def make_network(env):
with env.create_network() as net:
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
h, w, c = 28, 28, 1
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def forward(img):
_ = img
_ = O.conv2d('conv1',
_,
16, (3, 3),
padding='SAME',
nonlin=O.identity)
_ = O.batch_norm('bn1', _)
_ = O.relu(_)
_ = O.pooling2d('pool1', _, kernel=2)
_ = O.conv2d('conv2',
_,
32, (3, 3),
padding='SAME',
nonlin=O.identity)
_ = O.batch_norm('bn2', _)
_ = O.relu(_)
_ = O.pooling2d('pool2', _, kernel=2)
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('fc1', _, 64)
_ = O.fc('fc2', _, 10)
prob = O.softmax(_, name='prob')
pred = _.argmax(axis=1).astype('int32', name='pred')
net.add_output(prob)
net.add_output(pred)
if env.phase is env.Phase.TRAIN:
label = O.placeholder('label', shape=(None, ), dtype='int32')
loss = O.sparse_softmax_cross_entropy_with_logits(
logits=_, labels=label).mean()
loss = O.identity(loss, name='loss')
net.set_loss(loss)
accuracy = O.eq(label, pred).astype('float32').mean()
error = 1. - accuracy
summary.scalar('accuracy', accuracy)
summary.scalar('error', error)
summary.inference.scalar('loss', loss)
summary.inference.scalar('accuracy', accuracy)
summary.inference.scalar('error', error)
def make_network(env):
with env.create_network() as net:
nr_classes = get_env('dataset.nr_classes')
conv_bn_relu = functools.partial(O.conv2d, nonlin=O.bn_relu)
conv2d = conv_bn_relu
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
h, w, c = 32, 32, 3
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def forward(img):
_ = img
_ = conv2d('conv1.1', _, 16, (3, 3), padding='SAME')
_ = conv2d('conv1.2', _, 16, (3, 3), padding='SAME')
_ = O.pooling2d('pool1', _, kernel=3, stride=2)
_ = conv2d('conv2.1', _, 32, (3, 3), padding='SAME')
_ = conv2d('conv2.2', _, 32, (3, 3), padding='SAME')
_ = O.pooling2d('pool2', _, kernel=3, stride=2)
_ = conv2d('conv3.1', _, 64, (3, 3), padding='VALID')
_ = conv2d('conv3.2', _, 64, (3, 3), padding='VALID')
_ = conv2d('conv3.3', _, 64, (3, 3), padding='VALID')
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('fc1', _, 128, nonlin=O.relu)
_ = O.fc('fc2', _, 64, nonlin=O.relu)
_ = O.fc('linear', _, nr_classes)
prob = O.softmax(_, name='prob')
pred = _.argmax(axis=1).astype('int32', name='pred')
net.add_output(prob)
net.add_output(pred)
if env.phase is env.Phase.TRAIN:
label = O.placeholder('label', shape=(None, ), dtype='int32')
loss = O.sparse_softmax_cross_entropy_with_logits(logits=_, labels=label).mean()
loss = O.identity(loss, name='loss')
net.set_loss(loss)
accuracy = O.eq(label, pred).astype('float32').mean()
error = 1. - accuracy
summary.scalar('accuracy', accuracy)
summary.scalar('error', error)
summary.inference.scalar('loss', loss)
summary.inference.scalar('accuracy', accuracy)
summary.inference.scalar('error', error)
def make_network(env):
with env.create_network() as net:
n = 2
nr_classes = get_env('dataset.nr_classes')
conv2d = functools.partial(O.conv2d,
kernel=3,
use_bias=False,
padding='SAME')
conv_bn_relu = functools.partial(conv2d, nonlin=O.bn_relu)
dpc = env.create_dpcontroller()
with dpc.activate():
def inputs():
h, w, c = 32, 32, 3
img = O.placeholder('img', shape=(None, h, w, c))
return [img]
def residual(name, x, first=False, inc_dim=False):
in_channel = x.static_shape[3]
out_channel = in_channel
stride = 1
if inc_dim:
out_channel = in_channel * 2
stride = 2
with env.variable_scope(name):
_ = x if first else O.bn_relu(x)
_ = conv_bn_relu('conv1', _, out_channel, stride=stride)
_ = conv2d('conv2', _, out_channel)
if inc_dim:
x = O.pooling2d('pool', x, kernel=2)
x = O.pad(x, [[0, 0], [0, 0], [0, 0],
[in_channel // 2, in_channel // 2]])
print(name, x.static_shape)
_ = _ + x
return _
def forward(img):
_ = img / 128.0 - 1.0
_ = conv_bn_relu('conv0', _, 16)
_ = residual('res1.0', _, first=True)
for i in range(1, n):
_ = residual('res1.{}'.format(i), _)
_ = residual('res2.0', _, inc_dim=True)
for i in range(1, n):
_ = residual('res2.{}'.format(i), _)
_ = residual('res3.0', _, inc_dim=True)
for i in range(1, n):
_ = residual('res3.{}'.format(i), _)
_ = O.batch_norm('bn_last', _)
_ = O.relu(_)
_ = _.mean(axis=[1, 2]) # global avg pool
dpc.add_output(_, name='feature')
dpc.set_input_maker(inputs).set_forward_func(forward)
_ = dpc.outputs['feature']
_ = O.fc('linear', _, nr_classes)
prob = O.softmax(_, name='prob')
pred = _.argmax(axis=1).astype('int32', name='pred')
net.add_output(prob)
net.add_output(pred)
if env.phase is env.Phase.TRAIN:
label = O.placeholder('label', shape=(None, ), dtype='int32')
loss = O.sparse_softmax_cross_entropy_with_logits(
logits=_, labels=label).mean()
loss = O.identity(loss, name='loss')
net.set_loss(loss)
accuracy = O.eq(label, pred).astype('float32').mean()
error = 1. - accuracy
summary.scalar('accuracy', accuracy)
summary.scalar('error', error)
summary.inference.scalar('loss', loss)
summary.inference.scalar('accuracy', accuracy)
summary.inference.scalar('error', error)