def cb_from_gene(x, n_channel: int, seq: List[bool]):
relu = netop.relu
sigmoid = netop.sigmoid
assert len(seq) == 1
# Segmentation
seg_a = seq[0:1]
# Translation
act_fn = sigmoid if seg_a[0] else relu
# Apply
return netop.ret_new(act_fn(netop.conv2d(x, (3, 3), n_channel)))
def ccg_from_gene(x, n_channel: int, seq: List[bool]):
assert len(seq) == 3
# Segmentation
seg_a = seq[0:1]
seg_cb_0 = seq[1:2]
seg_cb_1 = seq[2:3]
# Translation
use_res = seg_a[0]
# Apply
v = x
v = cb_from_gene(v, n_channel, seg_cb_0)
vv = cb_from_gene(v, n_channel, seg_cb_1)
return netop.ret_new(netop.add(v, vv) if use_res else vv)
def pb_from_gene(x, seq: List[bool]):
assert len(seq) == 4
size_table = (1, 4, 16, 64)
# Segmentation
seg_h = seq[0:2]
seg_w = seq[2:4]
# Translation
h_size = size_table[geneop.cvtlstgray(seg_h)]
w_size = size_table[geneop.cvtlstgray(seg_w)]
# Apply
v = x
if h_size == 1 and w_size == 1:
return v
v = netop.adv_pool(v, (h_size, w_size), (h_size, w_size),
("boxcar", "boxcar"), False)
return netop.ret_new(v)
def rb_from_gene(x, seq: List[bool]):
assert len(seq) == 9
cnv_size_table = (16, 32, 64, 128)
# Segmentation
pb_0 = seq[0:4]
seg_a = seq[4:6]
ccg_0 = seq[6:9]
# Translation
cnv_size = cnv_size_table[geneop.cvtlstgray(seg_a)]
# Apply
v = x
vv = pb_from_gene(v, pb_0)
if vv is v:
return v
v = vv
v = ccg_from_gene(v, cnv_size, ccg_0)
v = netop.resize_nearest(v, netop.get_shape(x)[1:3])
return netop.ret_new(v)
def b_from_gene(x, n_out_channel: int, seq: List[bool]):
assert len(seq) == 26
# Segmentation
cg_0 = seq[0:5]
rb_0 = seq[5:14]
rb_1 = seq[14:23]
ccg_0 = seq[23:26]
# Apply
v = x
v = cg_from_gene(v, cg_0)
v_rb_0 = rb_from_gene(v, rb_0)
v_rb_1 = rb_from_gene(v, rb_1)
l = [v]
if v_rb_0 is not v:
l.append(v_rb_0)
if v_rb_1 is not v:
l.append(v_rb_1)
if len(l) > 1:
v = netop.concat(l, axis=-1)
elif len(l) == 1:
v = l[0]
v = ccg_from_gene(v, n_out_channel, ccg_0)
return netop.ret_new(v)
def build_from_gene(x, n_out_channel: int, seq: List[bool]):
assert len(seq) == 142
cnv_channel_table = (32, 64, 128, 256)
# Segmentation
seg_a = seq[0:2]
seg_b = seq[2:12]
b_0 = seq[12:38]
b_1 = seq[38:64]
b_2 = seq[64:90]
b_3 = seq[90:116]
b_4 = seq[116:142]
# Translation
n_channel = cnv_channel_table[geneop.cvtlstgray(seg_a)]
MRL = seg_b # b-bb-bbb-bbbb | 0-12-345-6789
# Apply
v = x
b0 = b_from_gene(v, n_channel, b_0)
b1 = b_from_gene(b0, n_channel, b_1)
b1 = netop.add(b1, b0) if MRL[0] else b1
b2 = b_from_gene(b1, n_channel, b_2)
b2 = netop.add(b2, b0) if MRL[1] else b2
b2 = netop.add(b2, b1) if MRL[2] else b2
b3 = b_from_gene(b2, n_channel, b_3)
b3 = netop.add(b3, b0) if MRL[3] else b3
b3 = netop.add(b3, b1) if MRL[4] else b3
b3 = netop.add(b3, b2) if MRL[5] else b3
b4 = b_from_gene(b3, n_channel, b_4)
b4 = netop.add(b4, b0) if MRL[6] else b4
b4 = netop.add(b4, b1) if MRL[7] else b4
b4 = netop.add(b4, b2) if MRL[8] else b4
b4 = netop.add(b4, b3) if MRL[9] else b4
return netop.ret_new(netop.conv2d(b4, (1, 1), n_out_channel))
def build_from_gene(x, n_out_channel: int, seq: List[bool]):
import numpy as np
assert len(seq) == 142
cnv_channel_table = (32, 64, 128, 256)
# Segmentation
seg_a = seq[0:2]
seg_b = seq[2:12]
b_0 = seq[12:38]
b_1 = seq[38:64]
b_2 = seq[64:90]
b_3 = seq[90:116]
b_4 = seq[116:142]
# Translation
n_channel = cnv_channel_table[geneop.cvtlstgray(seg_a)]
MRL = seg_b # b-bb-bbb-bbbb | 0-12-345-6789
# Apply
v = x # (batch_size, n_frame, n_spec, n_channel)
batch_size, n_frame, n_spec, n_in_ch = [int(x) for x in v.get_shape()]
v = v[:, :, :512, :]
b0 = b_from_gene(v, n_channel, b_0)
b1 = b_from_gene(b0, n_channel, b_1)
b1 = netop.add(b1, b0) if MRL[0] else b1
b2 = b_from_gene(b1, n_channel, b_2)
b2 = netop.add(b2, b0) if MRL[1] else b2
b2 = netop.add(b2, b1) if MRL[2] else b2
b3 = b_from_gene(b2, n_channel, b_3)
b3 = netop.add(b3, b0) if MRL[3] else b3
b3 = netop.add(b3, b1) if MRL[4] else b3
b3 = netop.add(b3, b2) if MRL[5] else b3
b4 = b_from_gene(b3, n_channel, b_4)
b4 = netop.add(b4, b0) if MRL[6] else b4
b4 = netop.add(b4, b1) if MRL[7] else b4
b4 = netop.add(b4, b2) if MRL[8] else b4
b4 = netop.add(b4, b3) if MRL[9] else b4
v = netop.ret_new(netop.conv2d(
b4, (1, 1),
n_out_channel)) # (batch_size, n_frame, n_feature, n_channel)
l = []
sgv = tf.stop_gradient(v)
for i_ch in range(n_out_channel):
vv = tf.keras.layers.Dense(
n_spec - 512,
activation="relu",
use_bias=True,
kernel_initializer=tf.compat.v1.initializers.he_normal(
seed=0x12345678)).apply(sgv[:, :, :, i_ch])
vv = tf.reshape(vv, (
batch_size,
n_frame,
n_spec - 512,
1,
))
l.append(vv)
vv = tf.concat(l, axis=3)
v = tf.concat([v, vv], axis=2)
if sys.is_train:
v = v * x
else:
v = tf.clip_by_value(v, 0.0, 1.0) * x
return v