def _get_bbox_head_logit(self, conv_feat):
if self._head_feat is not None:
return self._head_feat
p = self.p
if p.normalizer.__name__ == "fix_bn":
conv_feat = X.convrelu(conv_feat, filter=256, kernel=3, name="bbox_conv1")
conv_feat = X.convrelu(conv_feat, filter=256, kernel=3, name="bbox_conv2")
conv_feat = X.convrelu(conv_feat, filter=256, kernel=3, name="bbox_conv3")
conv_feat = X.convrelu(conv_feat, filter=256, kernel=3, name="bbox_conv4")
elif p.normalizer.__name__ in ["sync_bn", "gn"]:
conv_feat = X.convnormrelu(p.normalizer, conv_feat, filter=256, kernel=3, name="bbox_conv1")
conv_feat = X.convnormrelu(p.normalizer, conv_feat, filter=256, kernel=3, name="bbox_conv2")
conv_feat = X.convnormrelu(p.normalizer, conv_feat, filter=256, kernel=3, name="bbox_conv3")
conv_feat = X.convnormrelu(p.normalizer, conv_feat, filter=256, kernel=3, name="bbox_conv4")
else:
raise NotImplementedError("Unsupported normalizer: {}".format(p.normalizer.__name__))
flatten = X.flatten(conv_feat, name="bbox_feat_flatten")
reshape = X.reshape(flatten, (0, 0, 1, 1), name="bbox_feat_reshape")
if p.normalizer.__name__ == "fix_bn":
fc1 = X.convrelu(reshape, filter=1024, name="bbox_fc1")
elif p.normalizer.__name__ in ["sync_bn", "gn"]:
fc1 = X.convnormrelu(p.normalizer, reshape, filter=1024, name="bbox_fc1")
else:
raise NotImplementedError("Unsupported normalizer: {}".format(p.normalizer.__name__))
self._head_feat = fc1
return self._head_feat
def _get_bbox_head_logit(self, conv_feat):
if self._head_feat is not None:
return self._head_feat
flatten = X.flatten(conv_feat, name="bbox_feat_flatten")
reshape = X.reshape(flatten, (0, 0, 1, 1), name="bbox_feat_reshape")
fc1 = X.convrelu(reshape, filter=1024, name="bbox_fc1")
fc2 = X.convrelu(fc1, filter=1024, name="bbox_fc2")
self._head_feat = fc2
return self._head_feat
def _cls_head(self, conv_feat):
xavier_init = mx.init.Xavier(factor_type="in", rnd_type="uniform", magnitude=3)
flatten = X.flatten(conv_feat, name="bbox_feat_flatten")
fc1 = X.fc(flatten, filter=1024, name="bbox_cls_fc1", init=xavier_init)
fc1 = self.add_norm(fc1)
fc1 = X.relu(fc1)
fc2 = X.fc(fc1, filter=1024, name="bbox_cls_fc2", init=xavier_init)
fc2 = self.add_norm(fc2)
fc2 = X.relu(fc2)
return fc2
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.flatten(conv_feat, name="bbox_feat_flatten")
fc1 = X.fc(flatten, filter=1024, name="bbox_fc1", init=xavier_init)
fc1 = X.relu(fc1)
fc2 = X.fc(fc1, filter=1024, name="bbox_fc2", init=xavier_init)
fc2 = X.relu(fc2)
self._head_feat = fc2
return self._head_feat
def _get_bbox_head_logit(self, conv_feat):
# comment this for re-infer in test stage
# if self._head_feat is not None:
# return self._head_feat
p = self.p
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)
if p.normalizer.__name__ == "fix_bn":
fc1 = X.convrelu(reshape,
filter=1024,
weight=self.fc1_weight,
bias=self.fc1_bias,
no_bias=False,
name="bbox_fc1_" + stage)
fc2 = X.convrelu(fc1,
filter=1024,
weight=self.fc2_weight,
bias=self.fc2_bias,
no_bias=False,
name="bbox_fc2_" + stage)
elif p.normalizer.__name__ in ["sync_bn", "gn"]:
fc1 = X.convnormrelu(p.normalizer,
reshape,
filter=1024,
weight=self.fc1_weight,
bias=self.fc1_bias,
no_bias=False,
name="bbox_fc1_" + stage)
fc2 = X.convnormrelu(p.normalizer,
fc1,
filter=1024,
weight=self.fc2_weight,
bias=self.fc2_bias,
no_bias=False,
name="bbox_fc2_" + stage)
else:
raise NotImplementedError("Unsupported normalizer: {}".format(
p.normalizer.__name__))
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_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
