def __init__(self, classes=1000):
super(alexnet, self).__init__()
self.conv1 = nn.NormalConv2D(filters=64,
kernel_size=11,
strides=4,
padding='same',
activation='relu',
use_bias=True)
self.bn1 = nn.NormalBatchNormalization(center=False, scale=False)
self.pool1 = nn.NormalMaxPool2D(pool_size=2, strides=2)
self.enter_int = nn.EnterInteger(scale=1.0)
self.conv2 = nn.BinaryConv2D(filters=192,
kernel_size=5,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn2 = nn.BatchNormalization(self.conv2)
self.pool2 = nn.MaxPool2D(pool_size=2, strides=2)
self.conv3 = nn.BinaryConv2D(filters=384,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn3 = nn.BatchNormalization(self.conv3)
self.conv4 = nn.BinaryConv2D(filters=384,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn4 = nn.BatchNormalization(self.conv4)
self.conv5 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn5 = nn.BatchNormalization(self.conv5)
self.pool5 = nn.MaxPool2D(pool_size=2, strides=2)
self.exit_int = nn.ExitInteger()
self.flatten = nn.Flatten()
self.dense6 = nn.NormalDense(4096, use_bias=True, activation='relu')
self.bn6 = nn.NormalBatchNormalization()
self.dense7 = nn.NormalDense(4096, use_bias=True, activation='relu')
self.bn7 = nn.NormalBatchNormalization()
self.dense8 = nn.NormalDense(classes, use_bias=True)
#self.scalu = nn.Scalu()
self.softmax = nn.Activation('softmax')
def __init__(self, classes=1000):
super(vggnet, self).__init__()
self.conv1 = nn.NormalConv2D(
filters=96,
kernel_size=7,
strides=2,
padding='same',
activation=None,
use_bias=False)
self.pool1 = nn.NormalMaxPool2D(pool_size=2, strides=2)
self.bn1 = nn.NormalBatchNormalization(center=False, scale=False)
self.quantize = nn.EnterInteger(1.0)
self.conv2 = nn.BinaryConv2D(
filters=256,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn2 = nn.BatchNormalization(self.conv2)
self.conv3 = nn.BinaryConv2D(
filters=256,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn3 = nn.BatchNormalization(self.conv3)
self.conv4 = nn.BinaryConv2D(
filters=256,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.pool4 = nn.MaxPool2D(pool_size=2, strides=2)
self.bn4 = nn.BatchNormalization(self.conv4)
self.conv5 = nn.BinaryConv2D(
filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn5 = nn.BatchNormalization(self.conv5)
self.conv6 = nn.BinaryConv2D(
filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn6 = nn.BatchNormalization(self.conv6)
self.conv7 = nn.BinaryConv2D(
filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.pool7 = nn.MaxPool2D(pool_size=2, strides=2)
self.bn7 = nn.BatchNormalization(self.conv7)
self.conv8 = nn.BinaryConv2D(
filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn8 = nn.BatchNormalization(self.conv8)
self.conv9 = nn.BinaryConv2D(
filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn9 = nn.BatchNormalization(self.conv9)
self.conv10 = nn.BinaryConv2D(
filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.pool10 = nn.MaxPool2D(pool_size=2, strides=2)
self.bn10 = nn.BatchNormalization(self.conv10)
self.flatten = nn.Flatten()
def __init__(self, classes=1000):
super(resnet18, self).__init__()
# Input Layer
self.conv1 = nn.NormalConv2D(filters=64,
kernel_size=7,
strides=2,
padding='same',
activation=None,
use_bias=False)
self.pool1 = nn.NormalMaxPool2D(pool_size=3, strides=2, padding='same')
self.bn1 = nn.NormalBatchNormalization(center=False, scale=False)
self.quantize = nn.EnterInteger(1.0)
# BasicBlock 1
self.block1_conv1 = nn.BinaryConv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block1_bn1 = nn.BatchNormalization()
self.block1_conv2 = nn.BinaryConv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block1_bn2 = nn.BatchNormalization()
self.block1_res = nn.ResidualConnect()
# BasicBlock 2
self.block2_conv1 = nn.BinaryConv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block2_bn1 = nn.BatchNormalization()
self.block2_conv2 = nn.BinaryConv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block2_bn2 = nn.BatchNormalization()
self.block2_res = nn.ResidualConnect()
# BasicBlock 3
self.block3_conv1 = nn.BinaryConv2D(filters=128,
kernel_size=3,
strides=2,
padding='same',
activation=None,
use_bias=False)
self.block3_bn1 = nn.BatchNormalization()
self.block3_conv2 = nn.BinaryConv2D(filters=128,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block3_bn2 = nn.BatchNormalization()
self.block3_down_conv = nn.BinaryConv2D(filters=128,
kernel_size=1,
strides=2,
padding='valid',
activation=None,
use_bias=False)
self.block3_down_bn = nn.BatchNormalization()
self.block3_res = nn.ResidualConnect()
# BasicBlock 4
self.block4_conv1 = nn.BinaryConv2D(filters=128,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block4_bn1 = nn.BatchNormalization()
self.block4_conv2 = nn.BinaryConv2D(filters=128,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block4_bn2 = nn.BatchNormalization()
self.block4_res = nn.ResidualConnect()
# BasicBlock 5
self.block5_conv1 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=2,
padding='same',
activation=None,
use_bias=False)
self.block5_bn1 = nn.BatchNormalization()
self.block5_conv2 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block5_bn2 = nn.BatchNormalization()
self.block5_down_conv = nn.BinaryConv2D(filters=256,
kernel_size=1,
strides=2,
padding='valid',
activation=None,
use_bias=False)
self.block5_down_bn = nn.BatchNormalization()
self.block5_res = nn.ResidualConnect()
# BasicBlock 6
self.block6_conv1 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block6_bn1 = nn.BatchNormalization()
self.block6_conv2 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block6_bn2 = nn.BatchNormalization()
self.block6_res = nn.ResidualConnect()
# BasicBlock 7
self.block7_conv1 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=2,
padding='same',
activation=None,
use_bias=False)
self.block7_bn1 = nn.BatchNormalization()
self.block7_conv2 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block7_bn2 = nn.BatchNormalization()
self.block7_down_conv = nn.BinaryConv2D(filters=512,
kernel_size=1,
strides=2,
padding='valid',
activation=None,
use_bias=False)
self.block7_down_bn = nn.BatchNormalization()
self.block7_res = nn.ResidualConnect()
# BasicBlock 8
self.block8_conv1 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block8_bn1 = nn.BatchNormalization()
self.block8_conv2 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=1,
padding='same',
activation=None,
use_bias=False)
self.block8_bn2 = nn.BatchNormalization()
self.block8_res = nn.ResidualConnect()
self.avg_pool = nn.GlobalAveragePooling2D()
self.flatten = nn.Flatten()
self.dense = nn.BinaryDense(classes, use_bias=False)
self.scalu = nn.Scalu()
def __init__(self, classes=1000):
super(SqueezeNet, self).__init__()
self.classes = classes
self.c0 = nn.NormalConv2D(kernel_size=7,
strides=4,
filters=96,
padding='same',
activation='relu')
self.b0 = nn.NormalBatchNormalization(momentum=bnmomemtum)
#self.mp0 = nn.MaxPool2D(pool_size=2)
self.enter_int = nn.EnterInteger(scale=1.0)
# Fire 1
self.f1c1 = nn.BinaryConv2D(filters=32,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f1b1 = nn.BatchNormalization(self.f1c1, momentum=bnmomemtum)
self.f1c2 = nn.BinaryConv2D(filters=64,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f1b2 = nn.BatchNormalization(self.f1c2, momentum=bnmomemtum)
self.f1c3 = nn.BinaryConv2D(filters=64,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f1b3 = nn.BatchNormalization(self.f1c3, momentum=bnmomemtum)
self.f1concat = tf.keras.layers.Concatenate(axis=-1)
# Fire 2
self.f2c1 = nn.BinaryConv2D(filters=32,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f2b1 = nn.BatchNormalization(self.f2c1, momentum=bnmomemtum)
self.f2c2 = nn.BinaryConv2D(filters=64,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f2b2 = nn.BatchNormalization(self.f2c2, momentum=bnmomemtum)
self.f2c3 = nn.BinaryConv2D(filters=64,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f2b3 = nn.BatchNormalization(self.f2c3, momentum=bnmomemtum)
self.f2concat = tf.keras.layers.Concatenate(axis=-1)
# Fire 3
self.f3c1 = nn.BinaryConv2D(filters=32,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f3b1 = nn.BatchNormalization(self.f3c1, momentum=bnmomemtum)
self.f3c2 = nn.BinaryConv2D(filters=128,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f3b2 = nn.BatchNormalization(self.f3c2, momentum=bnmomemtum)
self.f3c3 = nn.BinaryConv2D(filters=128,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f3b3 = nn.BatchNormalization(self.f3c3, momentum=bnmomemtum)
self.f3concat = tf.keras.layers.Concatenate(axis=-1)
self.mp3 = tf.keras.layers.MaxPooling2D(pool_size=2)
# Fire 4
self.f4c1 = nn.BinaryConv2D(filters=32,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f4b1 = nn.BatchNormalization(self.f4c1, momentum=bnmomemtum)
self.f4c2 = nn.BinaryConv2D(filters=128,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f4b2 = nn.BatchNormalization(self.f4c2, momentum=bnmomemtum)
self.f4c3 = nn.BinaryConv2D(filters=128,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f4b3 = nn.BatchNormalization(self.f4c3, momentum=bnmomemtum)
self.f4concat = tf.keras.layers.Concatenate(axis=-1)
# Fire 5
self.f5c1 = nn.BinaryConv2D(filters=64,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f5b1 = nn.BatchNormalization(self.f5c1, momentum=bnmomemtum)
self.f5c2 = nn.BinaryConv2D(filters=192,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f5b2 = nn.BatchNormalization(self.f5c2, momentum=bnmomemtum)
self.f5c3 = nn.BinaryConv2D(filters=192,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f5b3 = nn.BatchNormalization(self.f5c3, momentum=bnmomemtum)
self.f5concat = tf.keras.layers.Concatenate(axis=-1)
# Fire 6
self.f6c1 = nn.BinaryConv2D(filters=64,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f6b1 = nn.BatchNormalization(self.f6c1, momentum=bnmomemtum)
self.f6c2 = nn.BinaryConv2D(filters=192,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f6b2 = nn.BatchNormalization(self.f6c2, momentum=bnmomemtum)
self.f6c3 = nn.BinaryConv2D(filters=192,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f6b3 = nn.BatchNormalization(self.f6c3, momentum=bnmomemtum)
self.f6concat = tf.keras.layers.Concatenate(axis=-1)
# Fire 7
self.f7c1 = nn.BinaryConv2D(filters=64,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f7b1 = nn.BatchNormalization(self.f7c1, momentum=bnmomemtum)
self.f7c2 = nn.BinaryConv2D(filters=256,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f7b2 = nn.BatchNormalization(self.f7c2, momentum=bnmomemtum)
self.f7c3 = nn.BinaryConv2D(filters=256,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f7b3 = nn.BatchNormalization(self.f7c3, momentum=bnmomemtum)
self.f7concat = tf.keras.layers.Concatenate(axis=-1)
self.mp7 = tf.keras.layers.MaxPooling2D(pool_size=2)
# Fire 8
self.f8c1 = nn.BinaryConv2D(filters=64,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f8b1 = nn.BatchNormalization(self.f8c1, momentum=bnmomemtum)
self.f8c2 = nn.BinaryConv2D(filters=256,
kernel_size=1,
activation='relu',
padding='same',
use_bias=False)
self.f8b2 = nn.BatchNormalization(self.f8c2, momentum=bnmomemtum)
self.f8c3 = nn.BinaryConv2D(filters=256,
kernel_size=3,
activation='relu',
padding='same',
use_bias=False)
self.f8b3 = nn.BatchNormalization(self.f8c3, momentum=bnmomemtum)
self.f8concat = tf.keras.layers.Concatenate(axis=-1)
self.exit_int = nn.ExitInteger()
# Output
self.avgpool = tf.keras.layers.GlobalAveragePooling2D()
self.classifier = tf.keras.layers.Dense(1000, activation='softmax')
def __init__(self, classes=1000):
super(vgg11, self).__init__()
# Set up configurable maxpool or stride dimension reduction.
self.scope = nn.Config.current
use_maxpool = self.scope.use_maxpool
if use_maxpool:
reduce_stride = 1
else:
reduce_stride = 2
self.conv1 = nn.NormalConv2D(filters=64,
kernel_size=3,
strides=1,
padding='same',
activation='relu')
self.bn1 = nn.NormalBatchNormalization()
self.pool1 = nn.NormalMaxPool2D(pool_size=2, strides=2)
self.quantize = nn.EnterInteger(1.0)
self.conv2 = nn.BinaryConv2D(filters=128,
kernel_size=3,
strides=reduce_stride,
padding='same',
activation='relu',
use_bias=False)
self.bn2 = nn.BatchNormalization()
self.pool2 = nn.MaxPool2D(pool_size=2, strides=2)
self.conv3 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn3 = nn.BatchNormalization()
self.conv4 = nn.BinaryConv2D(filters=256,
kernel_size=3,
strides=reduce_stride,
padding='same',
activation='relu',
use_bias=False)
self.bn4 = nn.BatchNormalization()
self.pool3 = nn.MaxPool2D(pool_size=2, strides=2)
self.conv5 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn5 = nn.BatchNormalization()
self.conv6 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=reduce_stride,
padding='same',
activation='relu',
use_bias=False)
self.bn6 = nn.BatchNormalization()
self.pool4 = nn.MaxPool2D(pool_size=2, strides=2)
self.conv7 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=1,
padding='same',
activation='relu',
use_bias=False)
self.bn7 = nn.BatchNormalization()
self.conv8 = nn.BinaryConv2D(filters=512,
kernel_size=3,
strides=reduce_stride,
padding='same',
activation='relu',
use_bias=False)
self.bn8 = nn.BatchNormalization()
self.pool5 = nn.MaxPool2D(pool_size=2, strides=2)
self.avgpool = nn.GlobalAveragePooling2D()
self.classifier = nn.BinaryDense(classes, use_bias=False)
self.scalu = nn.Scalu()
self.softmax = nn.Activation('softmax')