def get_graph(train=False):
g = Graph('YOLOv2-Test: 16-bit',
dataset='imagenet',
log_level=logging.INFO)
batch_size = 1
with g.as_default():
with g.name_scope('inputs'):
# Input dimensions are (Batch_size, Height, Width, Channels)
i = get_tensor(shape=(batch_size, 28, 28, 1),
name='data',
dtype=FQDtype.FXP16,
trainable=False)
with g.name_scope('conv0'):
# Weight dimensions are (Output Channels, Kernel Height, Kernel Width, Input Channels)
weights = get_tensor(shape=(20, 5, 5, 1),
name='weights',
dtype=FixedPoint(16, 12))
# Bias dimensions are (Output Channels,)
biases = get_tensor(shape=(20),
name='biases',
dtype=FixedPoint(32, 20))
# Intermediate data dimensions are (Batch_size, Height, Width, Channels)
conv = conv2D(i,
weights,
biases,
pad='VALID',
dtype=FixedPoint(16, 12))
return g
def get_LSTM(name, size):
g = Graph(name, dataset='PTB', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, size * 4),
name='data',
dtype=FQDtype.FXP4,
trainable=False)
with g.name_scope('matmul'):
out = fc(i,
output_channels=4 * size,
w_dtype=FQDtype.FXP4,
f_dtype=None)
return g
def get_bench_nn(bench_name, WRPN=False):
g = Graph(bench_name, dataset='nop', log_level=logging.INFO)
batch_size = 16
hyparamater = parase_bench_name(bench_name)
width = hyparamater['width']
width = int(width)
height = hyparamater['height']
height = int(height)
cin = hyparamater['cin']
cin = int(cin)
cout = hyparamater['cout']
cout = int(cout)
kernel_size = hyparamater['kernel']
kernel_size = int(kernel_size)
stride = hyparamater['stride']
stride = (1, int(stride), int(stride), 1)
pad = hyparamater['pad'].upper()
group = hyparamater['group']
group = int(group)
fb = hyparamater['fb']
wb = hyparamater['wb']
fb = {'8': FQDtype.FXP8, '4': FQDtype.FXP4, '2': FQDtype.FXP2}[fb]
wb = {'8': FQDtype.FXP8, '4': FQDtype.FXP4, '2': FQDtype.FXP2}[wb]
with g.as_default():
with g.name_scope('inputs'):
prev = get_tensor(shape=(batch_size, width, height, cin),
name='data',
dtype=FQDtype.FXP8,
trainable=False)
with g.name_scope('conv1'): # prepare
prev = conv(prev,
filters=cin,
kernel_size=1,
stride=(1, 1, 1, 1),
pad='SAME',
c_dtype=FQDtype.FXP8,
w_dtype=FQDtype.FXP8)
if 'base' in bench_name:
return g
with g.name_scope('conv2'): # profile
prev = conv(prev,
filters=cout,
kernel_size=kernel_size,
stride=stride,
pad=pad,
group=group,
c_dtype=fb,
w_dtype=wb)
return g
def get_graph(train=False):
g = Graph('YOLOv2-Test: 16-bit',
dataset='imagenet',
log_level=logging.INFO)
batch_size = 1
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 416, 416, 3),
name='data',
dtype=FQDtype.FXP16,
trainable=False)
with g.name_scope('conv0'):
conv0 = yolo_convolution(i,
filters=16,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 12),
s_dtype=FixedPoint(16, 9),
bn_dtype=FixedPoint(16, 8))
with g.name_scope('pool0'):
pool0 = maxPool(conv0,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv1'):
conv1 = yolo_convolution(pool0,
filters=32,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 8),
s_dtype=FixedPoint(16, 14),
bn_dtype=FixedPoint(16, 8))
with g.name_scope('pool1'):
pool1 = maxPool(conv1,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2'):
conv2 = yolo_convolution(
pool1,
filters=64,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
# batch_normalize=False, act='linear',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 10),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 9))
with g.name_scope('pool2'):
pool2 = maxPool(conv2,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv3'):
conv3 = yolo_convolution(pool2,
filters=128,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 10),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 10))
with g.name_scope('pool3'):
pool3 = maxPool(conv3,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv4'):
conv4 = yolo_convolution(pool3,
filters=256,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 11),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 10))
with g.name_scope('pool4'):
pool4 = maxPool(conv4,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv5'):
conv5 = yolo_convolution(pool4,
filters=512,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 12),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 11))
with g.name_scope('pool5'):
pool5 = maxPool(conv5,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 1, 1, 1),
pad=((0, 0), (0, 1), (0, 1), (0, 0)))
with g.name_scope('conv6'):
conv6 = yolo_convolution(pool5,
filters=1024,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 12),
s_dtype=FixedPoint(16, 11),
bn_dtype=FixedPoint(16, 9))
with g.name_scope('conv7'):
conv7 = yolo_convolution(conv6,
filters=1024,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 11),
s_dtype=FixedPoint(16, 14),
bn_dtype=FixedPoint(16, 12))
with g.name_scope('conv8'):
conv8 = yolo_convolution(conv7,
filters=125,
kernel_size=1,
batch_normalize=False,
act='linear',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 11))
return g
def initialize_pass(self, node, ctx):
self.dnnw_ir['graph'] = Graph(node.name, None, log_level=logging.INFO)
return node
def get_resnet_18_twn():
'''
ResNet-18
Note that this isn't ResNet-18B
'''
g = Graph('ResNet-18-TWN', dataset='ImageNet', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 224, 224, 3),
name='data',
dtype=FQDtype.FXP8,
trainable=False)
with g.name_scope('conv1_a'):
conv1_a = conv(i,
filters=64,
kernel_size=7,
pad='SAME',
stride=(1, 2, 2, 1),
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP8)
with g.name_scope('pool1_a'):
pool1_a = maxPool(conv1_a,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('module2'):
for i in range(1, 3):
if i == 1:
prev = pool1_a
else:
prev = conv2_b
with g.name_scope('conv2_{}_a'.format(i)):
conv2_a = conv(prev,
filters=64,
kernel_size=1,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv2_{}_b'.format(i)):
conv2_b = conv(conv2_a,
filters=64,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('add2'):
add2 = add((conv2_a, conv2_b))
with g.name_scope('module3'):
for i in range(1, 3):
if i == 1:
prev = add2
stride = (1, 2, 2, 1)
else:
prev = conv3_b
stride = (1, 1, 1, 1)
with g.name_scope('conv3_{}_a'.format(i)):
conv3_a = conv(prev,
filters=128,
kernel_size=1,
pad='SAME',
stride=stride,
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv3_{}_b'.format(i)):
conv3_b = conv(conv3_a,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('add3'):
add3 = add((conv3_a, conv3_b))
with g.name_scope('module4'):
for i in range(1, 3):
if i == 1:
prev = add3
stride = (1, 2, 2, 1)
else:
prev = conv4_b
stride = (1, 1, 1, 1)
with g.name_scope('conv4_{}_a'.format(i)):
conv4_a = conv(prev,
filters=256,
kernel_size=1,
pad='SAME',
stride=stride,
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv4_{}_b'.format(i)):
conv4_b = conv(conv4_a,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('add4'):
add4 = add((conv4_a, conv4_b))
with g.name_scope('module5'):
for i in range(1, 3):
if i == 1:
prev = add4
stride = (1, 2, 2, 1)
else:
prev = conv5_b
stride = (1, 1, 1, 1)
with g.name_scope('conv5_{}_a'.format(i)):
conv5_a = conv(prev,
filters=512,
kernel_size=1,
pad='SAME',
stride=stride,
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv5_{}_b'.format(i)):
conv5_b = conv(conv5_a,
filters=512,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
return g
def get_vgg_7_twn():
'''
VGG-7
'''
g = Graph('VGG-7-CIFAR-10', dataset='CIFAR-10', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 32, 32, 3),
name='data',
dtype=FQDtype.FXP2,
trainable=False)
with g.name_scope('conv1_a'):
conv1_a = conv(i,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool1_a'):
pool1_a = maxPool(conv1_a,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv1_b'):
conv1_b = conv(i,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool1_b'):
pool1_b = maxPool(conv1_b,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2_a'):
conv2_a = conv(pool1_a,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool2_a'):
pool2_a = maxPool(conv2_a,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2_b'):
conv2_b = conv(pool1_b,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool2_b'):
pool2_b = maxPool(conv2_b,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv3_a'):
conv3_a = conv(pool2_a,
filters=512,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool3_a'):
pool3_a = maxPool(conv3_a,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv3_b'):
conv3_b = conv(pool2_b,
filters=512,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool3_b'):
pool3_b = maxPool(conv3_b,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('add3'):
add3 = add((pool3_a, pool3_b))
with g.name_scope('flatten3'):
flatten3 = flatten(add3)
with g.name_scope('fc1'):
fc1 = fc(flatten3,
output_channels=1024,
w_dtype=FQDtype.FXP2,
f_dtype=FQDtype.FXP2)
return g
def get_lenet_5_twn():
'''
LeNet-5
TWN
'''
g = Graph('LeNet-5', dataset='MNIST', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 32, 32, 1),
name='data',
dtype=FQDtype.FXP2,
trainable=False)
with g.name_scope('conv0'):
conv0 = conv(i,
filters=32,
kernel_size=5,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool0'):
pool0 = maxPool(conv0,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv1'):
conv1 = conv(pool0,
filters=64,
kernel_size=5,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool1'):
pool1 = maxPool(conv1,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('flatten1'):
flatten1 = flatten(pool1)
with g.name_scope('fc1'):
fc1 = fc(flatten1,
output_channels=512,
w_dtype=FQDtype.FXP2,
f_dtype=FQDtype.FXP2)
with g.name_scope('fc2'):
fc2 = fc(fc1,
output_channels=10,
w_dtype=FQDtype.FXP2,
f_dtype=FQDtype.FXP2)
return g
def get_cifar10_qnn():
'''
CIFAR-10
QNN
'''
g = Graph('CIFAR-10-QNN', dataset='Cifar-10', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 32, 32, 3),
name='data',
dtype=FQDtype.FXP8,
trainable=False)
with g.name_scope('conv0'):
conv0 = conv(i,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP8)
with g.name_scope('conv1'):
conv1 = conv(conv0,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool1'):
pool1 = maxPool(conv1,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2'):
conv2 = conv(pool1,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('conv3'):
conv3 = conv(conv2,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool3'):
pool3 = maxPool(conv3,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv4'):
conv4 = conv(pool3,
filters=512,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('conv5'):
conv5 = conv(conv4,
filters=512,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP2,
w_dtype=FQDtype.FXP2)
with g.name_scope('pool5'):
pool5 = maxPool(conv5,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('flatten5'):
flatten5 = flatten(pool5)
with g.name_scope('fc6'):
fc6 = fc(flatten5,
output_channels=1024,
w_dtype=FQDtype.FXP2,
f_dtype=FQDtype.FXP2)
with g.name_scope('fc7'):
fc7 = fc(fc6,
output_channels=1024,
w_dtype=FQDtype.FXP2,
f_dtype=FQDtype.FXP2)
with g.name_scope('fc8'):
fc8 = fc(fc7,
output_channels=10,
w_dtype=FQDtype.FXP2,
f_dtype=None)
return g
def get_svhn_qnn():
'''
SVHN
QNN
Weights are 1-bit
Inputs are 8-bit for first layer, 1-bit for the rest
'''
g = Graph('SVHN-QNN', dataset='SVHN', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 32, 32, 3),
name='data',
dtype=FQDtype.FXP8,
trainable=False)
with g.name_scope('conv0'):
conv0 = conv(i,
filters=64,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.Bin,
w_dtype=FQDtype.Bin)
with g.name_scope('conv1'):
conv1 = conv(conv0,
filters=64,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.Bin,
w_dtype=FQDtype.Bin)
with g.name_scope('pool1'):
pool1 = maxPool(conv1,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2'):
conv2 = conv(pool1,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.Bin,
w_dtype=FQDtype.Bin)
with g.name_scope('conv3'):
conv3 = conv(conv2,
filters=128,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.Bin,
w_dtype=FQDtype.Bin)
with g.name_scope('pool3'):
pool3 = maxPool(conv3,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv4'):
conv4 = conv(pool3,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.Bin,
w_dtype=FQDtype.Bin)
with g.name_scope('conv5'):
conv5 = conv(conv4,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.Bin,
w_dtype=FQDtype.Bin)
with g.name_scope('pool5'):
pool5 = maxPool(conv5,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('flatten5'):
flatten5 = flatten(pool5)
with g.name_scope('fc6'):
fc6 = fc(flatten5,
output_channels=1024,
w_dtype=FQDtype.Bin,
f_dtype=FQDtype.Bin)
with g.name_scope('fc7'):
fc7 = fc(fc6,
output_channels=1024,
w_dtype=FQDtype.Bin,
f_dtype=FQDtype.Bin)
with g.name_scope('fc8'):
fc8 = fc(fc7,
output_channels=10,
w_dtype=FQDtype.Bin,
f_dtype=None)
return g
def get_alex_net_wrpn():
'''
AlexNet
Krizhevsky, Sutskever, and Hinton, 2012
'''
g = Graph('AlexNet-WRPN (2x)', dataset='imagenet', log_level=logging.INFO)
batch_size = 16
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 227, 227, 3),
name='data',
dtype=FQDtype.FXP8,
trainable=False)
with g.name_scope('conv1_a'):
conv1_a = conv(i,
filters=96,
kernel_size=11,
stride=(1, 4, 4, 1),
pad='VALID',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP8)
with g.name_scope('pool1_a'):
pool1_a = maxPool(conv1_a,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv1_b'):
conv1_b = conv(i,
filters=96,
kernel_size=11,
stride=(1, 4, 4, 1),
pad='VALID',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP8)
with g.name_scope('pool1_b'):
pool1_b = maxPool(conv1_b,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2_a'):
conv2_a = conv(pool1_a,
filters=256,
kernel_size=5,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('pool2_a'):
pool2_a = maxPool(conv2_a,
pooling_kernel=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2_b'):
conv2_b = conv(pool1_b,
filters=256,
kernel_size=5,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('pool2_b'):
pool2_b = maxPool(conv2_b,
pooling_kernel=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('concat2'):
concat2 = concat((pool2_a, pool2_b),
concat_dim=-1,
dtype=FQDtype.FXP4)
with g.name_scope('conv3_a'):
conv3_a = conv(concat2,
filters=384,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv3_b'):
conv3_b = conv(concat2,
filters=384,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv4_a'):
conv4_a = conv(conv3_a,
filters=384,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv4_b'):
conv4_b = conv(conv3_b,
filters=384,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('conv5_a'):
conv5_a = conv(conv4_a,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('pool5_a'):
pool5_a = maxPool(conv5_a,
pooling_kernel=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv5_b'):
conv5_b = conv(conv4_b,
filters=256,
kernel_size=3,
pad='SAME',
c_dtype=FQDtype.FXP4,
w_dtype=FQDtype.FXP4)
with g.name_scope('pool5_b'):
pool5_b = maxPool(conv5_b,
pooling_kernel=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('concat5'):
concat5 = concat((pool5_a, pool5_b), concat_dim=-1)
with g.name_scope('flatten5'):
flatten5 = flatten(concat5)
with g.name_scope('fc1'):
fc1 = fc(flatten5,
output_channels=8192,
w_dtype=FQDtype.FXP4,
f_dtype=FQDtype.FXP4)
with g.name_scope('fc2'):
fc2 = fc(fc1,
output_channels=8192,
w_dtype=FQDtype.FXP4,
f_dtype=FQDtype.FXP8)
with g.name_scope('fc3'):
fc3 = fc(fc2,
output_channels=1000,
w_dtype=FQDtype.FXP8,
f_dtype=None)
return g