def darknet53():
nn.conv(1,32,3,1,1,'leaky')
darknet_block(32*ratio,1)
darknet_block(64*ratio,2)
darknet_block(128*ratio,8)
darknet_block(256*ratio,8)
darknet_block(512*ratio,4)
def inference(images, keep_probability, phase_train=True, weight_decay=0.0):
""" Define an inference network for face recognition based
on inception modules using batch normalization
Args:
images: The images to run inference on, dimensions batch_size x height x width x channels
phase_train: True if batch normalization should operate in training mode
"""
# images = np.reshape(images, [-1, IMAGE_SIZE, IMAGE_SIZE, 3])
# print("nn4, images : ", images.shape)
images = tf.reshape(images, [-1, IMAGE_SIZE, IMAGE_SIZE, 3])
endpoints = {}
net = network.conv(images, 3, 64, 7, 7, 2, 2, 'SAME', 'conv1_7x7', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['conv1'] = net
net = network.mpool(net, 3, 3, 2, 2, 'SAME', 'pool1')
endpoints['pool1'] = net
net = network.conv(net, 64, 64, 1, 1, 1, 1, 'SAME', 'conv2_1x1', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['conv2_1x1'] = net
net = network.conv(net, 64, 192, 3, 3, 1, 1, 'SAME', 'conv3_3x3', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['conv3_3x3'] = net
net = network.mpool(net, 3, 3, 2, 2, 'SAME', 'pool3')
endpoints['pool3'] = net
net = network.inception(net, 192, 1, 64, 96, 128, 16, 32, 3, 32, 1, 'MAX', 'incept3a', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept3a'] = net
net = network.inception(net, 256, 1, 64, 96, 128, 32, 64, 3, 64, 1, 'MAX', 'incept3b', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept3b'] = net
net = network.inception(net, 320, 2, 0, 128, 256, 32, 64, 3, 0, 2, 'MAX', 'incept3c', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept3c'] = net
net = network.inception(net, 640, 1, 256, 96, 192, 32, 64, 3, 128, 1, 'MAX', 'incept4a', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept4a'] = net
net = network.inception(net, 640, 1, 224, 112, 224, 32, 64, 3, 128, 1, 'MAX', 'incept4b', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept4b'] = net
net = network.inception(net, 640, 1, 192, 128, 256, 32, 64, 3, 128, 1, 'MAX', 'incept4c', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept4c'] = net
net = network.inception(net, 640, 1, 160, 144, 288, 32, 64, 3, 128, 1, 'MAX', 'incept4d', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept4d'] = net
net = network.inception(net, 640, 2, 0, 160, 256, 64, 128, 3, 0, 2, 'MAX', 'incept4e', phase_train=phase_train, use_batch_norm=True)
endpoints['incept4e'] = net
net = network.inception(net, 1024, 1, 384, 192, 384, 0, 0, 3, 128, 1, 'MAX', 'incept5a', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept5a'] = net
net = network.inception(net, 896, 1, 384, 192, 384, 0, 0, 3, 128, 1, 'MAX', 'incept5b', phase_train=phase_train, use_batch_norm=True, weight_decay=weight_decay)
endpoints['incept5b'] = net
net = network.apool(net, 3, 3, 1, 1, 'VALID', 'pool6')
endpoints['pool6'] = net
net = tf.reshape(net, [-1, 896])
endpoints['prelogits'] = net
net = tf.nn.dropout(net, keep_probability)
endpoints['dropout'] = net
return net, endpoints
def cnn_text(net_input, sequence_length, vocab_size, embedding_size, stride_h, filter_sizes, num_filters,
num_classes, keep_prob, l2_lambda, is_bn, bn_training):
embedding_layer = embedding(net_input, vocab_size, embedding_size, "embedding")
pools = list()
for i, filter_size in enumerate(filter_sizes):
convi = conv(embedding_layer, filter_size, embedding_size, 1, num_filters, stride_h, 1, "conv" + str(i))
if is_bn:
bni = batch_norm(convi, bn_training, "bn" + str(i))
convi = bni
pooli = max_pool(convi, (sequence_length - filter_size) // stride_h + 1, 1, 1, 1, "pool" + str(i))
pools.append(pooli)
num_filters_total = num_filters * len(filter_sizes)
h_pool = concat(pools, 3, "concat")
h_pool_flat = tf.reshape(h_pool, [-1, num_filters_total])
h_drop = dropout(h_pool_flat, keep_prob)
h_fc = fc(h_drop, num_filters_total, num_classes, l2_lambda, "fc")
return h_fc
def __init__(self):
super(Discriminator, self).__init__()
options = {
'leaky': True,
'bn': True,
'wn': False,
'pixel': False,
'gdrop': True
}
self.from_feature = layer.linear(feature_size,
64 * 64,
leaky=options['leaky'],
wn=options['wn'])
self.feature_conv = layer.conv(1, 64, 4, 2, 1, **options)
# 64 x 64
self.from_rgb = layer.conv(nc,
64,
4,
2,
1,
leaky=True,
bn=False,
gdrop=True)
# 32 x 32
self.conv1 = layer.conv(128, 128, 4, 2, 1, **options)
# 16 x 16
self.conv2 = layer.conv(128, 256, 4, 2, 1, **options)
# 8 x 8
self.conv3 = layer.conv(256, 512, 4, 2, 1, **options)
# 4 x 4
self.conv4 = nn.Sequential(
minibatch_std_concat_layer(),
layer.conv(513, 1, 4, 1, 0, gdrop=options['gdrop'], only=True))
# 1 x 1
# self.linear = layer.linear(512, 1, sig=False, wn=options['wn'])
self.convs = [self.conv1, self.conv2, self.conv3, self.conv4]
def darknet_block(filters,n):
nn.conv(1,filters*2,3,2,1,'leaky')
for i in range(1,n+1):
darknet_unit(filters)
def darknet_unit(filters):
nn.conv(1,filters,1,1,1,'leaky')
nn.conv(1,filters*2,3,1,1,'leaky')
nn.shortcut(-3,'linear')
def Fuse(filters,layer_id):
nn.route(-4)
nn.conv(1,filters,1,1,1,'leaky')
#nn.upsample(2)
nn.deconv('null',filters,2,2,0,'linear')
nn.route('-1, '+str(layer_id))
def detection_moudle(filters):
nn.conv(1,filters,1,1,1,'leaky')
nn.conv(1,filters*2,3,1,1,'leaky')
nn.conv(1,filters,1,1,1,'leaky')
nn.conv(1,filters*2,3,1,1,'leaky')
nn.conv(1,filters,1,1,1,'leaky')
nn.conv(1,filters*2,3,1,1,'leaky')
nn.conv('null',config.num/config.layer_num*(config.num_class+5),1,1,1,'linear')
return config.net_id