def _setup_prediction(self):
self.batch_size = self.inputs['data'].get_shape().as_list()[0]
# LAYER 0: INPUT layer: 2D tensor. Three rows (channels) x window size
current_layer = self.inputs['data']
d1, d2, d3 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] LAYER input shape = "+str(d1)+"x"+str(d2)+"x"+str(d3))
# LAYER 1-8: 8 CONVOLUTINAL layers, 32 1D kernels of size 3 each, stride 2, zero padding
c = 32 # number of channels per conv layer
ksize = self.config.ksize # size of the convolution kernel
depth = self.config.num_conv_layers
for i in range(depth):
current_layer = layers.conv1(current_layer, c, ksize, stride=2, scope='conv{}'.format(i+1), padding='SAME')
d1, d2, d3 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] LAYER "+'conv{}'.format(i+1)+" shape = "+str(d1)+"x"+str(d2)+"x"+str(d3))
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['conv{}'.format(i+1)] = current_layer
if self.config.pooling:
# Max Pooling (down-sampling) with strides of 2 and kernel size of 2
#https://www.tensorflow.org/api_docs/python/tf/nn/pool
current_layer = tf.nn.pool(current_layer, window_shape=[self.config.pooling_window], pooling_type='MAX', padding='SAME', strides=[self.config.pooling_stride])
d1, d2, d3 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] LAYER "+'max_pool{}'.format(i+1)+" shape = "+str(d1)+"x"+str(d2)+"x"+str(d3))
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['pool{}'.format(i+1)] = current_layer
bs, width, _ = current_layer.get_shape().as_list()
current_layer = tf.reshape(current_layer, [bs, width*c], name="reshape")
# FULLY CONNECTED LAYERS
for i in range(0, self.config.num_fc_layers-1):
# LAYER 10: FULLY CONNECTED
current_layer = layers.fc(current_layer, self.config.fc_size, scope='fc{}'.format(i+1), activation_fn=None)
d1, d2 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] FC LAYER "+'{}'.format(i+1)+" shape = "+str(d1)+"x"+str(d2))
self.layers['fc{}'.format(i+1)] = current_layer
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
# LAYER: LAST FULLY CONNECTED (LOGITS)
current_layer = layers.fc(current_layer, self.config.n_clusters, scope='logits', activation_fn=None)
d1, d2 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] FC LAYER LOGITS shape = "+str(d1)+"x"+str(d2))
self.layers['logits'] = current_layer
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['class_prob'] = tf.nn.softmax(current_layer, name='class_prob')
self.layers['class_prediction'] = tf.argmax(self.layers['class_prob'], 1, name='class_pred')
tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(self.config.regularization),
weights_list=tf.get_collection(tf.GraphKeys.WEIGHTS))
def _setup_prediction(self):
self.batch_size = self.inputs['data'].get_shape().as_list()[0]
current_layer = self.inputs['data']
d1, d2, d3 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] INPUT LAYER shape = "+str(d1)+"x"+str(d2)+"x"+str(d3))
c = 32 # number of channels per conv layer
ksize = 3 # size of the convolution kernel
depth = 8
for i in range(depth):
current_layer = layers.conv1(current_layer, c, ksize, stride=2, scope='conv{}'.format(i+1), padding='SAME')
d1, d2, d3 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] LAYER "+'conv{}'.format(i+1)+" shape = "+str(d1)+"x"+str(d2)+"x"+str(d3))
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['conv{}'.format(i+1)] = current_layer
bs, width, _ = current_layer.get_shape().as_list()
#print ("[\033[94mINFO\033[0m] LAST LAYER CONV shape = "+str(bs)+"x"+str(width)+"x"+str(_))
current_layer = tf.reshape(current_layer, [bs, width*c], name="reshape")
d1, d2 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] LAST LAYER CONV RESHAPED shape = "+str(d1)+"x"+str(d2))
current_layer = layers.fc(current_layer, self.config.n_clusters, scope='logits', activation_fn=None)
d1, d2 = current_layer.get_shape().as_list()
print ("[\033[94mINFO\033[0m] FC LAYER shape = "+str(d1)+"x"+str(d2))
self.layers['logits'] = current_layer
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['class_prob'] = tf.nn.softmax(current_layer, name='class_prob')
self.layers['class_prediction'] = tf.argmax(self.layers['class_prob'], 1, name='class_pred')
tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(self.config.regularization),
weights_list=tf.get_collection(tf.GraphKeys.WEIGHTS))
def _setup_prediction(self):
self.batch_size = self.inputs['data'].get_shape().as_list()[0]
current_layer = self.inputs['data']
print current_layer
c = 64 # number of channels per conv layer
ksize = 3 # size of the convolution kernel
depth = 6
for i in range(depth):
current_layer = layers.conv1(current_layer,
c,
ksize,
stride=1,
scope='conv{}'.format(i + 1),
padding='SAME')
current_layer = layers.max_pooling(current_layer, 2)
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['conv{}'.format(i + 1)] = current_layer
bs, width, _ = current_layer.get_shape().as_list()
current_layer = tf.reshape(current_layer, [bs, width * c],
name="reshape")
current_layer = layers.fc(current_layer,
3,
scope='det_logits',
activation_fn=None)
self.layers['logits'] = current_layer
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['class_prob'] = tf.nn.softmax(current_layer,
name='det_class_prob')
self.layers['class_prediction'] = tf.argmax(self.layers['class_prob'],
1,
name='det_class_pred')
tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(self.config.regularization),
weights_list=tf.get_collection(tf.GraphKeys.WEIGHTS))
def _setup_prediction(self):
# def inputs
self.batch_size = self.inputs['data'].get_shape().as_list()[0]
current_layer = self.inputs['data']
current_layer = tf.squeeze(current_layer, [1])
# def hypo-params
c = 32 # number of channels per conv layer
ksize = 3 # size of the convolution kernel
depth = 8
lamb = 1e-4 # l2 reg
# def model structure
for i in range(depth):
current_layer = layers.conv1(current_layer, c, ksize, stride=1,
activation_fn=None, scope='conv{}'.format(i+1))
current_layer = layers.batch_norm(current_layer,
'conv{}_batch_norm'.format(i+1),
self.is_training)
current_layer = tf.nn.relu(current_layer)
self.layers['conv0_{}'.format(i+1)] = current_layer
current_layer = layers.max_pooling(current_layer, 2)
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['conv{}'.format(i+1)] = current_layer
# softmax regression
bs, width, _ = current_layer.get_shape().as_list()
current_layer = tf.reshape(current_layer, [bs, width*c], name="reshape")
current_layer = layers.fc(current_layer, 2, scope='det_logits', activation_fn=None)
self.layers['logits'] = current_layer
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
# output prediction
self.layers['class_prob'] = tf.nn.softmax(current_layer, name='det_class_prob')
self.layers['class_prediction'] = tf.argmax(self.layers['class_prob'], 1, name='det_class_pred')
# add L2 regularization
tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(lamb),
weights_list=tf.get_collection(tf.GraphKeys.WEIGHTS))
def _setup_prediction(self):
self.batch_size = self.inputs['data'].get_shape().as_list()[0]
current_layer = self.inputs['data']
#n_channels = 32 # number of channels per conv layer
ksize = 3 # size of the convolution kernel
for i in range(self.n_conv_layers):
current_layer = layers.conv1(current_layer, self.n_channels, ksize, stride=2, scope='conv{}'.format(i + 1), padding='SAME')
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
self.layers['conv{}'.format(i + 1)] = current_layer
bs, width, _ = current_layer.get_shape().as_list()
n_fc_nodes = width * self.n_channels
current_layer = tf.reshape(current_layer, [bs, n_fc_nodes], name="reshape")
# 20180916 AJL - include stram_max in fc layers
stream_max_tensor = tf.expand_dims(self.inputs['stream_max'], 1)
current_layer = tf.concat([current_layer, stream_max_tensor], 1)
n_fc_nodes += 1
for i in range(self.n_fc_layers - 1):
current_layer = layers.fc(current_layer, n_fc_nodes, scope='fc{}'.format(i + 1), activation_fn=tf.nn.relu)
current_layer = layers.fc(current_layer, self.cfg.n_distances + self.cfg.n_magnitudes + self.cfg.n_depths + self.cfg.n_azimuths, \
scope='logits', activation_fn=None)
istart = 0
iend = self.cfg.n_distances
self.layers['distance_logits'] = current_layer[ : , istart : iend]
self.layers['distance_prob'] = tf.nn.softmax(self.layers['distance_logits'], name='distance_prob')
self.layers['distance_prediction'] = tf.argmax(self.layers['distance_prob'], 1, name='distance_pred')
istart = iend
self.layers['magnitude_logits'] = tf.constant(-1)
self.layers['magnitude_prob'] = tf.constant(-1)
self.layers['magnitude_prediction'] = tf.constant(-1)
if self.cfg.n_magnitudes > 0:
iend += self.cfg.n_magnitudes
self.layers['magnitude_logits'] = current_layer[ : , istart : iend]
self.layers['magnitude_prob'] = tf.nn.softmax(self.layers['magnitude_logits'], name='magnitude_prob')
self.layers['magnitude_prediction'] = tf.argmax(self.layers['magnitude_prob'], 1, name='magnitude_pred')
istart = iend
self.layers['depth_logits'] = tf.constant(-1)
self.layers['depth_prob'] = tf.constant(-1)
self.layers['depth_prediction'] = tf.constant(-1)
if self.cfg.n_depths > 0:
iend += self.cfg.n_depths
self.layers['depth_logits'] = current_layer[ : , istart : iend]
self.layers['depth_prob'] = tf.nn.softmax(self.layers['depth_logits'], name='depth_prob')
self.layers['depth_prediction'] = tf.argmax(self.layers['depth_prob'], 1, name='depth_pred')
istart = iend
self.layers['azimuth_logits'] = tf.constant(-1)
self.layers['azimuth_prob'] = tf.constant(-1)
self.layers['azimuth_prediction'] = tf.constant(-1)
if self.cfg.n_azimuths > 0:
iend += self.cfg.n_azimuths
self.layers['azimuth_logits'] = current_layer[ : , istart : iend]
self.layers['azimuth_prob'] = tf.nn.softmax(self.layers['azimuth_logits'], name='azimuth_prob')
self.layers['azimuth_prediction'] = tf.argmax(self.layers['azimuth_prob'], 1, name='azimuth_pred')
istart = iend
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, current_layer)
tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(self.cfg.regularization),
weights_list=tf.get_collection(tf.GraphKeys.WEIGHTS))
