def reductionA(name, x):
with tf.variable_scope(name) as scope:
layer_ = max_pool('max_pool_0', x, k=3, s=2, padding='VALID')
layer__ = convolution2d('cnn__0', x, 192, k=3, s=2, padding='VALID')
layer___ = convolution2d('cnn___0', x, 224, k=1, s=1, padding='SAME')
layer___ = convolution2d('cnn___1',
layer___,
256,
k=3,
s=1,
padding='SAME')
layer___ = convolution2d('cnn___2',
layer___,
385,
k=3,
s=2,
padding='VALID')
layer_join = tf.concat([layer_, layer__, layer___],
axis=3,
name='join')
print 'layer_name :', 'join'
print 'layer_shape :', layer_join.get_shape()
return layer_join
def stem(name , x):
with tf.variable_scope(name) as scope:
layer=convolution2d('cnn_0',x,32,k=3,s=2 , padding='VALID')
layer = convolution2d('cnn_1',layer, 32, k = 3, s = 1, padding = 'VALID')
layer = convolution2d('cnn_2', layer, 64, k=3, s=1, padding='SAME')
layer_1 = max_pool('max_3', layer, k=3, s=2, padding='VALID')
layer_2 = convolution2d('cnn_3_1', layer, 96, k=3, s=2, padding='VALID')
layer_join=tf.concat([layer_1,layer_2] , axis=3 , name='join')
print 'layer_name :','join'
print 'layer_shape :',layer_join.get_shape()
return layer_join
def blockC(name, x):
with tf.variable_scope(name) as scope:
layer = avg_pool('avg_pool', x, k=2, s=1)
layer = convolution2d('cnn', layer, 256, k=1, s=1)
layer_ = convolution2d('cnn_0', x, 256, k=1, s=1)
layer__ = convolution2d('cnn__0', x, 384, k=1, s=1)
layer__0 = convolution2d_manual('cnn__1_0',
layer__,
256,
k_h=1,
k_w=3,
s=1)
layer__1 = convolution2d_manual('cnn__1_1',
layer__,
256,
k_h=3,
k_w=1,
s=1)
layer___ = convolution2d('cnn___0', x, 384, k=1, s=1)
layer___ = convolution2d_manual('cnn___1',
layer___,
448,
k_h=1,
k_w=3,
s=1)
layer___ = convolution2d_manual('cnn___2',
layer___,
512,
k_h=3,
k_w=1,
s=1)
layer___0 = convolution2d_manual('cnn___3_0',
layer___,
256,
k_h=3,
k_w=1,
s=1)
layer___1 = convolution2d_manual('cnn___3_1',
layer___,
256,
k_h=1,
k_w=3,
s=1)
layer_join = tf.concat(
[layer, layer_, layer__0, layer__1, layer___0, layer___1],
axis=3,
name='join')
print 'layer_name :', 'join'
print 'layer_shape :', layer_join.get_shape()
return layer_join
def blockB(name, x):
with tf.variable_scope(name) as scope:
layer = avg_pool('avg_pool', x, k=2, s=1)
layer = convolution2d('cnn', layer, 128, k=1, s=1)
layer_ = convolution2d('cnn_0', x, 384, k=1, s=1)
layer__ = convolution2d('cnn__0', x, 192, k=1, s=1)
layer__ = convolution2d_manual('cnn__1',
layer__,
224,
k_h=1,
k_w=7,
s=1)
layer__ = convolution2d_manual('cnn__2',
layer__,
256,
k_h=1,
k_w=7,
s=1)
layer___ = convolution2d('cnn___0', x, 192, k=1, s=1)
layer___ = convolution2d_manual('cnn___1',
layer___,
192,
k_h=1,
k_w=7,
s=1)
layer___ = convolution2d_manual('cnn___2',
layer___,
224,
k_h=7,
k_w=1,
s=1)
layer___ = convolution2d_manual('cnn___3',
layer___,
224,
k_h=1,
k_w=7,
s=1)
layer___ = convolution2d_manual('cnn___4',
layer___,
256,
k_h=7,
k_w=1,
s=1)
layer_join = tf.concat([layer, layer_, layer__, layer___],
axis=3,
name='join')
print 'layer_name :', 'join'
print 'layer_shape :', layer_join.get_shape()
return layer_join
def stem_1(name , x ):
with tf.variable_scope(name) as scope:
layer = convolution2d('cnn_0', x, 64, k=1, s=1)
layer = convolution2d('cnn_1', layer, 96, k=3, s=1, padding='VALID')
layer_ = convolution2d('cnn__0', x, 64, k=1, s=1)
layer_ = convolution2d_manual('cnn__1', layer_, 64, k_h=7,k_w=1, s=1)
layer_ = convolution2d_manual('cnn__2', layer_, 64, k_h=1,k_w=7,s=1 )
layer_ = convolution2d('cnn__3', layer_, 96, k=3, s=1, padding='VALID')
layer_join = tf.concat([layer, layer_], axis=3, name='join')
print 'layer_name :','join'
print 'layer_shape :',layer_join.get_shape()
return layer_join
def resnet_blockC(name, x):
with tf.variable_scope(name) as scope:
layer = convolution2d('cnn0', x, 192, 1, 1)
layer_ = convolution2d('cnn_0', x, 192, 1, 1)
layer_ = convolution2d_manual('cnn_1', layer_, 192, k_h=1, k_w=3, s=1)
layer_ = convolution2d_manual('cnn_2', layer_, 192, k_h=3, k_w=1, s=1)
layer_join = tf.concat([layer, layer_], axis=3, name='join')
layer_join = convolution2d('layer_join_cnn', layer_join, 1792, 1, 1)
if x.get_shape()[-1] != layer_join.get_shape()[-1]:
x=convolution2d('upscale_dimension',x, layer_join.get_shape()[-1] , k=1,s=1)
layer_join = tf.add(x, layer_join, 'add')
print 'layer_name :', 'join'
print 'layer_shape :', layer_join.get_shape()
return layer_join
def stem_2(name ,x ):
with tf.variable_scope(name) as scope:
layer= convolution2d('cnn_0' , x, 192,k=3,s=2,padding='VALID')
layer_=max_pool('max__0' , x, k=3 , s=2 , padding = 'VALID')
layer_join = tf.concat([layer , layer_] , axis = 3 ,name='join')
print 'layer_name :','join'
print 'layer_shape :',layer_join.get_shape()
return layer_join
def rpn_bbox_layer(layer, n_anchors=9):
with tf.variable_scope('bbox'):
layer = convolution2d('rpn_bbox_conv',
layer,
out_ch=n_anchors * 4,
k=1,
act=None,
s=1)
layer = tf.identity(layer, name='cls_output')
print '** cls layer shape : {}'.format(np.shape(layer)) #(1, ?, ?, 18)
return layer
def simple_convnet(x_):
print '###### Convolution Network building.... '
print
kernels=[5, 3, 3, 3, 3]
out_channels=[16, 16, 32, 64, 128]
strides = [2, 2, 1, 2 ,1 ]
layer=x_
for i in range(5):
layer = convolution2d(name='conv_{}'.format(i), x=layer, out_ch=out_channels[i], k=kernels[i], s=strides[i],
padding='SAME')
top_conv = tf.identity(layer , 'top_conv')
_feat_stride = np.prod(strides)
return top_conv , _feat_stride
def simple_convnet(self, out_chs, kernels, strides, n_classes):
layer = apply_aug_lv0(self.x_,
aug_lv0,
self.phase_train,
crop_h=self.img_h,
crop_w=self.img_w)
layer = layer
assert len(out_chs) == len(kernels) == len(strides)
for i in range(len(out_chs)):
layer = convolution2d('conv{}'.format(i), layer, out_chs[i],
kernels[i], strides[i])
top_conv = tf.identity(layer, 'top_conv')
self.logits = gap('gap', top_conv, n_classes)
def resnet_blockB(name , x):
with tf.variable_scope(name) as scope:
layer = convolution2d('cnn0', x, 32, 1, 1)
layer_ = convolution2d('cnn_0', x, 32, 1, 1)
layer_ = convolution2d('cnn_1', layer_, 32, 3, 1)
layer__ = convolution2d('cnn__0', x, 32, 1, 1)
layer__ = convolution2d('cnn__1', layer__, 32, 3, 1)
layer__ = convolution2d('cnn__2', layer__, 32, 3, 1)
layer_join=tf.concat([layer , layer_ , layer__] , axis=3 , name='join')
layer_join=convolution2d('layer_join_cnn' , layer_join ,256, 1,1 )
if x.get_shape()[-1] != layer_join.get_shape()[-1]:
x=convolution2d('upscale_dimension',x, layer_join.get_shape()[-1] , k=1,s=1)
layer_join=tf.add(x,layer_join )
return layer_join
def reductionC(name ,x ):
with tf.variable_scope(name) as scope:
layer=max_pool('max_pool0' , x ,3,2 , padding='VALID')
layer_=convolution2d('cnn_0',x , 256,1,1)
layer_=convolution2d('cnn_1',layer_, 384 , 3,1 , padding='VALID')
layer__ = convolution2d('cnn_0', x, 256, 1, 1)
layer__ = convolution2d('cnn_1', layer__, 256, 3, 1 , padding='VALID')
layer___= convolution2d('cnn__0',x, 256 , 1,1)
layer___= convolution2d('cnn__1',layer___,256 , 3,1)
layer___= convolution2d('cnn__2',layer___,256 , 3,1 , padding='VALID')
layer_join= tf.concat([layer , layer_ , layer__ , layer___], axis=3 , name='join' )
return layer_join
def blockA(name , x):
with tf.variable_scope(name) as scope:
layer = avg_pool('avg_pool', x, k=2, s=1)
layer = convolution2d('cnn', layer, 96, k=1, s=1)
layer_ = convolution2d('cnn_0', x, 96, k=1, s=1)
layer__ = convolution2d('cnn__0', x, 64, k=1, s=1)
layer__ = convolution2d('cnn__1', layer__,96, k=3, s=1)
layer___ = convolution2d('cnn___0', x,64, k=1, s=1)
layer___ = convolution2d('cnn___1',layer___,96, k=3, s=1)
layer___ = convolution2d('cnn___2',layer___,96, k=3, s=1)
layer_join = tf.concat([layer, layer_, layer__, layer___], axis=3, name='join')
print 'layer_name :', 'join'
print 'layer_shape :', layer_join.get_shape()
return layer_join
import numpy as np
import utils
from inception_v4 import stem, stem_1, stem_2, reductionB, reductionA, blockA, blockB, blockC, resnet_blockA, resnet_blockB, resnet_blockC
##########################setting############################
image_height, image_width, image_color_ch, n_classes, train_imgs, train_labs, test_imgs, test_labs = data.mnist_28x28(
)
x_ = tf.placeholder(dtype=tf.float32,
shape=[None, image_height, image_width, image_color_ch],
name='x_')
y_ = tf.placeholder(dtype=tf.int32, shape=[None, n_classes], name='y_')
phase_train = tf.placeholder(dtype=tf.bool, name='phase_train')
batch_size = 60
##########################structure##########################
layer, conv1_summary_tensor = convolution2d('conv1', x_, 64)
#layer = batch_norm_1(layer, phase_train, 'bn0')
layer = batch_norm_2(layer, phase_train, 'bn0') # best learning rate ==> 0.1
layer = max_pool('max_pool1', layer)
layer, topconv_summary_tensor = convolution2d('top_conv', layer, 128)
layer = batch_norm_2(layer, phase_train,
'bn1') # best learning rate ==> 0.01 batch norm이 많아지면
layer, fc_summary_tensor = affine('fully_connect',
layer,
1024,
keep_prob=0.5,
phase_train=phase_train)
layer = batch_norm_1(layer, phase_train, 'bn2')
y_conv = logits('end_layer', layer, n_classes)
merged = tf.summary.merge_all()
out_chs = [32, 32, 64, 64, 64, 128, 128, 128, 256, 256, 256, 512, 512]
filters = [5, 3, 5, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3]
strides = [2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
#for Mnist version
pool_indices = [1, 4, 7, 10]
out_chs = [32, 32, 64, 64, 64]
filters = [5, 3, 5, 3, 3]
strides = [2, 1, 2, 1, 1] #,1,1,1,1,1,1,1,1]
#Building Network
layer = x_
assert len(out_chs) == len(filters)
n_layers = len(out_chs)
for i in range(n_layers):
layer = convolution2d('conv_{}'.format(i), layer, out_chs[i], s=1)
if i in pool_indices:
'max pool'
layer = max_pool('maxPool_{}'.format(i), layer)
top_conv = tf.identity(layer, 'top_conv')
logits = ram('ram', top_conv)
#logits=ram('ram' , top_conv)
# Build Optimizer
pred,pred_cls , cost , train_op,correct_pred ,accuracy = \
algorithm(y_conv=logits , y_=y_ ,learning_rate=lr , optimizer='sgd' , use_l2_loss=True ,activation='sigmoid' , cost_func='mse')
sess = tf.Session()
init = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
