name=name + "_conv3", filter=filter)
bn3 = norm(conv3, name=name + "_bn3")
if proj:
shortcut = conv(input, name=name + "_sc", filter=filter, stride=stride)
shortcut = norm(shortcut, name=name + "_sc_bn")
else:
shortcut = input
eltwise = add(bn3, shortcut, name=name + "_plus")
return relu(eltwise, name=name + "_relu")
def trident_resnet_v1b_unit(input, name, id, filter, stride, dilate, proj, **kwargs):
"""
Compared with v1, v1b moves stride=2 to the 3x3 conv instead of the 1x1 conv and use std in pre-processing
This is also known as the facebook re-implementation of ResNet(a.k.a. the torch ResNet)
"""
p = kwargs["params"]
share_bn = p.branch_bn_shared
share_conv = p.branch_conv_shared
norm = p.normalizer
######################### prepare names #########################
if id is not None:
conv_postfix = ("_shared%s" if share_conv else "_branch%s") % id
bn_postfix = ("_shared%s" if share_bn else "_branch%s") % id
other_postfix = "_branch%s" % id
else:
conv_postfix = ""
bn_postfix = ""
other_postfix = ""
######################### prepare parameters #########################
conv_params = lambda x: dict(
weight=X.shared_var(name + "_%s_weight" % x) if share_conv else None,
name=name + "_%s" % x + conv_postfix
)
def bn_params(x):
ret = dict(
gamma=X.shared_var(name + "_%s_gamma" % x) if share_bn else None,
beta=X.shared_var(name + "_%s_beta" % x) if share_bn else None,
moving_mean=X.shared_var(name + "_%s_moving_mean" % x) if share_bn else None,
moving_var=X.shared_var(name + "_%s_moving_var" % x) if share_bn else None,
name=name + "_%s" % x + bn_postfix
)
if norm.__name__ == "gn":
del ret["moving_mean"], ret["moving_var"]
return ret
######################### construct graph #########################
conv1 = conv(input, filter=filter // 4, **conv_params("conv1"))
bn1 = norm(conv1, **bn_params("bn1"))
relu1 = relu(bn1, name=name + other_postfix)
conv2 = conv(relu1, filter=filter // 4, kernel=3, stride=stride, dilate=dilate, **conv_params("conv2"))
bn2 = norm(conv2, **bn_params("bn2"))
relu2 = relu(bn2, name=name + other_postfix)
conv3 = conv(relu2, filter=filter, **conv_params("conv3"))
bn3 = norm(conv3, **bn_params("bn3"))
if proj:
shortcut = conv(input, filter=filter, stride=stride, **conv_params("sc"))
shortcut = norm(shortcut, **bn_params("sc_bn"))
else:
shortcut = input
eltwise = add(bn3, shortcut, name=name + "_plus" + other_postfix)
return relu(eltwise, name=name + "_relu" + other_postfix)
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 sepc_conv(x,
name,
out_channels,
kernel_size,
i,
stride=1,
padding=0,
dilation=1,
groups=1,
deformable_groups=1,
part_deform=False,
start_level=1,
weight=None,
bias=None,
weight_offset=None,
bias_offset=None):
assert weight is not None and bias is not None
if part_deform:
assert weight_offset is not None and bias_offset is not None
if i < start_level or not part_deform:
return conv(x,
name,
filter=out_channels,
kernel=kernel_size,
stride=stride,
pad=kernel_size // 2,
dilate=dilation,
num_group=groups,
no_bias=False,
weight=weight,
bias=bias)
offset = conv(x,
name + 'offset',
filter=deformable_groups * 2 * kernel_size * kernel_size,
kernel=kernel_size,
stride=stride,
pad=kernel_size // 2,
dilate=dilation,
num_group=groups,
no_bias=False,
weight=weight_offset,
bias=bias_offset)
return DeformConv(x,
offset,
name,
out_channels,
kernel_size,
stride,
padding=padding,
dilation=dilation,
groups=groups,
deformable_groups=deformable_groups,
no_bias=False,
weight=weight,
bias=bias)
def se(input, prefix, f_down, f_up):
with mx.name.Prefix(prefix + "_"):
gap = mx.sym.mean(input, axis=-1, keepdims=True)
gap = mx.sym.mean(gap, axis=-2, keepdims=True)
fc1 = conv(gap, name="fc1", filter=f_down)
fc1 = relu6(fc1, name="fc1_relu")
fc2 = conv(fc1, name="fc2", filter=f_up)
att = sigmoid(fc2, name="sigmoid")
input = mx.sym.broadcast_mul(input, att, name="mul")
return input
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 trident_resnet_v1_unit(input, name, id, filter, stride, dilate, proj, **kwargs):
p = kwargs["params"]
share_bn = p.branch_bn_shared
share_conv = p.branch_conv_shared
norm = p.normalizer
######################### prepare names #########################
if id is not None:
conv_postfix = ("_shared%s" if share_conv else "_branch%s") % id
bn_postfix = ("_shared%s" if share_bn else "_branch%s") % id
other_postfix = "_branch%s" % id
else:
conv_postfix = ""
bn_postfix = ""
other_postfix = ""
######################### prepare parameters #########################
conv_params = lambda x: dict(
weight=X.shared_var(name + "_%s_weight" % x) if share_conv else None,
name=name + "_%s" % x + conv_postfix
)
bn_params = lambda x: dict(
gamma=X.shared_var(name + "_%s_gamma" % x) if share_bn else None,
beta=X.shared_var(name + "_%s_beta" % x) if share_bn else None,
moving_mean=X.shared_var(name + "_%s_moving_mean" % x) if share_bn else None,
moving_var=X.shared_var(name + "_%s_moving_var" % x) if share_bn else None,
name=name + "_%s" % x + bn_postfix
)
######################### construct graph #########################
conv1 = conv(input, filter=filter // 4, stride=stride, **conv_params("conv1"))
bn1 = norm(conv1, **bn_params("bn1"))
relu1 = relu(bn1, name=name + other_postfix)
conv2 = conv(relu1, filter=filter // 4, kernel=3, dilate=dilate, **conv_params("conv2"))
bn2 = norm(conv2, **bn_params("bn2"))
relu2 = relu(bn2, name=name + other_postfix)
conv3 = conv(relu2, filter=filter, **conv_params("conv3"))
bn3 = norm(conv3, **bn_params("bn3"))
if proj:
shortcut = conv(input, filter=filter, stride=stride, **conv_params("sc"))
shortcut = norm(shortcut, **bn_params("sc_bn"))
else:
shortcut = input
eltwise = add(bn3, shortcut, name=name + "_plus" + other_postfix)
return relu(eltwise, name=name + "_relu" + other_postfix)
def _refine_pts(self, cls_feat, reg_feat, dcn_offset, pts_init_out):
p = self.p
point_conv_channel = p.head.point_conv_channel
num_class = p.num_class
output_channel = num_class - 1
pts_output_channel = p.point_generate.num_points * 2
cls_conv = mx.symbol.contrib.DeformableConvolution(
data=cls_feat,
offset=dcn_offset,
kernel=(self.dcn_kernel, self.dcn_kernel),
pad=(self.dcn_pad, self.dcn_pad),
stride=(1, 1),
dilate=(1, 1),
num_filter=point_conv_channel,
weight=self.cls_conv_weight,
bias=self.cls_conv_bias,
no_bias=False,
name="cls_conv")
cls_conv_relu = X.relu(cls_conv)
cls_out = X.conv(data=cls_conv_relu,
kernel=1,
filter=output_channel,
weight=self.cls_out_weight,
bias=self.cls_out_bias,
no_bias=False,
name="cls_out")
pts_refine_conv = mx.symbol.contrib.DeformableConvolution(
data=reg_feat,
offset=dcn_offset,
kernel=(self.dcn_kernel, self.dcn_kernel),
pad=(self.dcn_pad, self.dcn_pad),
stride=(1, 1),
dilate=(1, 1),
num_filter=point_conv_channel,
weight=self.pts_refine_conv_weight,
bias=self.pts_refine_conv_bias,
no_bias=False,
name="pts_refine_conv")
pts_refine_conv_relu = X.relu(pts_refine_conv)
pts_refine_out = X.conv(data=pts_refine_conv_relu,
kernel=1,
filter=pts_output_channel,
weight=self.pts_refine_out_weight,
bias=self.pts_refine_out_bias,
no_bias=False,
name="pts_refine_out")
pts_refine_out = pts_refine_out + X.block_grad(pts_init_out)
return pts_refine_out, cls_out
def _get_bbox_head_logit(self, conv_feat):
if self._head_feat is not None:
return self._head_feat
xavier_init = mx.init.Xavier(factor_type="in", rnd_type="uniform", magnitude=3)
flatten = X.reshape(conv_feat, shape=(0, -1, 1, 1), name="bbox_feat_reshape")
fc1 = X.conv(flatten, filter=1024, name="bbox_fc1", init=xavier_init)
fc1 = self.add_norm(fc1)
fc1 = X.relu(fc1)
fc2 = X.conv(fc1, filter=1024, name="bbox_fc2", init=xavier_init)
fc2 = self.add_norm(fc2)
fc2 = X.relu(fc2)
self._head_feat = fc2
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_P0_features(c_features, p_names, dim_reduced, init, norm):
p_features = {}
for c_feature, p_name in zip(c_features, p_names):
p = X.conv(data=c_feature,
filter=dim_reduced,
no_bias=False,
weight=X.var(name=p_name + "_weight", init=init),
bias=X.var(name=p_name + "_bias", init=X.zero_init()),
name=p_name)
p = norm(p, name=p_name + '_bn')
p_features[p_name] = p
return p_features
def convnormrelu(input, prefix, kernel, f_in, f_out, stride, proj, norm,
**kwargs):
with mx.name.Prefix(prefix + "_"):
conv1 = conv(input,
name="conv1",
filter=f_out,
kernel=kernel,
stride=stride,
no_bias=False)
bn1 = norm(conv1, name="bn1")
relu1 = relu6(bn1, name="relu1")
return relu1
def _init_pts(self, reg_feat):
p = self.p
point_conv_channel = p.head.point_conv_channel
pts_output_channel = p.point_generate.num_points * 2
pts_init_conv = X.conv(data=reg_feat,
kernel=3,
filter=point_conv_channel,
weight=self.pts_init_conv_weight,
bias=self.pts_init_conv_bias,
no_bias=False,
name="pts_init_conv")
pts_init_conv_relu = X.relu(pts_init_conv)
pts_init_out = X.conv(data=pts_init_conv_relu,
kernel=1,
filter=pts_output_channel,
weight=self.pts_init_out_weight,
bias=self.pts_init_out_bias,
no_bias=False,
name="pts_init_out")
return pts_init_out
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 _convs_and_fcs(self, x, num_convs, num_fcs, name, conv_init, fc_init):
'''
Args:
x: [N, C, H, W] feature maps
num_convs: int
num_fcs: int
conv_init: mx initializer
Returns:
x: [N, C, H, W] or [N, C, 1, 1]
'''
if num_convs == 0 and num_fcs == 0:
return x
out_channels = self.p.TSD.conv_out_channels
out_fc_channels = self.p.TSD.fc_out_channels
if num_convs > 0:
for i in range(num_convs):
x = X.relu(
X.conv(x,
kernel=3,
filter=out_channels,
no_bias=False,
name=name + '_conv%s' % i,
init=conv_init))
if num_fcs > 0:
x = X.reshape(x,
shape=(0, -1, 1, 1),
name=name + '_conv_fc_flatten')
for i in range(num_fcs):
x = X.relu(
X.conv(x,
kernel=1,
filter=out_fc_channels,
no_bias=False,
name=name + '_fc%s' % i,
init=fc_init))
return x
def _cls_subnet(self, conv_feat, stride):
p = self.p
norm = p.normalizer
conv_channel = p.head.conv_channel
# classification subset
cls_conv1 = X.conv(data=conv_feat,
kernel=3,
filter=conv_channel,
weight=self.cls_conv1_weight,
bias=self.cls_conv1_bias,
no_bias=False,
name="cls_conv1")
cls_conv1 = norm(cls_conv1, name="cls_conv1_bn_s{}".format(stride))
cls_conv1_relu = X.relu(cls_conv1)
cls_conv2 = X.conv(data=cls_conv1_relu,
kernel=3,
filter=conv_channel,
weight=self.cls_conv2_weight,
bias=self.cls_conv2_bias,
no_bias=False,
name="cls_conv2")
cls_conv2 = norm(cls_conv2, name="cls_conv2_bn_s{}".format(stride))
cls_conv2_relu = X.relu(cls_conv2)
cls_conv3 = X.conv(data=cls_conv2_relu,
kernel=3,
filter=conv_channel,
weight=self.cls_conv3_weight,
bias=self.cls_conv3_bias,
no_bias=False,
name="cls_conv3")
cls_conv3 = norm(cls_conv3, name="cls_conv3_bn_s{}".format(stride))
cls_conv3_relu = X.relu(cls_conv3)
if p.fp16:
cls_conv3_relu = X.to_fp32(cls_conv3_relu, name="cls_conv3_fp32")
return cls_conv3_relu
def _reg_subnet(self, conv_feat, stride):
p = self.p
norm = p.normalizer
conv_channel = p.head.conv_channel
# regression subnet
reg_conv1 = X.conv(data=conv_feat,
kernel=3,
filter=conv_channel,
weight=self.reg_conv1_weight,
bias=self.reg_conv1_bias,
no_bias=False,
name="reg_conv1")
reg_conv1 = norm(reg_conv1, name="reg_conv1_bn_s{}".format(stride))
reg_conv1_relu = X.relu(reg_conv1)
reg_conv2 = X.conv(data=reg_conv1_relu,
kernel=3,
filter=conv_channel,
weight=self.reg_conv2_weight,
bias=self.reg_conv2_bias,
no_bias=False,
name="reg_conv2")
reg_conv2 = norm(reg_conv2, name="reg_conv2_bn_s{}".format(stride))
reg_conv2_relu = X.relu(reg_conv2)
reg_conv3 = X.conv(data=reg_conv2_relu,
kernel=3,
filter=conv_channel,
weight=self.reg_conv3_weight,
bias=self.reg_conv3_bias,
no_bias=False,
name="reg_conv3")
reg_conv3 = norm(reg_conv3, name="reg_conv3_bn_s{}".format(stride))
reg_conv3_relu = X.relu(reg_conv3)
if p.fp16:
reg_conv3_relu = X.to_fp32(reg_conv3_relu, name="reg_conv3_fp32")
return reg_conv3_relu
def _get_bbox_head_logit(self, conv_feat):
#if self._head_feat is not None:
# return self._head_feat
stage = self.stage
flatten = X.flatten(conv_feat, name="bbox_feat_flatten_" + stage)
reshape = X.reshape(flatten, (0, 0, 1, 1),
name="bbox_feat_reshape_" + stage)
fc1 = X.conv(reshape,
filter=1024,
weight=self.fc1_weight,
name="bbox_fc1_" + stage)
fc1_relu = X.relu(fc1, name="bbox_fc1_relu_" + stage)
fc2 = X.conv(fc1_relu,
filter=1024,
weight=self.fc2_weight,
name="bbox_fc2_" + stage)
fc2_relu = X.relu(fc2, name="bbox_fc2_" + stage)
self._head_feat = fc2_relu
return self._head_feat
def get_P0_features(c_features,
p_names,
dim_reduced,
init,
norm,
kernel=1):
p_features = []
for c_feature, p_name in zip(c_features, p_names):
p = X.conv(data=c_feature,
filter=dim_reduced,
kernel=kernel,
no_bias=False,
weight=X.var(name=p_name + "_weight", init=init),
bias=X.var(name=p_name + "_bias", init=X.zero_init()),
name=p_name)
p_features.append(p)
return p_features
def _get_mask_head_logit(self, conv_feat):
if self._head_feat is not None:
return self._head_feat
up_stride = int(self.pMask.resolution // self.pMaskRoi.out_size)
dim_reduced = self.pMask.dim_reduced
msra_init = mx.init.Xavier(rnd_type="gaussian", factor_type="out", magnitude=2)
current = conv_feat
for i in range(4):
current = X.conv(
current,
name="mask_fcn_conv{}".format(i + 1),
filter=dim_reduced,
kernel=3,
no_bias=False,
init=msra_init
)
current = self.add_norm(current)
current = X.relu(current)
mask_up = current
for i in range(up_stride // 2):
weight = X.var(
name="mask_up{}_weight".format(i),
init=msra_init,
lr_mult=1,
wd_mult=1)
mask_up = mx.sym.Deconvolution(
mask_up,
kernel=(2, 2),
stride=(2, 2),
num_filter=dim_reduced,
no_bias=False,
weight=weight,
name="mask_up{}".format(i)
)
mask_up = X.relu(
mask_up,
name="mask_up{}_relu".format(i))
mask_up = X.to_fp32(mask_up, name='mask_up_to_fp32')
self._head_feat = mask_up
return self._head_feat
def get_output(self, conv_feat):
pBbox = self.pBbox
num_class = pBbox.num_class
head_feat = self._get_mask_head_logit(conv_feat)
msra_init = mx.init.Xavier(rnd_type="gaussian", factor_type="out", magnitude=2)
if self.pMask:
head_feat = X.to_fp32(head_feat, name="mask_head_to_fp32")
mask_fcn_logit = X.conv(
head_feat,
filter=num_class,
name="mask_fcn_logit",
no_bias=False,
init=msra_init
)
return mask_fcn_logit
def _bbox_subnet(self,
conv_feat,
conv_channel,
num_base_anchor,
num_class,
stride,
nb_conv=0):
p = self.p
if nb_conv <= 0:
bbox_conv4_relu = conv_feat
if p.fp16:
bbox_conv4_relu = X.to_fp32(bbox_conv4_relu,
name="bbox_conv4_fp32")
output_channel = num_base_anchor * 4
output = X.conv(data=bbox_conv4_relu,
kernel=3,
filter=output_channel,
weight=self.bbox_pred_weight,
bias=self.bbox_pred_bias,
no_bias=False,
name="bbox_pred")
return output
return super()._bbox_subnet(conv_feat, conv_channel, num_base_anchor,
num_class, stride)
def fpn_neck(self, data):
if self.fpn_feat is not None:
return self.fpn_feat
c2, c3, c4, c5 = data
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 = self.add_norm(p5)
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")
p5_conv = self.add_norm(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")
p4_la = self.add_norm(p4_la)
p5_clip = mx.sym.slice_like(p5_up, p4_la, name="P4_clip")
p4 = mx.sym.add_n(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")
p4_conv = self.add_norm(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")
p3_la = self.add_norm(p3_la)
p4_clip = mx.sym.slice_like(p4_up, p3_la, name="P3_clip")
p3 = mx.sym.add_n(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")
p3_conv = self.add_norm(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")
p2_la = self.add_norm(p2_la)
p3_clip = mx.sym.slice_like(p3_up, p2_la, name="P2_clip")
p2 = mx.sym.add_n(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")
p2_conv = self.add_norm(p2_conv)
# P6
p6 = X.max_pool(
p5_conv,
name="P6_subsampling",
kernel=1,
stride=2,
)
conv_fpn_feat = dict(stride64=p6,
stride32=p5_conv,
stride16=p4_conv,
stride8=p3_conv,
stride4=p2_conv)
self.fpn_feat = conv_fpn_feat
return self.fpn_feat
def get_retinanet_neck(self, data):
if self.neck is not None:
return self.neck
c2, c3, c4, c5 = data
import mxnet as mx
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.slice_like(p5_up, p4_la, name="P4_clip")
p4 = mx.sym.add_n(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.slice_like(p4_up, p3_la, name="P3_clip")
p3 = mx.sym.add_n(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")
# P6
p6 = X.conv(data=c5,
kernel=3,
stride=2,
filter=256,
no_bias=False,
weight=X.var(name="P6_conv_weight", init=xavier_init),
bias=X.var(name="P6_conv_bias", init=X.zero_init()),
name="P6_conv")
# P7
p6_relu = X.relu(data=p6, name="P6_relu")
p7 = X.conv(data=p6_relu,
kernel=3,
stride=2,
filter=256,
no_bias=False,
weight=X.var(name="P7_conv_weight", init=xavier_init),
bias=X.var(name="P7_conv_bias", init=X.zero_init()),
name="P7_conv")
self.neck = dict(stride8=p3_conv,
stride16=p4_conv,
stride32=p5_conv,
stride64=p6,
stride128=p7)
return self.neck