def preresnet_basicblock(l, ch_out, stride, preact, block_type='B'):
l, shortcut = apply_preactivation(l, preact, 'basic')
l = Conv2DQuant('conv1', l, ch_out, 3, stride=stride, nl=BNReLU)
l = QuantizedActiv('quant2', l)
l = Conv2DQuant('conv2', l, ch_out, 3, nl=get_bn(zero_init=False))
return l + resnet_shortcut(shortcut,
ch_out,
stride,
nl=get_bn(zero_init=False),
block_type=block_type)
def add_layer(name, l):
shape = l.get_shape().as_list()
in_channel = shape[1]
with tf.variable_scope(name) as scope:
c = Conv2DQuant('conv1x1', l, 4 * GROWTH_RATE, 1)
c = BNReLU('bnrelu_2', c)
c = QuantizedActiv('quant2', c)
c = Conv2DQuant('conv3x3', c, GROWTH_RATE, 3)
c = BNReLU('bnrelu_3', c)
c = QuantizedActiv('quant3', c)
l = tf.concat([c, l], 1)
return l
def preresnet_bottleneck(l, ch_out, stride, preact, block_type='A'):
# stride is applied on the second conv, following fb.resnet.torch
l, shortcut = apply_preactivation(l, preact, 'basic')
l = Conv2DQuant('conv1', l, ch_out, 1, nl=BNReLU)
l = QuantizedActiv('quant2', l)
l = Conv2DQuant('conv2', l, ch_out, 3, stride=stride, nl=BNReLU)
l = QuantizedActiv('quant3', l)
l = Conv2DQuant('conv3', l, ch_out * 4, 1, nl=get_bn(zero_init=False))
return l + resnet_shortcut(shortcut,
ch_out * 4,
stride,
nl=get_bn(zero_init=False),
block_type=block_type)
def add_transition(name, l):
shape = l.get_shape().as_list()
in_channel = shape[1]
out_channel = math.floor(in_channel * REDUCTION)
with tf.variable_scope(name) as scope:
l = Conv2DQuant('conv1', l, out_channel, 1, stride=1, use_bias=False)
l = AvgPooling('pool', l, 2)
return l
def resnet_bottleneck(l, ch_out, stride, stride_first=False):
"""
stride_first: original resnet put stride on first conv. fb.resnet.torch put stride on second conv.
"""
shortcut = l
l = Conv2DQuant('conv1',
l,
ch_out,
1,
stride=stride if stride_first else 1,
nl=BNReLU)
l = Conv2DQuant('conv2',
l,
ch_out,
3,
stride=1 if stride_first else stride,
nl=BNReLU)
l = Conv2DQuant('conv3', l, ch_out * 4, 1, nl=get_bn(zero_init=True))
return l + resnet_shortcut(
shortcut, ch_out * 4, stride, nl=get_bn(zero_init=False))
def resnet_shortcut(l, n_out, stride, nl=tf.identity, block_type='B'):
data_format = get_arg_scope()['Conv2DQuant']['data_format']
n_in = l.get_shape().as_list()[1 if data_format == 'NCHW' else 3]
if n_in != n_out: # change dimension when channel is not the same
if block_type == 'B':
return Conv2DQuant('convshortcut',
l,
n_out,
1,
stride=stride,
nl=nl)
else:
l = AvgPooling('poolshortcut', l, stride, stride, padding='VALID')
if data_format == 'NCHW':
paddings = [[0, 0], [0, n_out - n_in], [0, 0], [0, 0]]
else:
paddings = [[0, 0], [0, 0], [0, 0], [0, n_out - n_in]]
l = tf.pad(l, paddings, 'CONSTANT')
return l
else:
return l
def inception_block(l,
name,
ch_1x1,
ch_3x3,
ch_5x5,
is_last_block=False,
is_last=False):
data_format = get_arg_scope()['Conv2DQuant']['data_format']
with tf.variable_scope(name):
conv1x1 = Conv2DQuant(
'1x1',
l,
ch_1x1,
1,
nl=getBNReLUQuant if not is_last_block else tf.identity)
conv3x3_reduce = Conv2DQuant('3x3_reduce',
l,
ch_3x3,
1,
nl=getBNReLUQuant)
conv3x3 = Conv2DQuant(
'3x3',
conv3x3_reduce,
ch_3x3,
3,
nl=getBNReLUQuant if not is_last_block else tf.identity)
conv5x5_reduce = Conv2DQuant('5x5_reduce',
l,
ch_5x5,
1,
nl=getBNReLUQuant)
conv5x5 = Conv2DQuant(
'5x5',
conv5x5_reduce,
ch_5x5,
5,
nl=getBNReLUQuant if not is_last_block else tf.identity)
if is_last_block and not is_last:
conv1x1 = MaxPooling('pool_1x1',
conv1x1,
shape=3,
stride=2,
padding='SAME')
conv1x1 = BNReLU('conv1x1_bn', conv1x1)
conv1x1 = QuantizedActiv('conv1x1_quant', conv1x1)
conv3x3 = MaxPooling('pool_3x3',
conv3x3,
shape=3,
stride=2,
padding='SAME')
conv3x3 = BNReLU('conv3x3_bn', conv3x3)
conv3x3 = QuantizedActiv('conv3x3_quant', conv3x3)
conv5x5 = MaxPooling('pool_5x5',
conv5x5,
shape=3,
stride=2,
padding='SAME')
conv5x5 = BNReLU('conv5x5_bn', conv5x5)
conv5x5 = QuantizedActiv('conv5x5_quant', conv5x5)
l = tf.concat([conv1x1, conv3x3, conv5x5],
1 if data_format == 'NCHW' else 3,
name='concat')
if is_last:
l = BNReLU('output_bn', l)
return l
def resnet_basicblock(l, ch_out, stride):
shortcut = l
l = Conv2DQuant('conv1', l, ch_out, 3, stride=stride, nl=BNReLU)
l = Conv2DQuant('conv2', l, ch_out, 3, nl=get_bn(zero_init=True))
return l + resnet_shortcut(
shortcut, ch_out, stride, nl=get_bn(zero_init=False))