def __init__(self) -> None:
size = img_size
self.conv1 = convolution.Convolution(input_shape=(size, size, ci),
out_channel=oc1,
kernel_size=(ks1, ks1),
stride=(stride1, stride1),
learning_rate=conv1_learning_rate,
activate_func='relu')
assert (size - ks1) % stride1 == 0
size = (size - ks1) // stride1 + 1
self.pool1 = pooling.Pooling(input_shape=(size, size, oc1),
pool_size=ps1)
assert size % 2 == 0
size //= 2
self.conv2 = convolution.Convolution(input_shape=(size, size, oc1),
out_channel=oc2,
kernel_size=(ks2, ks2),
stride=(stride2, stride2),
learning_rate=conv2_learning_rate,
activate_func='relu')
assert (size - ks2) % stride2 == 0
size = (size - ks2) // stride2 + 1
self.pool2 = pooling.Pooling(input_shape=(size, size, oc2),
pool_size=ps2)
assert size % 2 == 0
size //= 2
self.conv3 = convolution.Convolution(input_shape=(size, size, oc2),
out_channel=oc3,
kernel_size=(ks3, ks3),
stride=(stride3, stride3),
learning_rate=conv3_learning_rate,
activate_func='relu')
assert (size - ks3) % stride3 == 0
size = (size - ks3) // stride3 + 1
self.pool3 = pooling.Pooling(input_shape=(size, size, oc3),
pool_size=ps3)
assert size % 2 == 0
size //= 2
self.conv4 = convolution.Convolution(input_shape=(size, size, oc3),
out_channel=oc4,
kernel_size=(ks4, ks4),
stride=(stride4, stride4),
learning_rate=conv4_learning_rate,
activate_func='relu')
assert (size - ks4) % stride4 == 0
size4 = (size - ks4) // stride4 + 1
self.fc1 = full_connection.FullConnection(
dim_in=size * size * oc3,
dim_out=deepid_dim // 2,
learning_rate=fc1_learning_rate,
activate_func='relu')
self.fc2 = full_connection.FullConnection(
dim_in=size4 * size4 * oc4,
dim_out=deepid_dim // 2,
learning_rate=fc2_learning_rate,
activate_func='relu')
def _layers_init(self):
return sequential.Sequential([
convolution.Convolution(kernel_size=4,
output_depth=32,
input_depth=4,
input_dim=27,
act='relu',
stride_size=1,
pad='VALID'),
convolution.Convolution(kernel_size=3,
output_depth=32,
stride_size=1,
act='relu',
pad='SAME'),
maxpool.MaxPool(),
convolution.Convolution(kernel_size=3,
output_depth=64,
stride_size=1,
act='relu',
pad='SAME'),
convolution.Convolution(kernel_size=3,
output_depth=64,
stride_size=1,
act='relu',
pad='SAME'),
convolution.Convolution(kernel_size=3,
output_depth=64,
stride_size=1,
act='relu',
pad='SAME'),
maxpool.MaxPool(),
linear.Linear(64,
act='relu',
keep_prob=self.dropout_prob,
use_dropout=True),
linear.Linear(64,
act='relu',
keep_prob=self.dropout_prob,
use_dropout=True),
avgCyclicPool2.CyclicAvgPool(),
linear.Linear(2, act='linear')
])
def convolutional(X,
channel_in,
channel_out,
filter_height=3,
filter_width=3,
stride=1,
pad=1):
W = Parameter(
np.random.randn(channel_out, channel_in, filter_height, filter_width),
'Weights')
b = Parameter(np.random.randn(channel_out), 'Bias')
return convolution.Convolution(X, W, b, stride, pad)
def __init__(self) -> None:
self.conv1 = convolution.Convolution((130, 130, 3),
out_channel=96,
kernel_size=(5, 5),
stride=(1, 1),
learning_rate=learning_rate,
padding=(1, 1),
activate_func='mfm')
self.conv2a = convolution.Convolution((64, 64, 48),
out_channel=96,
kernel_size=(1, 1),
stride=(1, 1),
learning_rate=learning_rate,
activate_func='mfm')
self.conv2 = convolution.Convolution((64, 64, 48),
out_channel=192,
kernel_size=(3, 3),
stride=(1, 1),
learning_rate=learning_rate,
padding=(1, 1),
activate_func='mfm')
self.conv3a = convolution.Convolution((32, 32, 96),
out_channel=192,
kernel_size=(1, 1),
stride=(1, 1),
learning_rate=learning_rate,
activate_func='mfm')
self.conv3 = convolution.Convolution((32, 32, 96),
out_channel=384,
kernel_size=(3, 3),
stride=(1, 1),
learning_rate=learning_rate,
padding=(1, 1),
activate_func='mfm')
self.conv4a = convolution.Convolution((16, 16, 192),
out_channel=384,
kernel_size=(1, 1),
stride=(1, 1),
learning_rate=learning_rate,
activate_func='mfm')
self.conv4 = convolution.Convolution((16, 16, 192),
out_channel=256,
kernel_size=(3, 3),
stride=(1, 1),
learning_rate=learning_rate,
padding=(1, 1),
activate_func='mfm')
self.conv5a = convolution.Convolution((16, 16, 128),
out_channel=256,
kernel_size=(1, 1),
stride=(1, 1),
learning_rate=learning_rate,
activate_func='mfm')
self.conv5 = convolution.Convolution((16, 16, 128),
out_channel=256,
kernel_size=(3, 3),
stride=(1, 1),
learning_rate=learning_rate,
padding=(1, 1),
activate_func='mfm')
self.pool1 = pooling.Pooling((128, 128, 48), pool_size=2)
self.pool2 = pooling.Pooling((64, 64, 96), pool_size=2)
self.pool3 = pooling.Pooling((32, 32, 192), pool_size=2)
self.pool4 = pooling.Pooling((16, 16, 128), pool_size=2)
self.fc1 = full_connection.FullConnection(8 * 8 * 128,
512,
learning_rate=learning_rate,
activate_func='mfm')
self.fc_out = full_connection.FullConnection(
256, class_num, learning_rate=learning_rate)