def create_resnet(depth=18):
'''Original resnet, where the there is a relue after the addition layer'''
net = ffnet.FeedForwardNet()
net.add(
Conv2D('input-conv',
64,
7,
2,
pad=3,
use_bias=False,
input_sample_shape=(3, 224, 224)))
net.add(BatchNormalization('input-bn'))
net.add(Activation('input_relu'))
net.add(MaxPooling2D('input_pool', 3, 2, pad=1))
conf = cfg[depth]
if depth > 34:
stage(0, net, conf[0], 64, 64, 256, 1, bottleneck)
stage(1, net, conf[1], 256, 128, 512, 2, bottleneck)
stage(2, net, conf[2], 512, 256, 1024, 2, bottleneck)
stage(3, net, conf[3], 1024, 512, 2048, 2, bottleneck)
else:
stage(0, net, conf[0], 64, 64, 64, 1, basicblock)
stage(1, net, conf[1], 64, 128, 128, 2, basicblock)
stage(2, net, conf[2], 128, 256, 256, 2, basicblock)
stage(3, net, conf[3], 256, 512, 512, 2, basicblock)
net.add(AvgPooling2D('avg', 7, 1, pad=0))
net.add(Flatten('flat'))
net.add(Dense('dense', 1000))
return net
def create_preact_resnet(depth=200):
'''Resnet with the batchnorm and relu moved to before the conv layer for each block'''
net = ffnet.FeedForwardNet()
net.add(
Conv2D('input-conv',
64,
7,
2,
pad=3,
use_bias=False,
input_sample_shape=(3, 224, 224)))
net.add(BatchNormalization('input-bn'))
net.add(Activation('input_relu'))
net.add(MaxPooling2D('input_pool', 3, 2, pad=1))
conf = cfg[depth]
if depth > 34:
stage(0, net, conf[0], 64, 64, 256, 1, bottleneck, preact=True)
stage(1, net, conf[1], 256, 128, 512, 2, bottleneck, preact=True)
stage(2, net, conf[2], 512, 256, 1024, 2, bottleneck, preact=True)
stage(3, net, conf[3], 1024, 512, 2048, 2, bottleneck, preact=True)
else:
stage(0, net, conf[0], 64, 64, 64, 1, basicblock, preact=True)
stage(1, net, conf[1], 64, 128, 128, 2, basicblock, preact=True)
stage(2, net, conf[2], 128, 256, 256, 2, basicblock, preact=True)
stage(3, net, conf[3], 256, 512, 512, 2, basicblock, preact=True)
net.add(BatchNormalization('final-bn'))
net.add(Activation('final-relu'))
net.add(AvgPooling2D('avg', 7, 1, pad=0))
net.add(Flatten('flat'))
net.add(Dense('dense', 1000))
return net
def create_net(depth, nb_classes, batchnorm=False, use_cpu=False):
if use_cpu:
layer.engine = 'singacpp'
net = ffnet.FeedForwardNet()
net = create_layers(net, cfg[depth], (3, 224, 224), batchnorm)
net.add(Flatten('flat'))
net.add(Dense('dense/classifier.0', 4096))
net.add(Activation('act/classifier.1'))
net.add(Dropout('dropout/classifier.2'))
net.add(Dense('dense/classifier.3', 4096))
net.add(Activation('act/classifier.4'))
net.add(Dropout('dropout/classifier.5'))
net.add(Dense('dense/classifier.6', nb_classes))
return net
def add_transition(name, net, n_channels, last=False):
net.add(BatchNormalization('%s/norm' % name))
lyr = net.add(Activation('%s/relu' % name))
if last:
net.add(
AvgPooling2D('%s/pool' % name,
lyr.get_output_sample_shape()[1:3],
pad=0))
net.add(Flatten('flat'))
else:
net.add(
Conv2D('%s/conv' % name,
n_channels,
1,
1,
pad=0,
use_bias=conv_bias))
net.add(AvgPooling2D('%s/pool' % name, 2, 2, pad=0))
def create_net(shape, weight_path='bvlc_googlenet.pickle'):
net = ffnet.FeedForwardNet()
net.add(Conv2D('conv1/7x7_s2', 64, 7, 2, pad=3, input_sample_shape=shape))
c1 = net.add(Activation('conv1/relu_7x7'))
pool1 = pool(net, c1, 'pool1/3x3_s2', 3, 2)
norm1 = net.add(LRN('pool1/norm1', 5, 0.0001, 0.75))
c3x3r = conv(net, norm1, 'conv2', 64, 1, suffix='3x3_reduce')
c3x3 = conv(net, c3x3r, 'conv2', 192, 3, pad=1, suffix='3x3')
norm2 = net.add(LRN('conv2/norm2', 5, 0.0001, 0.75))
pool2 = pool(net, norm2, 'pool2/3x3_s2', 3, 2)
i3a = inception(net, pool2, 'inception_3a', 64, 96, 128, 16, 32, 32)
i3b = inception(net, i3a, 'inception_3b', 128, 128, 192, 32, 96, 64)
pool3 = pool(net, i3b, 'pool3/3x3_s2', 3, 2)
i4a = inception(net, pool3, 'inception_4a', 192, 96, 208, 16, 48, 64)
i4b = inception(net, i4a, 'inception_4b', 160, 112, 224, 24, 64, 64)
i4c = inception(net, i4b, 'inception_4c', 128, 128, 256, 24, 64, 64)
i4d = inception(net, i4c, 'inception_4d', 112, 144, 288, 32, 64, 64)
i4e = inception(net, i4d, 'inception_4e', 256, 160, 320, 32, 128, 128)
pool4 = pool(net, i4e, 'pool4/3x3_s2', 3, 2)
i5a = inception(net, pool4, 'inception_5a', 256, 160, 320, 32, 128, 128)
i5b = inception(net, i5a, 'inception_5b', 384, 192, 384, 48, 128, 128)
pool5 = net.add(AvgPooling2D('pool5/7x7_s1', 7, 1, pad=0))
drop5 = net.add(Dropout('drop', 0.4))
flat = net.add(Flatten('flat'))
dense = net.add(Dense('loss3/classifier', 1000))
# prob=net.add(Softmax('softmax'))
net.load(weight_path, use_pickle=True)
print('total num of params %d' % (len(net.param_names())))
# SINGA and Caffe have different layout for the weight matrix of the dense
# layer
for key, val in zip(net.param_names(), net.param_values()):
# print key
if key == 'loss3/classifier_weight' or key == 'loss3/classifier/weight':
tmp = tensor.to_numpy(val)
tmp = tmp.reshape(tmp.shape[::-1])
val.copy_from_numpy(np.transpose(tmp))
return net
def create_wide_resnet(depth=50):
'''Similar original resnet except that a<=b<=c for the bottleneck block'''
net = ffnet.FeedForwardNet()
net.add(
Conv2D('input-conv',
64,
7,
2,
pad=3,
use_bias=False,
input_sample_shape=(3, 224, 224)))
net.add(BatchNormalization('input-bn'))
net.add(Activation('input_relu'))
net.add(MaxPooling2D('input_pool', 3, 2, pad=1))
stage(0, net, 3, 64, 128, 256, 1, bottleneck)
stage(1, net, 4, 256, 256, 512, 2, bottleneck)
stage(2, net, 6, 512, 512, 1024, 2, bottleneck)
stage(3, net, 3, 1024, 1024, 2048, 2, bottleneck)
net.add(AvgPooling2D('avg_pool', 7, 1, pad=0))
net.add(Flatten('flag'))
net.add(Dense('dense', 1000))
return net
