def fpn(feats, fpn_filters, strides):
# lateral connections(1x1 conv)
feats = [Conv_BN(i, fpn_filters, kernel_size=1, strides=1, activation=None) for i in feats]
C3, C4, C5 = feats
# top-down connections(upSampling)
P5 = C5
P5_up = UpSampling2D(size=2, interpolation='nearest')(P5)
P4 = add([C4, P5_up])
P4_up = UpSampling2D(size=2, interpolation='nearest')(P4)
P3 = add([C3, P4_up])
# p6p7
P6 = Conv_BN(P5, fpn_filters, 3, strides=2, activation='relu')
P7 = Conv_BN(P6, fpn_filters, 3, strides=2, activation='relu')
feature_dict = {8:P3, 16:P4, 32:P5, 64:P6, 128:P7}
return [feature_dict[s] for s in strides]
def build_bifpn(feats, config):
# Node id starts from the input features and monotonically increase whenever
# a new node is added. Here is an example for level P3 - P7:
# P7 (4) P7" (12)
# P6 (3) P6' (5) P6" (11)
# P5 (2) P5' (6) P5" (10)
# P4 (1) P4' (7) P4" (9)
# P3 (0) P3" (8)
fpn_nodes = [{'feat_level': 6, 'inputs_offsets': [3, 4]}, # for P6'
{'feat_level': 5, 'inputs_offsets': [2, 5]}, # for P5'
{'feat_level': 4, 'inputs_offsets': [1, 6]}, # for P4'
{'feat_level': 3, 'inputs_offsets': [0, 7]}, # for P3"
{'feat_level': 4, 'inputs_offsets': [1, 7, 8]}, # for P4"
{'feat_level': 5, 'inputs_offsets': [2, 6, 9]}, # for P5"
{'feat_level': 6, 'inputs_offsets': [3, 5, 10]}, # for P6"
{'feat_level': 7, 'inputs_offsets': [4, 11]}] # for P7"
for i, fpn_node in enumerate(fpn_nodes):
new_node_h, new_node_w = feats[fpn_node['feat_level']-config['min_level']]._keras_shape[1:3]
nodes_in = []
for idx, inputs_offset in enumerate(fpn_node['inputs_offsets']):
input_node = feats[inputs_offset]
input_node = resample_feature_map(input_node, new_node_h, new_node_w, config['fpn_num_filters'])
nodes_in.append(input_node)
new_node = Lambda(fuse_features, arguments={'weight_method': 'fast'})(nodes_in)
new_node = Conv_BN(new_node, config['fpn_num_filters'], activation=None)
feats.append(new_node)
fpn_feats = feats[-5:]
return fpn_feats # [P3", P4", P5", P6", P7"]
def shared_box_head(n_filters, n_anchors):
inpt = Input((None,None,n_filters))
x = inpt
for i in range(4):
x = Conv_BN(x, n_filters, 3, strides=1, activation='relu')
x = Conv2D(4*n_anchors, 3, strides=1, padding='same')(x)
model = Model(inpt, x, name='shared_box_head')
return model
def resample_feature_map(x, target_h, target_w, target_c):
# 1x1 conv if channel not match, conv-bn-swish
x = Conv_BN(x, target_c, kernel_size=1, strides=1)
# resize
h, w = x._keras_shape[1:3]
if h > target_h and w > target_w:
h_stride = int((h - 1)//target_h + 1)
w_stride = int((w - 1)//target_w + 1)
x = MaxPooling2D(pool_size=(h_stride+1, w_stride+1), strides=(h_stride,w_stride), padding='same')(x)
elif h <= target_h and w <= target_w:
x = Lambda(tf.image.resize_nearest_neighbor, arguments={'size': [target_h, target_w]})(x)
else:
print("Incompatible target feature map size")
return x
def fpn_node(x, n_filters, up_filters, out_filters):
# shared conv
x = Conv_BN(x, n_filters, 1, strides=1, activation='leaky')
x = Conv_BN(x, n_filters * 2, 3, strides=1, activation='leaky')
x = Conv_BN(x, n_filters, 1, strides=1, activation='leaky')
x = Conv_BN(x, n_filters * 2, 3, strides=1, activation='leaky')
x = Conv_BN(x, n_filters, 1, strides=1, activation='leaky')
# up branch 1x1 conv
up = Conv_BN(x, up_filters, 1, strides=1,
activation='leaky') if up_filters else x
# out branch 3x3 conv + 1x1 conv head
out = Conv_BN(x, n_filters * 2, 3, strides=1, activation='leaky')
out = Conv2D(out_filters, 1, strides=1, padding='same')(out)
return up, out