def CNNLight(X, Training, Scope):
with tf.variable_scope(Scope):
ConvLayer1 = ConvLayer(X, 1, 64, Training, 'ConvLayer1')
MPool1 = max_pool(ConvLayer1, ksize=(2, 2), stride=(2, 2))
ConvLayer2 = ConvLayer(MPool1, 64, 128, Training, 'ConvLayer2')
MPool2 = max_pool(ConvLayer2, ksize=(2, 2), stride=(2, 2))
ConvLayer3 = ConvLayer(MPool2, 128, 256, Training, 'ConvLayer3')
ConvLayer4 = ConvLayer(ConvLayer3, 256, 256, Training, 'ConvLayer4')
MPool4 = max_pool(ConvLayer4, ksize=(2, 1), stride=(2, 1))
ConvLayer5 = ConvLayer(MPool4, 256, 512, Training, 'ConvLayer5')
ConvLayer6 = ConvLayer(ConvLayer5, 512, 512, Training, 'ConvLayer6')
MPool6 = max_pool(ConvLayer6, ksize=(2, 1), stride=(2, 1))
ConvLayer7 = ConvLayer(MPool6, 512, 512, Training, 'ConvLayer7')
MPool7 = max_pool(ConvLayer7, ksize=(2, 1), stride=(2, 1))
MPool7_T = tf.transpose(MPool7, perm=[0, 2, 1, 3])
MPool7_T_RSH = tf.reshape(MPool7_T, [-1, FV, LastFilters])
return tf.reshape(MPool7_T_RSH, [-1, NFeatures])
def net_real(self, x, y, training=True, get_activ=False):
##
## mean subtract the image inputs
use_eval_mean = (not training and self.eval_mean)
mean_to_subtract = self.dataset_synth.val_image_channel_mean if use_eval_mean else self.dataset_real.train_image_channel_mean
x = tf.subtract(x, mean_to_subtract)
# VGG16 takes bgr format
conv1_1 = conv2d_relu2(x[..., ::-1], name='conv1_1', training=False)
conv1_2 = conv2d_relu2(conv1_1, name='conv1_2', training=False)
pool1 = max_pool(conv1_2, 'pool1')
##
conv2_1 = conv2d_relu2(pool1, name='conv2_1', training=False)
conv2_2 = conv2d_relu2(conv2_1, name='conv2_2', training=False)
##
pool2 = max_pool(conv2_2, 'pool2')
##
conv3_1 = conv2d_relu2(pool2, name='conv3_1', training=False)
conv3_2 = conv2d_relu2(conv3_1, name='conv3_2', training=False)
conv3_3 = conv2d_relu2(conv3_2, name='conv3_3', training=False)
pool3 = max_pool(conv3_3, 'pool3')
####
#if get_activ:
# return conv1_1, conv1_2, conv2_1, conv2_2, conv3_1, conv3_2, conv3_3
# ##
conv4_1 = conv2d_relu2(pool3, name='conv4_1', training=False)
conv4_2 = conv2d_relu2(conv4_1, name='conv4_2', training=False)
conv4_3 = conv2d_relu2(conv4_2, name='conv4_3', training=False)
#pool4 = max_pool(conv4_3, 'pool4')
###
if get_activ:
return conv1_1, conv1_2, conv2_1, conv2_2, conv3_1, conv3_2, conv3_3, conv4_1, conv4_2, conv4_3
def classifier(x):
#parameters for convolution kernels
IMG_DEPTH = 1
C1_KERNEL_SIZE, C2_KERNEL_SIZE, C3_KERNEL_SIZE = 5, 5, 5
C1_OUT_CHANNELS, C2_OUT_CHANNELS, C3_OUT_CHANNELS = 6, 16, 120
C1_STRIDES, C2_STRIDES, C3_STRIDES = 1, 1, 1
P1_SIZE, P2_SIZE = 2, 2
P1_STRIDE, P2_STRIDE = 2, 2
F4_SIZE, F5_SIZE = 84, 10
C1_kernel = util.weights(
[C1_KERNEL_SIZE, C1_KERNEL_SIZE, IMG_DEPTH, C1_OUT_CHANNELS], 0.1,
'C1_kernel')
C2_kernel = util.weights(
[C2_KERNEL_SIZE, C2_KERNEL_SIZE, C1_OUT_CHANNELS, C2_OUT_CHANNELS],
0.1, 'C2_kernel')
C3_kernel = util.weights(
[C3_KERNEL_SIZE, C3_KERNEL_SIZE, C2_OUT_CHANNELS, C3_OUT_CHANNELS],
0.1, 'C3_kernel')
C1_bias = util.bias([C1_OUT_CHANNELS], 'C1_bias')
C2_bias = util.bias([C2_OUT_CHANNELS], 'C2_bias')
C3_bias = util.bias([C3_OUT_CHANNELS], 'C3_bias')
#LeNet-5 structure
C1 = util.convLayer(x, C1_kernel, C1_STRIDES, 'SAME')
ReLU1 = tf.nn.relu(C1 + C1_bias)
P1 = util.max_pool(ReLU1, P1_SIZE, P1_STRIDE)
C2 = util.convLayer(P1, C2_kernel, C2_STRIDES, 'SAME')
ReLU2 = tf.nn.relu(C2 + C2_bias)
P2 = util.max_pool(ReLU2, P2_SIZE, P2_STRIDE)
C3 = util.convLayer(P2, C3_kernel, C3_STRIDES, 'SAME')
ReLU3 = tf.nn.relu(C3 + C3_bias)
num_F4_in = (int)(ReLU3.shape[1] * ReLU3.shape[2] * ReLU3.shape[3])
F4_in = tf.reshape(ReLU3, [-1, num_F4_in])
F4_weights = util.weights([num_F4_in, F4_SIZE], 0.1, 'F4_weights')
F4_bias = util.bias([F4_SIZE], 'F4_bias')
F4 = tf.matmul(F4_in, F4_weights)
ReLU4 = tf.nn.relu(F4 + F4_bias)
F5_weights = util.weights([F4_SIZE, F5_SIZE], 0.1, 'F5_weights')
F5_bias = util.bias([F5_SIZE], 'F5_bias')
F5 = tf.matmul(ReLU4, F5_weights) + F5_bias
return F5
def CNN(X, Training, Scope): with tf.variable_scope(Scope): ConvLayer1 = ConvLayer(X, 1, 64, Training, 'ConvLayer1') ConvLayer2 = ConvLayer(ConvLayer1, 64, 64, Training, 'ConvLayer2') MPool2 = max_pool(ConvLayer2, ksize=(2, 2), stride=(2, 2)) ConvLayer3 = ConvLayer(MPool2, 64, 128, Training, 'ConvLayer3') ConvLayer4 = ConvLayer(ConvLayer3, 128, 128, Training, 'ConvLayer4') MPool4 = max_pool(ConvLayer4, ksize=(2, 2), stride=(2, 2)) ConvLayer5 = ConvLayer(MPool4, 128, 256, Training, 'ConvLayer5') ConvLayer6 = ConvLayer(ConvLayer5, 256, 256, Training, 'ConvLayer6') ConvLayer7 = ConvLayer(ConvLayer6, 256, 256, Training, 'ConvLayer7') MPool7 = max_pool(ConvLayer7, ksize=(2, 1), stride=(2, 1)) ConvLayer8 = ConvLayer(MPool7, 256, 512, Training, 'ConvLayer8') ConvLayer9 = ConvLayer(ConvLayer8, 512, 512, Training, 'ConvLayer9') ConvLayer10 = ConvLayer(ConvLayer9, 512, 512, Training, 'ConvLayer10') MPool10 = max_pool(ConvLayer10, ksize=(2, 1), stride=(2, 1)) ConvLayer11 = ConvLayer(MPool10, 512, 512, Training, 'ConvLayer11') ConvLayer12 = ConvLayer(ConvLayer11, 512, 512, Training, 'ConvLayer12') ConvLayer13 = ConvLayer(ConvLayer12, 512, LastFilters, Training, 'ConvLayer13') MPool13 = max_pool(ConvLayer13, ksize=(2, 1), stride=(2, 1)) MPool13_T = tf.transpose(MPool13, perm=[0,2,1,3]) MPool13_T_RSH = tf.reshape(MPool13_T, [-1, FV, LastFilters]) cnn_res = tf.reshape(MPool13_T_RSH, [-1, NFeatures]) return cnn_res
def classifier(x):
#parameters for convolution kernels
IMG_DEPTH = 1
C1_KERNEL_SIZE, C2_KERNEL_SIZE, C3_KERNEL_SIZE, C4_KERNEL_SIZE, C5_KERNEL_SIZE = 11, 5, 3, 3, 3
C1_OUT_CHANNELS, C2_OUT_CHANNELS, C3_OUT_CHANNELS, C4_OUT_CHANNELS, C5_OUT_CHANNELS = 96, 256, 384, 384, 256
#C1_OUT_CHANNELS,C2_OUT_CHANNELS,C3_OUT_CHANNELS,C4_OUT_CHANNELS,C5_OUT_CHANNELS=48,128,197,197,128
C1_STRIDES, C2_STRIDES, C3_STRIDES, C4_STRIDES, C5_STRIDES = 1, 1, 1, 1, 1
P1_SIZE, P2_SIZE, P5_SIZE = 3, 3, 3
P1_STRIDE, P2_STRIDE, P5_STRIDE = 2, 2, 2
F6_SIZE, F7_SIZE = 4096, 4096
F8_SIZE = 10
#convolution kernels and bias
C1_kernel = util.weights(
[C1_KERNEL_SIZE, C1_KERNEL_SIZE, IMG_DEPTH, C1_OUT_CHANNELS], 0.01,
'C1_kernel')
C2_kernel = util.weights(
[C2_KERNEL_SIZE, C2_KERNEL_SIZE, C1_OUT_CHANNELS, C2_OUT_CHANNELS],
0.01, 'C2_kernel')
C3_kernel = util.weights(
[C3_KERNEL_SIZE, C3_KERNEL_SIZE, C2_OUT_CHANNELS, C3_OUT_CHANNELS],
0.01, 'C3_kernel')
C4_kernel = util.weights(
[C4_KERNEL_SIZE, C4_KERNEL_SIZE, C3_OUT_CHANNELS, C4_OUT_CHANNELS],
0.01, 'C4_kernel')
C5_kernel = util.weights(
[C5_KERNEL_SIZE, C5_KERNEL_SIZE, C4_OUT_CHANNELS, C5_OUT_CHANNELS],
0.01, 'C5_kernel')
C1_bias = util.bias([C1_OUT_CHANNELS], 'C1_bias')
C2_bias = util.bias([C2_OUT_CHANNELS], 'C2_bias')
C3_bias = util.bias([C3_OUT_CHANNELS], 'C3_bias')
C4_bias = util.bias([C4_OUT_CHANNELS], 'C4_bias')
C5_bias = util.bias([C5_OUT_CHANNELS], 'C5_bias')
#AlexNet network
#Conv layer 1
C1 = util.convLayer(x, C1_kernel, C1_STRIDES, 'SAME')
ReLU1 = tf.nn.relu(C1 + C1_bias)
P1 = util.max_pool(ReLU1, P1_SIZE, P1_STRIDE)
NORM1 = tf.nn.local_response_normalization(P1)
#Conv layer 2
C2 = util.convLayer(NORM1, C2_kernel, C2_STRIDES, 'SAME')
ReLU2 = tf.nn.relu(C2 + C2_bias)
P2 = util.max_pool(ReLU2, P2_SIZE, P2_STRIDE)
NORM2 = tf.nn.local_response_normalization(P2)
#Conv layer 3
C3 = util.convLayer(NORM2, C3_kernel, C3_STRIDES, 'SAME')
ReLU3 = tf.nn.relu(C3 + C3_bias)
#Conv layer 4
C4 = util.convLayer(ReLU3, C4_kernel, C4_STRIDES, 'SAME')
ReLU4 = tf.nn.relu(C4 + C4_bias)
#Conv layer 5
C5 = util.convLayer(ReLU4, C5_kernel, C5_STRIDES, 'SAME')
ReLU5 = tf.nn.relu(C5 + C5_bias)
P5_pre = util.max_pool(ReLU5, P5_SIZE, P5_STRIDE)
num_P5_out = (int)(P5_pre.shape[1] * P5_pre.shape[2] * P5_pre.shape[3])
P5 = tf.reshape(P5_pre, [-1, num_P5_out])
#Fully connected layer 6
F6_weights = util.weights([num_P5_out, F6_SIZE], 0.01, 'F6_weights')
F6_bias = util.bias([F6_SIZE], 'F6_bias')
F6 = tf.matmul(P5, F6_weights)
ReLU6 = tf.nn.relu(F6 + F6_bias)
DROP6 = tf.nn.dropout(ReLU6, 0.5)
#Fully connected layer 7
F7_weights = util.weights([F6_SIZE, F7_SIZE], 0.01, 'F7_weights')
F7_bias = util.bias([F7_SIZE], 'F7_bias')
F7 = tf.matmul(DROP6, F7_weights)
ReLU7 = tf.nn.relu(F7 + F7_bias)
DROP7 = tf.nn.dropout(ReLU7, 0.5)
#Fully connected layer 8
F8_weights = util.weights([F7_SIZE, F8_SIZE], 0.01, 'F8_weights')
F8_bias = util.bias([F8_SIZE], 'F8_bias')
logits = tf.matmul(DROP7, F8_weights) + F8_bias
return logits
file_writer = tf.summary.FileWriter('LOGS', sess.graph)
# Create placeholders for independent and dependant variables once batch has been selected.
with tf.name_scope('Input_Image'):
x = tf.placeholder(tf.float32, shape=[None, None, None, 1], name='Image') # Independent variables.
# Reshape to amenable shape.
# x_image = tf.reshape(x, [-1, windowSize[0], windowSize[1], 1])
with tf.name_scope('Input_Synapse'):
y_syn = tf.placeholder(tf.float32, shape=[None, 2]) # Target values.
with tf.name_scope('First_Layer'):
# Create first convolutional layer. (No pooling.)
W_conv1 = util.weight_variable(firstLayerDimensions, "w_conv_1") # Weights in first layer.
b_conv1 = util.bias_variable([firstLayerDimensions[3]], "b_conv_1") # Biases in first layer.
h_conv1 = tf.nn.relu(util.conv2d(x, W_conv1, valid=True, stride=1) + b_conv1) # Perform convolution (with zero padding) and apply ReLU.
h_pool1 = util.max_pool(h_conv1, 1) #, kernelWidth=2)
with tf.name_scope('Second_Layer'):
# Create first convolutional layer. (No pooling.)
W_conv2 = util.weight_variable(secondLayerDimensions, "w_conv_2") # Weights in first layer.
b_conv2 = util.bias_variable([secondLayerDimensions[3]], "b_conv_2") # Biases in first layer.
h_conv2 = tf.nn.relu(util.atrous_conv2d(h_pool1, W_conv2, valid=True, rate=2) + b_conv2) # Perform convolution (with zero padding) and apply ReLU.
h_pool2 = util.atrous_max_pool(h_conv2, mask_size=2, rate=2)
with tf.name_scope('Third_Layer'):
# Create first convolutional layer. (No pooling.)
W_conv3 = util.weight_variable(thirdLayerDimensions, "w_conv_3") # Weights in first layer.
b_conv3 = util.bias_variable([thirdLayerDimensions[3]], "b_conv_3") # Biases in first layer.
h_conv3 = tf.nn.relu(util.atrous_conv2d(h_pool2, W_conv3, valid=True, rate=4) + b_conv3) # Perform convolution (with zero padding) and apply ReLU.
h_pool3 = util.atrous_max_pool(h_conv3, mask_size=2, rate=4)