def forward(self, net, batched_inputs):
del batched_inputs
cls_net = wnnl.non_local_blockv4(net,
scope=f"NonLocalROIHeadsHook_clsv2",
normalizer_fn=wnnl.evo_norm_s0,
activation_fn=None,
n_head=4,
weighed_sum=False)
reg_net = wnnl.non_local_blockv4(net,
scope=f"NonLocalROIHeadsHook_regv2",
normalizer_fn=wnnl.evo_norm_s0,
n_head=4,
activation_fn=None,
weighed_sum=False)
return cls_net, reg_net
def forward(self, features, batched_inputs):
del batched_inputs
res = OrderedDict()
normalizer_fn, normalizer_params = odt.get_norm(
"evo_norm_s0", is_training=self.is_training)
normalizer_params['G'] = 8
with tf.variable_scope("NonLocalBackboneHookV3"):
for k, v in features.items():
if k[0] not in ["C", "P"]:
continue
level = int(k[1:])
if level <= 2:
res[k] = v
continue
h = self.base_size // (2**level)
w = self.base_size // (2**level)
v = wnnl.non_local_blockv4(v,
inner_dims=[128, 128, 128],
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params,
n_head=2,
activation_fn=None,
weighed_sum=False,
scope=f"non_localv4_{level}",
size=[h, w])
res[k] = v
return res
def forward(self, net, batched_inputs, reuse=None):
del batched_inputs
net = wnnl.non_local_blockv4(net,
scope=f"NonLocalROIHeadsHookV2",
inner_dims_multiplier=[1, 1, 1],
normalizer_fn=wnnl.evo_norm_s0,
activation_fn=None,
weighed_sum=False,
n_head=4)
return net
def forward(self, net, batched_inputs):
del batched_inputs
cls_net = wnnl.non_local_blockv4(net,
scope=f"NonLocalROIHeadsHook_clsv3",
inner_dims_multiplier=[2, 2, 2],
normalizer_fn=wnnl.evo_norm_s0,
activation_fn=None,
n_head=2,
weighed_sum=False)
return cls_net, net
def forward(self, net, batched_inputs, reuse=None):
with tf.variable_scope("AddBBoxesSizeInfo"):
reg_net = net
shape = wmlt.combined_static_and_dynamic_shape(net)
C = shape[-1]
K = 4
with tf.variable_scope("pos_embedding"):
pos_embs_shape = [1, shape[1], shape[2], K * C]
pos_embedding = tf.get_variable(
"pos_embs",
shape=pos_embs_shape,
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=0.02))
bboxes = self.parent.t_proposal_boxes
with tf.name_scope("trans_bboxes"):
_, H, W, _ = btf.combined_static_and_dynamic_shape(
batched_inputs[IMAGE])
bboxes = odb.tfrelative_boxes_to_absolutely_boxes(bboxes, W, H)
bymin, bxmin, bymax, bxmax = tf.unstack(bboxes, axis=-1)
bh = bymax - bymin
bw = bxmax - bxmin
br0 = bh / (bw + 1e-8)
br1 = bw / (bh + 1e-8)
bboxes = tf.stack([bh, bw, br0, br1], axis=-1)
B, BN, BC = btf.combined_static_and_dynamic_shape(bboxes)
bboxes = tf.reshape(bboxes, [B * BN, BC])
bboxes = tf.stop_gradient(bboxes)
bboxes = slim.fully_connected(bboxes,
C,
activation_fn=self.activation_fn,
normalizer_fn=self.normalizer_fn,
normalizer_params=self.norm_params)
bboxes = slim.fully_connected(bboxes,
K * C,
activation_fn=tf.nn.sigmoid,
normalizer_fn=None)
pos_embedding = tf.reshape(bboxes,
[B * BN, 1, 1, K * C]) * pos_embedding
pos_embedding = tf.layers.dense(
pos_embedding,
C,
kernel_initializer=tf.truncated_normal_initializer(
stddev=0.02))
cls_net = wnnl.non_local_blockv4(net,
scope=f"non_local",
normalizer_fn=wnnl.evo_norm_s0,
activation_fn=None,
n_head=4,
weighed_sum=False,
pos_embedding=pos_embedding)
return cls_net, reg_net
def forward(self, net, batched_inputs):
del batched_inputs
if isinstance(net, (list, tuple)):
cls_x = net[0]
box_x = net[1]
else:
cls_x = net
box_x = net
cls_x = wnnl.non_local_blockv4(cls_x,
scope=f"NonLocalROIHeadsHook_clsv2",
normalizer_fn=wnnl.evo_norm_s0,
activation_fn=None,
n_head=4,
weighed_sum=False)
return cls_x, box_x