def res_block_bottleneck(x,
ksize_list,
strides,
is_training,
data_format='NHWC',
w_project=None,
no_activation=False):
"""
Computes y = x + F(x), where F(x) is the residual block function. At downsample layers, applies
a downsample function on x as well.
"""
if w_project is not None:
x_ = tf.conv2d(x,
w_project,
strides,
padding='SAME',
data_format=data_format)
else:
x_ = x
return x_ + _bottleneck_residual(x,
ksize_list,
strides,
'SAME',
is_training,
data_format=data_format,
no_activation=no_activation)
def tf_detect_glassframe(rgb_img, seg_img):
# input range [0,1]
dX = seg_img[:, :, 1:] - seg_img[:, :, :-1]
dX = tf.pad(dX, ((0, 0), (0, 0), (0, 1)), "constant")
dY = seg_img[:, 1:, :] - seg_img[:, :-1, :]
dY = tf.pad(dY, ((0, 0), (0, 1), (0, 0)), "constant")
G = tf.sqrt(tf.square(dX) + tf.square(dY))
G = tf.where(tf.greater(G, 0.1), tf.ones_like(G), tf.zeros_like(G))
G = tf.expand_dims(G, axis=3)
k = 10
kernel = np.ones((k, k), np.float32) / (k * k)
kernel = tf.reshape(kernel, [k, k, 1, 1])
G_list = tf.split(G, G_list.get_shape().as_list()[-1], axis=-1)
G_out_list = []
for g_channel in G_list:
g_out_channel = tf.conv2d(g_channel,
kernel,
strides=[1, 1, 1, 1],
padding='SAME')
G_out_list.append(g_out_channel)
mask = tf.concat(G_out_list, axis=-1)
mask = tf.where(tf.greater(mask, 0.01), tf.ones_like(mask),
tf.zeros_like(mask))
mask = tf.squeeze(mask)
# convert rgb to hsv
hsv_img = tf_rgb_to_hsv(rgb_img)
v_img = hsv_img[:, :, :, 2]
v_mask = tf.where(tf.less(v_img, 0.6), tf.ones_like(v_img),
tf.zeros_like(v_img))
glassframe = v_mask * mask * seg_img
return glassframe
def conv2d_op(x,name,shape,strides,pams): with tf.name_scope(name) as scope: W_conv=tf.get_variable(scope+'w',shape=shape,dtype=tf.float32, initializer=tf.contrib.layers.xavier_initializer()) b_conv=tf.Variable(tf.constant(0.0,shape=[shape[-1]],dtype=tf.float32),trainable=True,name='b') conv=tf.conv2d(x,W_conv,strides=strides,padding='SAME') h_conv=tf.nn.relu(tf.bias_add(conv,b_conv),name=scope) pams+=[W_conv,b_conv] return h_conv
def buildLayer(self, layers, layerDescription):
layerType = layerDescription["type"]
layerName = layerDescription["name"]
output = None
if layerType == "fully-connected":
inputs = self.getSimpleInputs(layers, layerDescription)
layerSize = int(layerDescription["size"])
output = tf.nn.fully_connected(inputs, layerSize)
else if layerType == "relu":
inputs = self.getSimpleInputs(layers, layerDescription)
output = tf.nn.relu(inputs)
else if layerType == "tanh":
inputs = self.getSimpleInputs(layers, layerDescription)
output = tf.nn.tanh(inputs)
else if layerType == "sigmoid":
inputs = self.getSimpleInputs(layers, layerDescription)
output = tf.nn.sigmoid(inputs)
else if layerType == "mul":
inputs = self.getBinaryInputs(layers, layerDescription)
output = tf.mul(inputs[0], inputs[1])
else if layerType == "add":
inputs = self.getBinaryInputs(layers, layerDescription)
output = tf.add(inputs[0], inputs[1])
else if layerType == "ones":
layerSize = int(layerDescription["size"])
output = tf.ones(layerSize)
else if layerType == "zeros":
layerSize = int(layerDescription["size"])
output = tf.zeros(layerSize)
else if layerType == "conv2d":
layerSize = int(layerDescription["size"])
layerStride = int(layerDescription["stride"])
layerFilterSize = int(layerDescription["filter"])
inputs, padding = self.getSimple2DInputsAndPadding(layers,
layerDescription, layerFilterSize)
self.fixInputsForConv2d(inputs)
filterWeights = self.makeConv2dFilterWeights(inputs, layerSize,
layerStride, layerFilterSize)
output = tf.conv2d(inputs, filterWeights, self.getConv2dStrides(layerStride), padding)
layer[layerName] = output
def neural_network():
with tf.device('/cpu:0'), tf.name_scope("embedding"):
embedding_size = 128
W = tf.Variable(
tf.random_uniform([input_size, embedding_size], -1.0, 1.0))
embedding_chars = tf.embedding_lookup(W, X)
embedded_chars_expanded = tf.expand_dims(embedding_chars, -1)
num_filter = 128
filter_size = [3, 4, 5]
pooled_output = []
for i, filter_size in enumerate(filter_sizes):
with tf.name_scope("conv_maxpool_%s" % filter_size):
filter_shape = [filter_size, embedding_size, 1, num_filter]
W = tf.Variable(tf.truncated_normal(filter_shape, stdev=0.1))
b = tf.Variable(tf.constant(0.1, shape=[num_filter]))
conv = tf.conv2d(embedded_chars_expanded,
W,
strides=[1, 1, 1, 1],
padding="VALID")
h = tf.nn.relu(tf.nn.bias_add(conv, b))
pooled = tf.nn.max_pool(h,
ksize=[1, input_size - file_size, 1, 1],
strides=[1, 1, 1, 1],
padding="VALID")
pooled_output.append(pooled)
num_filters_total = num_filters * len(filter_sizes)
h_pool = tf.concat(3, pooled_outputs)
h_pool_flat = tf.reshape(h_pool, [-1, num_filters_total])
# dropout
with tf.name_scope("dropout"):
h_drop = tf.nn.dropout(h_pool_flat, dropout_keep_prob)
# output
with tf.name_scope("output"):
W = tf.get_variable("W",
shape=[num_filters_total, num_classes],
initializer=tf.contrib.layers.xavier_initializer())
b = tf.Variable(tf.constant(0.1, shape=[num_classes]))
output = tf.nn.xw_plus_b(h_drop, W, b)
return output
def forward(self, input_image):
input_image = tf.reshape(input_image, [-1, 32, 32, 3])
activ = tf.nn.conv2d(input_image,
filters=self.convW1,
strides=[1, 4, 4, 1],
padding='VALID')
activ = tf.nn.conv2d(activ,
filters=self.convW2,
strides=[1, 2, 2, 1],
padding='VALID')
activ = tf.nn.conv2d(activ,
filters=self.convW3,
strides=[1, 2, 2, 1],
padding='VALID')
activ = tf.conv2d(activ,
filters=self.conW4,
strides=[1, 1, 1, 1],
padding='VALID')
activ = tf.slim.flatten(activ)
activ = tf.matmul(activ, self.W1) + self.b1
activ = tf.matmul(activ, self.W2) + self.b2
return activ
def sparse_res_block_bottleneck(x,
ksize_list,
indices,
block_params,
strides,
is_training,
data_format='NHWC',
w_project=None,
no_activation=False,
use_var=False):
"""
Computes y = x + F(x), where F(x) is the residual block function. At downsample layers, applies
a downsample function on x as well.
:param x: [Tensor] [N, H, W, C]. Input activation tensor, dtype float32.
:param ksize_list: [list] List of list of 4 int. Kernel size for each convolution
layer in the residual block.
:param indices: [tuple] Non-sparse locations returned by reduce_mask.
:param block_params: [tuple] BlockParam namedtuple.
:param
:return
"""
transpose = True if data_format == 'NCHW' else False
p = sbnet_module.sparse_gather(x,
indices.bin_counts,
indices.active_block_indices,
bsize=block_params.bsize,
boffset=block_params.boffset,
bstride=block_params.bstrides,
transpose=transpose)
if w_project is not None:
x = tf.conv2d(x, w_project, strides, padding='SAME')
# Set shape for BN in the residual function.
if transpose:
p.set_shape([
None,
x.get_shape()[3], block_params.bsize[0], block_params.bsize[1]
])
else:
p.set_shape([
None, block_params.bsize[0], block_params.bsize[1],
x.get_shape()[3]
])
q = _bottleneck_residual(p,
ksize_list,
strides,
'VALID',
is_training,
data_format=data_format,
no_activation=no_activation)
if use_var:
y = sbnet_module.sparse_scatter_var(q,
indices.bin_counts,
indices.active_block_indices,
x,
bsize=block_params.bsize_out,
boffset=[0, 0],
bstride=block_params.bsize_out,
add=True,
transpose=transpose)
else:
y = sbnet_module.sparse_scatter(q,
indices.bin_counts,
indices.active_block_indices,
x,
bsize=block_params.bsize_out,
boffset=[0, 0],
bstride=block_params.bsize_out,
add=True,
transpose=transpose)
return y
def conv2d(x, W, strides, padding='SAME'):
return tf.conv2d(x, W, strides=strides, padding=padding)
def inference():
# zero mean input
mean = tf.constant(VGG_MEAN, dtype=tf.float32, shape=[1,1,1,3], name='img_mean')
imgs = self.imgs - mean
# color transform
with tf.name_scope('color_space') as scope:
kernel = tf.Variable(tf.truncated_normal([1,1,1,3]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[3],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv = tf.nn.relu(out, name=scope)
with tf.name_scope('conv1_1') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,3,64]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[64],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv1_1 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv1_2') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,64,64]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[64],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv1_2 = tf.nn.relu(out, name=scope)
with tf.name_scope('pool_1') as scope:
pool_1 = tf.nn.max_pool(conv1_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool1')
with tf.name_scope('conv2_1') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,64,128]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[128],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv2_1 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv2_2') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,128,128]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[128],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv2_2 = tf.nn.relu(out, name=scope)
with tf.name_scope('pool_2') as scope:
pool_2 = tf.nn.max_pool(conv2_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool2')
with tf.name_scope('conv3_1') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,128,256]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv2_1 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv3_2') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,256,256]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv3_2 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv3_3') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,256,256]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[256],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv3_3 = tf.nn.relu(out, name=scope)
with tf.name_scope('pool_3') as scope:
pool_4 = tf.nn.max_pool(conv3_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool3')
with tf.name_scope('conv4_1') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,256,512]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[512],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv2_1 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv4_2') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,512,512]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[512],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv2_2 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv4_3') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,512,512]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[512],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv2_1 = tf.nn.relu(out, name=scope)
with tf.name_scope('pool_4') as scope:
pool_4 = tf.nn.max_pool(conv4_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool4')
with tf.name_scope('conv5_1') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,512,512]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[512],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv5_1 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv5_2') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,512,512]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[512],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv5_2 = tf.nn.relu(out, name=scope)
with tf.name_scope('conv5_3') as scope:
kernel = tf.Variable(tf.truncated_normal([3,3,512,512]), dtype=tf.float32, stddev=1e-1, name='weights')
conv = tf.conv2d(imgs, kernel, [1,1,1,1], padding='SAME')
biases = tf.Variable(tf.constant(0.0,shape=[512],dtype=tf.float32), trainable=True,name='biases')
out = tf.nn.bias_add(conv, biases)
conv5_3 = tf.nn.relu(out, name=scope)
with tf.name_scope('pool_5') as scope:
pool_5 = tf.nn.max_pool(conv2_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1],
padding='SAME', name='pool5')
return pool5
import tensorflow as tf
import cv2
# Create a 3x3 Gabor filter
params = {
'ksize': (3, 3),
'sigma': 1.0,
'theta': 0,
'lambd': 15.0,
'gamma': 0.02
}
filter = cv2.getGaborKernel(**params)
# make the filter to have 4 dimensions.
filter = tf.expand_dims(filter, 2)
filter = tf.expand_dims(filter, 3)
# Apply the filter on `image`
answer = tf.conv2d(image, filter, strides=[1, 1, 1, 1], padding='SAME')
def convLayer(self, input, x):
# conv and 최적화
featureMap = tf.conv2d(input (x, x))
return featureMap