def cnn(mini_timeslice):
""" input: event tensor numpy shape 1, 13, 13, 18, 3"""
conv1 = tf.nn.relu(
conv3d(mini_timeslice, weights["l1"]) + biases["l1"])
conv1 = tf.contrib.layers.batch_norm(conv1)
conv2 = tf.nn.relu(
conv3d(conv1, weights["l2"]) + biases["l2"])
conv1 = tf.contrib.layers.batch_norm(conv1)
conv2 = maxpool3d(conv2)
conv2 = tf.contrib.layers.batch_norm(conv2)
conv3 = tf.nn.relu(
conv3d(conv2, weights["l3"]) + biases["l3"])
conv3 = tf.contrib.layers.batch_norm(conv3)
conv3 = maxpool3d(conv3)
conv4 = tf.nn.relu(
conv3d(conv3, weights["l4"]) + biases["l4"])
return conv4
def cnn(mini_timeslice):
""" input: event tensor numpy shape 1, 13, 13, 18, 3"""
conv1 = tf.nn.relu(
conv3d(mini_timeslice, weights["l1"]) + biases["l1"])
conv1 = tf.contrib.layers.batch_norm(conv1)
conv2 = tf.nn.relu(
conv3d(conv1, weights["l2"]) + biases["l2"])
conv1 = tf.contrib.layers.batch_norm(conv1)
conv2 = maxpool3d(conv2)
conv2 = tf.contrib.layers.batch_norm(conv2)
#elements = np.prod(conv2._shape_as_list()[1:])
fc = tf.reshape(conv2, [-1, 11025])
fc = tf.nn.relu(
tf.matmul(fc, weights["l3"]) + biases["l3"])
fc = tf.nn.dropout(fc, keep_prob)
conv1 = tf.contrib.layers.batch_norm(fc)
fc = tf.nn.relu(
tf.matmul(fc, weights["l4"]) + biases["l4"])
return fc
def cnn(input_slice):
""" input: event tensor numpy shape 1, 13, 13, 18, 3"""
conv1 = tf.nn.relu(conv3d(input_slice, weights["l1"]) + biases["l1"])
conv2 = tf.nn.relu(conv3d(conv1, weights["l2"]) + biases["l2"])
conv2 = maxpool3d(conv2)
conv3 = tf.nn.relu(conv3d(conv2, weights["l3"] + biases["l3"]))
fc = tf.reshape(conv3, [-1, 7 * 7 * 9 * nodes["l3"]])
fc = tf.nn.sigmoid(tf.matmul(fc, weights["l4"]) + biases["l4"])
fc = tf.nn.sigmoid(tf.matmul(fc, weights["l5"]) + biases["l5"])
return fc
def km3nnet(x):
""" input: event tensor numpy shape num_minitimeslices, 18, 18, 13, 3
output: label prediction shape 3 (one hot encoded)"""
# loop over mini time slices
nodes = {"l1": 25,
"l2": 25,
"l3": 80,
"l4": 40,
"l5": 20}
weights = {"l1": weight([4, 4, 4, 3, nodes["l1"]]),
"l2": weight([3, 3, 3, nodes["l1"], nodes["l2"]]),
"l3": weight([11025, nodes["l3"]]),
"l4": weight([nodes["l3"], nodes["l4"]])}
biases = {"l1": bias(nodes["l1"]),
"l2": bias(nodes["l2"]),
"l3": bias(nodes["l3"]),
"l4": bias(nodes["l4"])}
conv1 = tf.nn.relu(
conv3d(x, weights["l1"]) + biases["l1"])
conv2 = tf.nn.relu(
conv3d(conv1, weights["l2"]) + biases["l2"])
conv2 = maxpool3d(conv2)
elements = np.prod(conv2._shape_as_list()[1:])
fc = tf.reshape(conv2, [-1, elements])
fc = tf.nn.relu(
tf.matmul(fc, weights["l3"]) + biases["l3"])
fc = tf.nn.relu(
tf.matmul(fc, weights["l4"]) + biases["l4"])
fc = tf.reshape(fc, [50, 1, 40])
c = tf.unstack(fc, num_mini_timeslices, 0)
lstm_layer = tf.contrib.rnn.BasicLSTMCell(nodes["l5"], forget_bias=1.)
outputs, _ = tf.contrib.rnn.static_rnn(lstm_layer, c, dtype=tf.float32)
def cnn(input_slice):
""" input: event tensor numpy shape 1, 13, 13, 18, 3"""
conv1 = tf.nn.relu(
conv3d(input_slice, weights["l1"]) + biases["l1"])
conv2 = tf.nn.relu(
conv3d(conv1, weights["l2"]) + biases["l2"])
conv2 = maxpool3d(conv2)
elements = np.prod(conv2._shape_as_list()[1:])
fc = tf.reshape(conv2, [-1, elements])
fc = tf.nn.sigmoid(
tf.matmul(fc, weights["l3"]) + biases["l3"])
fc = tf.nn.dropout(fc, keep_prob)
fc = tf.nn.sigmoid(
tf.matmul(fc, weights["l4"]) + biases["l4"])
return fc