def define_inner_modules(self, name, is_training, activations, conv_filter_shapes, bias_shapes, ksizes, pool_strides, topdown_filter_shapes, topdown_output_shapes, keep_prob, FLAGS):
# create all modules of the network
# -----
self.layers = {}
with tf.name_scope('input_normalization'):
self.layers["inp_norm"] = m.NormalizationModule("inp_norm")
with tf.name_scope('convolutional_layer_0'):
#self.layers["conv0"] = m.TimeConvolutionalLayerWithBatchNormalizationModule("conv0", bias_shapes[0][-1], is_training, 0.0, 1.0, 0.5, activations[0], [5,5,1,32], [1,1,1,1], [1,28,28,32])
self.layers["conv0"] = m.TimeConvolutionalLayerModule("conv0", activations[0], [3,3,1,32], [1,1,1,1], [1,28,28,32])
with tf.name_scope('pooling_layer_0'):
self.layers["pool0"] = m.MaxPoolingModule("pool0", ksizes[0], pool_strides[0])
with tf.name_scope('convolutional_layer_1'):
#self.layers["conv1"] = m.TimeConvolutionalLayerWithBatchNormalizationModule("conv1", bias_shapes[1][-1], is_training, 0.0, 1.0, 0.5, activations[1], [5,5,32,64], [1,1,1,1], [1,28,28,64])
self.layers["conv1"] = m.TimeConvolutionalLayerModule("conv1", activations[1], [5,5,32,64], [1,1,1,1], [1,14,14,64])
with tf.name_scope('pooling_layer_1'):
self.layers["pool1"] = m.MaxPoolingModule("pool1", ksizes[0], pool_strides[1])
with tf.name_scope('global_average_pooling'):
self.layers['gap'] = m.GapModule('gap')
with tf.name_scope('fully_connected_layer_0'):
self.layers["fc0"] = m.FullyConnectedModule("fc0", bias_shapes[1][-1], np.prod(bias_shapes[2]))
# connect all modules of the network in a meaningful way
# -----
with tf.name_scope('wiring_of_modules'):
self.layers["conv0"].add_input(self.layers["inp_norm"], 0)
self.layers["pool0"].add_input(self.layers["conv0"])
self.layers["conv1"].add_input(self.layers["pool0"], 0)
self.layers["pool1"].add_input(self.layers["conv1"])
self.layers["gap"].add_input(self.layers["pool1"])
self.layers["fc0"].add_input(self.layers["gap"])
with tf.name_scope('input_output'):
self.input_module = self.layers["inp_norm"]
self.output_module = self.layers["fc0"]
def define_inner_modules(self, name, is_training, trainable_input, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS, keep_prob, FLAGS):
# book-keeping of dimensions
self.shapes = {}
# create all modules of the network
# -----
self.layers = {}
with tf.name_scope('input_normalization'):
self.layers["inp_norm"] = m.NormalizationModule("inp_norm")
with tf.name_scope('convolutional_layer_0'):
self.layers["conv0"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv0", \
32, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [4,4,IMAGE_CHANNELS,32], [1,2,2,1], [1, IMAGE_HEIGHT//2, IMAGE_WIDTH//2, 32])
with tf.name_scope('convolutional_layer_1'):
self.layers["conv1"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv1", \
16, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [2,2, 32, 16], [1,2,2,1], [1, IMAGE_HEIGHT//4, IMAGE_WIDTH//4, 16])
with tf.name_scope('convolutional_layer_2'):
self.layers["conv2"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv2", \
4, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [1, 1, 16, 4], [1,1,1,1], [1, IMAGE_HEIGHT//4, IMAGE_HEIGHT//4, 4])
with tf.name_scope('trainable_input_canvas'):
self.layers['bottleneck_switch'] = m.SwitchModule('bottleneck_switch', trainable_input)
self.layers['bottleneck_canvas'] = m.BiasModule('bottleneck_canvas', [FLAGS.batchsize,IMAGE_HEIGHT//4, IMAGE_HEIGHT//4, 4])
with tf.name_scope('convolutional_layer_3'):
self.layers["conv3"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv3", \
16, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [1, 1, 4, 16], [1,1,1,1], [1, IMAGE_HEIGHT//4, IMAGE_HEIGHT//4, 16])
with tf.name_scope('convolutional_layer_4'):
self.layers["conv4"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv4", \
32, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [1, 1, 16, 32], [1,1,1,1], [1, IMAGE_HEIGHT//4, IMAGE_HEIGHT//4, 32])
with tf.name_scope('deconvolutional_layer_0'):
self.layers["deconv0"] = m.Conv2DTransposeModule("deconv0", \
[4,4,IMAGE_CHANNELS,32], [1,4,4,1], [FLAGS.batchsize,IMAGE_HEIGHT,IMAGE_WIDTH,IMAGE_CHANNELS])
self.layers["deconv0_act"] = m.ActivationModule("deconv0_act", \
lrn_relu)
# connect all modules of the network in a meaningful way
# -----
with tf.name_scope('wiring_of_modules'):
self.layers["conv0"].add_input(self.layers["inp_norm"])
self.layers["conv1"].add_input(self.layers["conv0"])
self.layers["conv2"].add_input(self.layers["conv1"])
self.layers["bottleneck_switch"].add_input(self.layers["conv2"])
self.layers["bottleneck_switch"].add_input(self.layers["bottleneck_canvas"])
self.layers["conv3"].add_input(self.layers["bottleneck_switch"])
self.layers["conv4"].add_input(self.layers["conv3"])
self.layers["deconv0"].add_input(self.layers["conv4"])
self.layers["deconv0_act"].add_input(self.layers["deconv0"])
with tf.name_scope('input_output'):
self.input_module = self.layers["inp_norm"]
self.output_module = self.layers["deconv0_act"]
def define_inner_modules(self, name, is_training, trainable_input, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS, keep_prob, FLAGS):
# book-keeping of dimensions
self.shapes = {}
# create all modules of the network
# -----
self.layers = {}
with tf.name_scope('input_normalization'):
self.layers["inp_norm"] = m.NormalizationModule("inp_norm")
with tf.name_scope('convolutional_layer_0'):
self.layers["conv0"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv0", \
32, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [3,3,IMAGE_CHANNELS,32], [1,2,2,1], [1, IMAGE_HEIGHT//2, IMAGE_WIDTH//2, 32])
# self.layers["conv0"] = m.ConvolutionalLayerModule("conv0", \
# lrn_relu, [3,3,IMAGE_CHANNELS,32], [1,2,2,1], [1, IMAGE_HEIGHT//2, IMAGE_WIDTH//2, 32])
# with tf.name_scope('pooling_layer_0'):
# self.layers["pool0"] = m.MaxPoolingModule("pool0", [1,2,2,1], [1,2,2,1])
with tf.name_scope('convolutional_layer_1'):
self.layers["conv1"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv1", \
16, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [3, 3, 32, 16], [1,2,2,1], [1, IMAGE_HEIGHT//4, IMAGE_WIDTH//4, 16])
# self.layers["conv1"] = m.ConvolutionalLayerModule("conv1", \
# lrn_relu, [3, 3, 32, 16], [1,2,2,1], [1, IMAGE_HEIGHT//4, IMAGE_WIDTH//4, 16])
# with tf.name_scope('pooling_layer_1'):
# self.layers["pool1"] = m.MaxPoolingModule("pool1", [1,2,2,1], [1,2,2,1])
with tf.name_scope('convolutional_layer_2'):
self.layers["conv2"] = m.ConvolutionalLayerWithBatchNormalizationModule("conv2", \
4, is_training, 0.0, 1.0, 0.5, \
lrn_relu, [3, 3, 16, 4], [1,2,2,1], [1, np.ceil(IMAGE_HEIGHT/8), np.ceil(IMAGE_WIDTH/8), 4])
# self.layers["conv2"] = m.ConvolutionalLayerModule("conv2", \
# lrn_relu, [3, 3, 16, 8], [1,2,2,1], [1, np.ceil(IMAGE_HEIGHT/8), np.ceil(IMAGE_WIDTH/8), 8])
# with tf.name_scope('pooling_layer_2'):
# self.layers["pool2"] = m.MaxPoolingModule("pool2", [1,2,2,1], [1,2,2,1])
with tf.name_scope('trainable_input_canvas'):
self.layers['bottleneck_switch'] = m.SwitchModule('bottleneck_switch', trainable_input)
self.layers['bottleneck_canvas'] = m.BiasModule('bottleneck_canvas', [FLAGS.batchsize,np.ceil(IMAGE_HEIGHT/8), np.ceil(IMAGE_WIDTH/8), 4])
with tf.name_scope('deconvolutional_layer_0'):
self.layers["deconv0"] = m.Conv2DTransposeModule("deconv0", \
[3,3,16,4], [1,2,2,1], [FLAGS.batchsize,IMAGE_HEIGHT//4,IMAGE_WIDTH//4,16])
# self.layers["deconv0_bn"] = m.BatchNormalizationModule("deconv0_batchnorm", \
# 8, is_training, beta_init=0.0, gamma_init=0.1, ema_decay_rate=0.5, moment_axes=[0,1,2], variance_epsilon=1e-3)
self.layers["deconv0_act"] = m.ActivationModule("deconv0_act", \
lrn_relu)
with tf.name_scope('deconvolutional_layer_1'):
self.layers["deconv1"] = m.Conv2DTransposeModule("deconv1", \
[3,3,32,16], [1,2,2,1], [FLAGS.batchsize,IMAGE_HEIGHT//2,IMAGE_WIDTH//2,32])
# self.layers["deconv0_bn"] = m.BatchNormalizationModule("deconv1_batchnorm", \
# 16, is_training, beta_init=0.0, gamma_init=0.1, ema_decay_rate=0.5, moment_axes=[0,1,2], variance_epsilon=1e-3)
self.layers["deconv1_act"] = m.ActivationModule("deconv1_act", \
lrn_relu)
with tf.name_scope('deconvolutional_layer_2'):
self.layers["deconv2"] = m.Conv2DTransposeModule("deconv2", \
[3,3,IMAGE_CHANNELS,32], [1,2,2,1], [FLAGS.batchsize,IMAGE_HEIGHT,IMAGE_WIDTH,IMAGE_CHANNELS])
self.layers["deconv2_act"] = m.ActivationModule("deconv2_act", \
lrn_relu)
# connect all modules of the network in a meaningful way
# -----
with tf.name_scope('wiring_of_modules'):
self.layers["conv0"].add_input(self.layers["inp_norm"])
self.layers["conv1"].add_input(self.layers["conv0"])
self.layers["conv2"].add_input(self.layers["conv1"])
self.layers["bottleneck_switch"].add_input(self.layers["conv2"])
self.layers["bottleneck_switch"].add_input(self.layers["bottleneck_canvas"])
self.layers["deconv0"].add_input(self.layers["bottleneck_switch"])
self.layers["deconv0_act"].add_input(self.layers["deconv0"])
self.layers["deconv1"].add_input(self.layers["deconv0_act"])
self.layers["deconv1_act"].add_input(self.layers["deconv1"])
self.layers["deconv2"].add_input(self.layers["deconv1_act"])
self.layers["deconv2_act"].add_input(self.layers["deconv2"])
with tf.name_scope('input_output'):
self.input_module = self.layers["inp_norm"]
self.output_module = self.layers["deconv2_act"]
def define_inner_modules(self, name, is_training, activations, conv_filter_shapes, bias_shapes, ksizes, pool_strides, topdown_filter_shapes, topdown_output_shapes, keep_prob, FLAGS):
# create all modules of the network
# -----
self.layers = {}
with tf.name_scope('input_normalization'):
self.layers["inp_norm"] = m.NormalizationModule("inp_norm")
with tf.name_scope('convolutional_layer_0'):
if FLAGS.batchnorm:
self.layers["conv0"] = m.TimeConvolutionalLayerWithBatchNormalizationModule("conv0", bias_shapes[0][-1], is_training, 0.0, 1.0, 0.5, activations[0], conv_filter_shapes[0], [1,1,1,1], bias_shapes[0])
else:
self.layers["conv0"] = m.TimeConvolutionalLayerModule("conv0", activations[0], conv_filter_shapes[0], [1,1,1,1], bias_shapes[0])
with tf.name_scope('lateral_layer_0'):
lateral_filter_shape = conv_filter_shapes[0]
tmp = lateral_filter_shape[2]
lateral_filter_shape[2] = lateral_filter_shape[3]
lateral_filter_shape[3] = tmp
self.layers["lateral0"] = m.Conv2DModule("lateral0", lateral_filter_shape, [1,1,1,1])
self.layers["lateral0_batchnorm"] = m.BatchNormalizationModule("lateral0_batchnorm", lateral_filter_shape[-1], is_training, beta_init=0.0, gamma_init=0.1, ema_decay_rate=0.5, moment_axes=[0,1,2], variance_epsilon=1e-3)
with tf.name_scope('pooling_layer_0'):
self.layers["pool0"] = m.MaxPoolingWithArgmaxModule("pool0", ksizes[0], pool_strides[0])
with tf.name_scope('dropout_layer_0'):
self.layers['dropoutc0'] = m.DropoutModule('dropoutc0', keep_prob=keep_prob)
with tf.name_scope('convolutional_layer_1'):
if FLAGS.batchnorm:
self.layers["conv1"] = m.TimeConvolutionalLayerWithBatchNormalizationModule("conv1", bias_shapes[1][-1], is_training, 0.0, 1.0, 0.5, activations[1], conv_filter_shapes[1], [1,1,1,1], bias_shapes[1])
else:
self.layers["conv1"] = m.TimeConvolutionalLayerModule("conv1", activations[1], conv_filter_shapes[1], [1,1,1,1], bias_shapes[1])
with tf.name_scope('topdown_layer_0'):
self.layers["topdown0"] = m.Conv2DTransposeModule("topdown0", topdown_filter_shapes[0], [1,2,2,1], topdown_output_shapes[0])
self.layers["topdown0_batchnorm"] = m.BatchNormalizationModule("topdown0_batchnorm",topdown_output_shapes[0][-1], is_training, beta_init=0.0, gamma_init=0.1, ema_decay_rate=0.5, moment_axes=[0,1,2], variance_epsilon=1e-3)
with tf.name_scope('lateral_layer_1'):
lateral_filter_shape = conv_filter_shapes[1]
tmp = lateral_filter_shape[2]
lateral_filter_shape[2] = lateral_filter_shape[3]
lateral_filter_shape[3] = tmp
self.layers["lateral1"] = m.Conv2DModule("lateral1", lateral_filter_shape, [1,1,1,1])
self.layers["lateral1_batchnorm"] = m.BatchNormalizationModule("lateral1_batchnorm", lateral_filter_shape[-1], is_training, beta_init=0.0, gamma_init=0.1, ema_decay_rate=0.5, moment_axes=[0,1,2], variance_epsilon=1e-3)
with tf.name_scope('pooling_layer_1'):
self.layers["pool1"] = m.MaxPoolingWithArgmaxModule("pool1", ksizes[0], pool_strides[1])
self.layers["flatpool1"] = m.FlattenModule("flatpool1")
with tf.name_scope('dropout_layer_1'):
self.layers['dropoutc1'] = m.DropoutModule('dropoutc1', keep_prob=keep_prob)
with tf.name_scope('fully_connected_layer_0'):
if FLAGS.batchnorm:
self.layers["fc0"] = m.FullyConnectedLayerWithBatchNormalizationModule("fc0", bias_shapes[-1][-1], is_training, 0.0, 1.0, 0.5, activations[2], int(np.prod(np.array(bias_shapes[1]) / np.array(pool_strides[1]))), np.prod(bias_shapes[2]))
else:
self.layers["fc0"] = m.FullyConnectedLayerModule("fc0", activations[2], int(np.prod(np.array(bias_shapes[1]) / np.array(pool_strides[1]))), np.prod(bias_shapes[2]))
with tf.name_scope('unpooling_layer_1'):
self.layers["unpool1"] = m.UnpoolingModule("unpool1", ksizes[0], pool_strides[0])
with tf.name_scope('unconvolution_layer_1'):
self.layers["unconv1"] = m.UnConvolutionModule("unconv1", conv_filter_shapes[1],[1,1,1,1], topdown_output_shapes[0][:1]+ bias_shapes[1][1:])
with tf.name_scope('unpooling_layer_0'):
self.layers["unpool0"] = m.UnpoolingModule("unpool0", ksizes[0], pool_strides[0])
with tf.name_scope('unconvolution_layer_0'):
self.layers["unconv0"] = m.UnConvolutionModule("unconv0", conv_filter_shapes[0],[1,1,1,1],topdown_output_shapes[0])
# connect all modules of the network in a meaningful way
# -----
with tf.name_scope('wiring_of_modules'):
self.layers["conv0"].add_input(self.layers["inp_norm"], 0)
self.layers["pool0"].add_input(self.layers["conv0"])
self.layers["dropoutc0"].add_input(self.layers["pool0"])
self.layers["conv1"].add_input(self.layers["dropoutc0"], 0)
self.layers["pool1"].add_input(self.layers["conv1"])
self.layers["dropoutc1"].add_input(self.layers["pool1"])
self.layers["flatpool1"].add_input(self.layers["dropoutc1"])
self.layers["fc0"].add_input(self.layers["flatpool1"])
#try out unpooling
self.layers["unpool1"].add_input(self.layers["pool1"])
self.layers["unconv1"].add_input(self.layers["unpool1"])
self.layers["unconv1"].add_input(self.layers["conv1"])
self.layers["unpool0"].add_input(self.layers["unconv1"])
self.layers["unpool0"].add_input(self.layers["pool0"])
self.layers["unconv0"].add_input(self.layers["unpool0"])
self.layers["unconv0"].add_input(self.layers["conv0"])
if "L" in FLAGS.architecture:
if FLAGS.batchnorm:
self.layers["lateral0"].add_input(self.layers["conv0"].preactivation)
self.layers["lateral0_batchnorm"].add_input(self.layers["lateral0"])
self.layers["conv0"].add_input(self.layers["lateral0_batchnorm"], -1)
self.layers["lateral1"].add_input(self.layers["conv1"].preactivation)
self.layers["lateral1_batchnorm"].add_input(self.layers["lateral1"])
self.layers["conv1"].add_input(self.layers["lateral1_batchnorm"], -1)
else:
self.layers["lateral0"].add_input(self.layers["conv0"].preactivation)
self.layers["conv0"].add_input(self.layers["lateral0"], -1)
self.layers["lateral1"].add_input(self.layers["conv1"].preactivation)
self.layers["conv1"].add_input(self.layers["lateral1"], -1)
if "T" in FLAGS.architecture:
if FLAGS.batchnorm:
self.layers["topdown0_batchnorm"].add_input(self.layers["topdown0"])
self.layers["conv0"].add_input(self.layers["topdown0_batchnorm"], -1)
self.layers["topdown0"].add_input(self.layers["conv1"].preactivation)
else:
self.layers["conv0"].add_input(self.layers["topdown0"], -1)
self.layers["topdown0"].add_input(self.layers["conv1"].preactivation)
with tf.name_scope('input_output'):
self.input_module = self.layers["inp_norm"]
self.output_module = self.layers["fc0"]