def _build_graph(self, inputs):
image, label = inputs
blocks = CFG[DEPTH]
bottleneck = functools.partial(resnet_bottleneck, stride_first=True)
# tensorflow with padding=SAME will by default pad [2,3] here.
# but caffe conv with stride will pad [3,3]
image = tf.pad(image, [[0, 0], [3, 3], [3, 3], [0, 0]])
image = tf.transpose(image, [0, 3, 1, 2])
with argscope([Conv2D, MaxPooling, GlobalAvgPooling, BatchNorm],
data_format='NCHW'), \
argscope(Conv2D, nl=tf.identity, use_bias=False,
W_init=variance_scaling_initializer(mode='FAN_OUT')):
logits = (LinearWrap(image).Conv2D(
'conv0', 64, 7, stride=2, nl=BNReLU,
padding='VALID').MaxPooling(
'pool0', shape=3, stride=2, padding='SAME').apply(
resnet_group, 'group0',
bottleneck, 64, blocks[0], 1).apply(
resnet_group, 'group1',
bottleneck, 128, blocks[1], 2).apply(
resnet_group, 'group2', bottleneck,
256, blocks[2], 2).apply(
resnet_group, 'group3', bottleneck, 512,
blocks[3],
2).GlobalAvgPooling('gap').FullyConnected(
'linear', 1000, nl=tf.identity)())
prob = tf.nn.softmax(logits, name='prob')
ImageNetModel.compute_loss_and_error(logits, label)
def build_graph(self, image, label):
blocks = CFG[DEPTH]
bottleneck = functools.partial(resnet_bottleneck, stride_first=True)
# tensorflow with padding=SAME will by default pad [2,3] here.
# but caffe conv with stride will pad [3,2]
image = tf.pad(image, [[0, 0], [3, 2], [3, 2], [0, 0]])
image = tf.transpose(image, [0, 3, 1, 2])
with argscope([Conv2D, MaxPooling, GlobalAvgPooling, BatchNorm],
data_format='channels_first'), \
argscope(Conv2D, use_bias=False):
logits = (LinearWrap(image).Conv2D(
'conv0', 64, 7, strides=2, activation=BNReLU,
padding='VALID').MaxPooling(
'pool0', 3, strides=2, padding='SAME').apply(
resnet_group, 'group0',
bottleneck, 64, blocks[0], 1).apply(
resnet_group, 'group1',
bottleneck, 128, blocks[1], 2).apply(
resnet_group, 'group2', bottleneck,
256, blocks[2], 2).apply(
resnet_group, 'group3', bottleneck, 512,
blocks[3],
2).GlobalAvgPooling('gap').FullyConnected(
'linear', 1000)())
tf.nn.softmax(logits, name='prob')
ImageNetModel.compute_loss_and_error(logits, label)
def build_graph(self, image, label):
blocks = CFG[DEPTH]
bottleneck = functools.partial(resnet_bottleneck, stride_first=True)
# tensorflow with padding=SAME will by default pad [2,3] here.
# but caffe conv with stride will pad [3,2]
image = tf.pad(image, [[0, 0], [3, 2], [3, 2], [0, 0]])
image = tf.transpose(image, [0, 3, 1, 2])
with argscope([Conv2D, MaxPooling, GlobalAvgPooling, BatchNorm],
data_format='channels_first'), \
argscope(Conv2D, use_bias=False):
logits = (LinearWrap(image)
.Conv2D('conv0', 64, 7, strides=2, activation=BNReLU, padding='VALID')
.MaxPooling('pool0', 3, strides=2, padding='SAME')
.apply(resnet_group, 'group0', bottleneck, 64, blocks[0], 1)
.apply(resnet_group, 'group1', bottleneck, 128, blocks[1], 2)
.apply(resnet_group, 'group2', bottleneck, 256, blocks[2], 2)
.apply(resnet_group, 'group3', bottleneck, 512, blocks[3], 2)
.GlobalAvgPooling('gap')
.FullyConnected('linear', 1000)())
tf.nn.softmax(logits, name='prob')
ImageNetModel.compute_loss_and_error(logits, label)
def image_preprocess(image):
image = ImageNetModel.image_preprocess(image)
image = tf.transpose(image, [0, 3, 1, 2])
return image
def build_graph(self, image, label):
image = image / 256.0
fw, fa, fg = get_dorefa(BITW, BITA, BITG)
def new_get_variable(v):
name = v.op.name
# don't binarize first and last layer
if not name.endswith('W') or 'conv1' in name or 'fct' in name:
return v
else:
logger.info("Binarizing weight {}".format(v.op.name))
return fw(v)
def nonlin(x):
return tf.clip_by_value(x, 0.0, 1.0)
def activate(x):
return fa(nonlin(x))
def resblock(x, channel, stride):
def get_stem_full(x):
return (LinearWrap(x).Conv2D(
'c3x3a', channel,
3).BatchNorm('stembn').apply(activate).Conv2D(
'c3x3b', channel, 3)())
channel_mismatch = channel != x.get_shape().as_list()[3]
if stride != 1 or channel_mismatch or 'pool1' in x.name:
# handling pool1 is to work around an architecture bug in our model
if stride != 1 or 'pool1' in x.name:
x = AvgPooling('pool', x, stride, stride)
x = BatchNorm('bn', x)
x = activate(x)
shortcut = Conv2D('shortcut', x, channel, 1)
stem = get_stem_full(x)
else:
shortcut = x
x = BatchNorm('bn', x)
x = activate(x)
stem = get_stem_full(x)
return shortcut + stem
def group(x, name, channel, nr_block, stride):
with tf.variable_scope(name + 'blk1'):
x = resblock(x, channel, stride)
for i in range(2, nr_block + 1):
with tf.variable_scope(name + 'blk{}'.format(i)):
x = resblock(x, channel, 1)
return x
with remap_variables(new_get_variable), \
argscope(BatchNorm, decay=0.9, epsilon=1e-4), \
argscope(Conv2D, use_bias=False, nl=tf.identity):
logits = (
LinearWrap(image)
# use explicit padding here, because our private training framework has
# different padding mechanisms from TensorFlow
.tf.pad([[0, 0], [3, 2], [3, 2], [0, 0]]).Conv2D(
'conv1', 64, 7, stride=2, padding='VALID',
use_bias=True).tf.pad(
[[0, 0], [1, 1], [1, 1], [0, 0]],
'SYMMETRIC').MaxPooling(
'pool1',
3, 2, padding='VALID').apply(
group, 'conv2', 64,
2, 1).apply(group, 'conv3', 128, 2, 2).apply(
group, 'conv4', 256, 2,
2).apply(group, 'conv5', 512, 2,
2).BatchNorm('lastbn').
apply(nonlin).GlobalAvgPooling('gap').tf.multiply(
49) # this is due to a bug in our model design
.FullyConnected('fct', 1000)())
tf.nn.softmax(logits, name='output')
ImageNetModel.compute_loss_and_error(logits, label)
def _build_graph(self, inputs):
image, label = inputs
image = image / 256.0
fw, fa, fg = get_dorefa(BITW, BITA, BITG)
def new_get_variable(v):
name = v.op.name
# don't binarize first and last layer
if not name.endswith('W') or 'conv1' in name or 'fct' in name:
return v
else:
logger.info("Binarizing weight {}".format(v.op.name))
return fw(v)
def nonlin(x):
return tf.clip_by_value(x, 0.0, 1.0)
def activate(x):
return fa(nonlin(x))
def resblock(x, channel, stride):
def get_stem_full(x):
return (LinearWrap(x)
.Conv2D('c3x3a', channel, 3)
.BatchNorm('stembn')
.apply(activate)
.Conv2D('c3x3b', channel, 3)())
channel_mismatch = channel != x.get_shape().as_list()[3]
if stride != 1 or channel_mismatch or 'pool1' in x.name:
# handling pool1 is to work around an architecture bug in our model
if stride != 1 or 'pool1' in x.name:
x = AvgPooling('pool', x, stride, stride)
x = BatchNorm('bn', x)
x = activate(x)
shortcut = Conv2D('shortcut', x, channel, 1)
stem = get_stem_full(x)
else:
shortcut = x
x = BatchNorm('bn', x)
x = activate(x)
stem = get_stem_full(x)
return shortcut + stem
def group(x, name, channel, nr_block, stride):
with tf.variable_scope(name + 'blk1'):
x = resblock(x, channel, stride)
for i in range(2, nr_block + 1):
with tf.variable_scope(name + 'blk{}'.format(i)):
x = resblock(x, channel, 1)
return x
with remap_variables(new_get_variable), \
argscope(BatchNorm, decay=0.9, epsilon=1e-4), \
argscope(Conv2D, use_bias=False, nl=tf.identity):
logits = (LinearWrap(image)
# use explicit padding here, because our training framework has
# different padding mechanisms from TensorFlow
.tf.pad([[0, 0], [3, 2], [3, 2], [0, 0]])
.Conv2D('conv1', 64, 7, stride=2, padding='VALID', use_bias=True)
.tf.pad([[0, 0], [1, 1], [1, 1], [0, 0]], 'SYMMETRIC')
.MaxPooling('pool1', 3, 2, padding='VALID')
.apply(group, 'conv2', 64, 2, 1)
.apply(group, 'conv3', 128, 2, 2)
.apply(group, 'conv4', 256, 2, 2)
.apply(group, 'conv5', 512, 2, 2)
.BatchNorm('lastbn')
.apply(nonlin)
.GlobalAvgPooling('gap')
.tf.multiply(49) # this is due to a bug in our model design
.FullyConnected('fct', 1000)())
tf.nn.softmax(logits, name='output')
ImageNetModel.compute_loss_and_error(logits, label)
def image_preprocess(image):
image = ImageNetModel.image_preprocess(image)
image = tf.transpose(image, [0, 3, 1, 2])
return image