def triage_wf(self, wf_tf, threshold):
"""
Run neural net triage on given spike waveforms
Parameters:
-----------
wf_tf: tf tensor (n_spikes, n_temporal_length, n_neigh)
tf tensor that produces spikes waveforms
threshold: int
threshold used on a probability obtained after nn to determine
whether it is a clear spike
Returns:
-----------
tf tensor (n_spikes,)
a boolean tensorflow tensor that produces indices of
clear spikes
"""
# get parameters
K1, K2 = self.filters_dict['filters']
# first layer: temporal feature
layer1 = tf.nn.relu(
conv2d_VALID(tf.expand_dims(wf_tf, -1), self.W1) + self.b1)
# second layer: feataure mapping
layer11 = tf.nn.relu(conv2d(layer1, self.W11) + self.b11)
# third layer: spatial convolution
o_layer = conv2d_VALID(layer11, self.W2) + self.b2
# thrshold it
return o_layer[:, 0, 0, 0] > np.log(threshold / (1 - threshold))
def _make_training_graph(cls, waveform_length, n_neighbors, vars_dict):
"""Make graph for training
Returns
-------
x_tf: tf.tensor
Input tensor
y_tf: tf.tensor
Labels tensor
o_layer: tf.tensor
Output tensor
"""
# x and y input tensors
x_tf = tf.placeholder("float", [None, waveform_length, n_neighbors])
y_tf = tf.placeholder("float", [None])
input_tf = tf.expand_dims(x_tf, -1)
vars_dict, layer11 = (NeuralNetDetector._make_network(input_tf,
vars_dict,
padding='VALID'))
W2 = vars_dict['W2']
b2 = vars_dict['b2']
# third layer: spatial convolution
o_layer = tf.squeeze(conv2d_VALID(layer11, W2) + b2)
# sigmoid
sigmoid = tf.sigmoid(o_layer)
return x_tf, y_tf, o_layer, sigmoid
def make_o_layer_tf_tensors(self, x_tf, channel_index):
"""
Make a tensorflow tensor that outputs spike index
Parameters
-----------
x_tf: tf.tensors (n_observations, n_channels)
placeholder of recording for running tensorflow
channel_index: np.array (n_channels, n_neigh)
Each row indexes its neighboring channels.
For example, channel_index[c] is the index of
neighboring channels (including itself)
If any value is equal to n_channels, it is nothing but
a space holder in a case that a channel has less than
n_neigh neighboring channels
Returns
-------
output_tf: tf tensor (n_observations, n_channels)
tensorflow tensor that produces spike_index
"""
# get parameters
K1, K2 = self.filters_dict['filters']
nneigh = self.filters_dict['n_neighbors']
# save neighbor channel index
self.channel_index = channel_index[:, :nneigh]
# Temporal shape of input
T = tf.shape(x_tf)[0]
# input tensor into CNN
x_cnn_tf = tf.expand_dims(tf.expand_dims(x_tf, -1), 0)
# NN structures
# first temporal layer
layer1 = tf.nn.relu(conv2d(x_cnn_tf, self.W1) + self.b1)
# second temporal layer
layer11 = tf.nn.relu(conv2d(layer1, self.W11) + self.b11)
# first spatial layer
zero_added_layer11 = tf.concat(
(tf.transpose(layer11, [2, 0, 1, 3]), tf.zeros((1, 1, T, K2))),
axis=0)
temp = tf.transpose(tf.gather(zero_added_layer11, self.channel_index),
[0, 2, 3, 1, 4])
temp2 = conv2d_VALID(tf.reshape(temp, [-1, T, nneigh, K2]),
self.W2) + self.b2
# output layer
# o_layer: [1, temporal, spatial, 1]
o_layer = tf.transpose(temp2, [2, 1, 0, 3])[0, :, :, 0]
output_tf = tf.sigmoid(o_layer)
return output_tf
def get_spikes(self, x_tf, T, nneigh, c_idx, temporal_window, th):
"""
Detects and indexes spikes from the recording. The recording will
be chopped to minibatches if its temporal length
exceeds 10000. A spike is detected at [t, c] when the output
probability of the neural network detector crosses
the detection threshold at time t and channel c. For temporal
duplicates within a certain temporal radius,
the temporal index corresponding to the largest output probability
is assigned. For spatial duplicates within
certain neighboring channels, the channel with the highest energy
is assigned.
Parameters:
-----------
X: np.array
[number of channels, temporal length] raw recording.
Returns:
-----------
index: np.array
[number of detected spikes, 3] returned indices for spikes.
First column corresponds to temporal location;
second column corresponds to spatial (channel) location.
"""
# get parameters
K1, K2 = self.filters_dict['filters']
# NN structures
layer1 = tf.nn.relu(
conv2d(tf.expand_dims(tf.expand_dims(x_tf, -1), 0), self.W1) +
self.b1)
layer11 = tf.nn.relu(conv2d(layer1, self.W11) + self.b11)
zero_added_layer11 = tf.concat(
(tf.transpose(layer11, [2, 0, 1, 3]), tf.zeros((1, 1, T, K2))),
axis=0)
temp = tf.transpose(tf.gather(zero_added_layer11, c_idx),
[0, 2, 3, 1, 4])
temp2 = conv2d_VALID(tf.reshape(temp, [-1, T, nneigh, K2]),
self.W2) + self.b2
o_layer = tf.transpose(temp2, [2, 1, 0, 3])
temporal_max = max_pool(o_layer, [1, temporal_window, 1, 1])
local_max_idx = tf.where(
tf.logical_and(o_layer[0, :, :, 0] >= temporal_max[0, :, :, 0],
o_layer[0, :, :, 0] > np.log(th / (1 - th))))
return local_max_idx
def _make_network(cls, input_tensor, filters_size, waveform_length,
n_neighbors):
"""Mates tensorflow network, from first layer to output layer
"""
K1, K2 = filters_size
# initialize and save nn weights
W1 = weight_variable([waveform_length, 1, 1, K1])
b1 = bias_variable([K1])
W11 = weight_variable([1, 1, K1, K2])
b11 = bias_variable([K2])
W2 = weight_variable([1, n_neighbors, K2, 1])
b2 = bias_variable([1])
# first layer: temporal feature
layer1 = tf.nn.relu(
conv2d_VALID(tf.expand_dims(input_tensor, -1), W1) + b1)
# second layer: feataure mapping
layer11 = tf.nn.relu(conv2d(layer1, W11) + b11)
# third layer: spatial convolution
o_layer = conv2d_VALID(layer11, W2) + b2
vars_dict = {
"W1": W1,
"W11": W11,
"W2": W2,
"b1": b1,
"b11": b11,
"b2": b2
}
return o_layer, vars_dict
def make_detection_tf_tensors(self, x_tf, channel_index, threshold):
"""
Make a tensorflow tensor that outputs spike index
Parameters
-----------
x_tf: tf.tensors (n_observations, n_channels)
placeholder of recording for running tensorflow
channel_index: np.array (n_channels, n_neigh)
Each row indexes its neighboring channels.
For example, channel_index[c] is the index of
neighboring channels (including itself)
If any value is equal to n_channels, it is nothing but
a space holder in a case that a channel has less than
n_neigh neighboring channels
threshold: int
threshold on a probability to determine
location of spikes
Returns
-------
spike_index_tf: tf tensor (n_spikes, 2)
tensorflow tensor that produces spike_index
"""
# get parameters
K1, K2 = self.filters_dict['filters']
nneigh = self.filters_dict['n_neighbors']
# save neighbor channel index
self.channel_index = channel_index[:, :nneigh]
# Temporal shape of input
T = tf.shape(x_tf)[0]
# input tensor into CNN
x_cnn_tf = tf.expand_dims(tf.expand_dims(x_tf, -1), 0)
# NN structures
# first temporal layer
layer1 = tf.nn.relu(conv2d(x_cnn_tf, self.W1) + self.b1)
# second temporal layer
layer11 = tf.nn.relu(conv2d(layer1, self.W11) + self.b11)
# first spatial layer
zero_added_layer11 = tf.concat(
(tf.transpose(layer11, [2, 0, 1, 3]), tf.zeros((1, 1, T, K2))),
axis=0)
temp = tf.transpose(tf.gather(zero_added_layer11, self.channel_index),
[0, 2, 3, 1, 4])
temp2 = conv2d_VALID(tf.reshape(temp, [-1, T, nneigh, K2]),
self.W2) + self.b2
# output layer
o_layer = tf.transpose(temp2, [2, 1, 0, 3])
# temporal max
temporal_max = max_pool(o_layer, [1, 3, 1, 1]) - 1e-8
# spike index is local maximum crossing a threshold
spike_index_tf = tf.cast(
tf.where(
tf.logical_and(
o_layer[0, :, :, 0] >= temporal_max[0, :, :, 0],
o_layer[0, :, :, 0] > np.log(threshold /
(1 - threshold)))), 'int32')
return spike_index_tf
def fit(self, x_train, y_train):
"""
Trains the neural network detector for spike detection
Parameters:
-----------
x_train: np.array
[number of training data, temporal length, number of channels]
augmented training data consisting of
isolated spikes, noise and misaligned spikes.
y_train: np.array
[number of training data] label for x_train. '1' denotes presence
of an isolated spike and '0' denotes
the presence of a noise data or misaligned spike.
path_to_model: string
name of the .ckpt to be saved
"""
######################
# Loading parameters #
######################
logger = logging.getLogger(__name__)
# get parameters
n_data, waveform_length_train, n_neighbors_train = x_train.shape
if self.waveform_length != waveform_length_train:
raise ValueError('waveform length from network ({}) does not '
'match training data ({})'
.format(self.waveform_length,
waveform_length_train))
if self.n_neighbors != n_neighbors_train:
raise ValueError('number of n_neighbors from network ({}) does '
'not match training data ({})'
.format(self.n_neigh,
n_neighbors_train))
####################
# Building network #
####################
# x and y input tensors
x_tf = tf.placeholder("float", [self.n_batch, self.waveform_length,
self.n_neighbors])
y_tf = tf.placeholder("float", [self.n_batch])
input_tf = tf.expand_dims(x_tf, -1)
vars_dict, layer11 = (NeuralNetDetector
._make_network(input_tf,
self.waveform_length,
self.filters_size,
self.n_neighbors,
padding='VALID'))
W2 = vars_dict['W2']
b2 = vars_dict['b2']
# third layer: spatial convolution
o_layer = tf.squeeze(conv2d_VALID(layer11, W2) + b2)
##########################
# Optimization objective #
##########################
# cross entropy
_ = tf.nn.sigmoid_cross_entropy_with_logits(logits=o_layer,
labels=y_tf)
cross_entropy = tf.reduce_mean(_)
weights = tf.trainable_variables()
# regularization term
l2_regularizer = (tf.contrib.layers
.l2_regularizer(scale=self.l2_reg_scale))
regularization = tf.contrib.layers.apply_regularization(l2_regularizer,
weights)
regularized_loss = cross_entropy + regularization
# train step
train_step = (tf.train.AdamOptimizer(self.train_step_size)
.minimize(regularized_loss))
############
# Training #
############
# saver
saver = tf.train.Saver(vars_dict)
logger.debug('Training detector network...')
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
pbar = trange(self.n_iter)
for i in pbar:
# sample n_batch observations from 0, ..., n_data
idx_batch = np.random.choice(n_data, self.n_batch,
replace=False)
res = sess.run([train_step, regularized_loss],
feed_dict={x_tf: x_train[idx_batch],
y_tf: y_train[idx_batch]})
if i % 100 == 0:
pbar.set_description('Loss: %s' % res[1])
logger.debug('Saving network: %s', self.path_to_model)
saver.save(sess, self.path_to_model)
# estimate tp and fp with a sample
idx_batch = np.random.choice(n_data, self.n_batch, replace=False)
output = sess.run(o_layer, feed_dict={x_tf: x_train[idx_batch]})
y_test = y_train[idx_batch]
tp = np.mean(output[y_test == 1] > 0)
fp = np.mean(output[y_test == 0] > 0)
logger.debug('Approximate training true positive rate: '
+ str(tp) + ', false positive rate: ' + str(fp))
path_to_params = change_extension(self.path_to_model, 'yaml')
logger.debug('Saving network parameters: %s', path_to_params)
save_detect_network_params(filters_size=self.filters_size,
waveform_length=self.waveform_length,
n_neighbors=self.n_neighbors,
output_path=path_to_params)
def _make_graph(cls, threshold, channel_index, waveform_length,
filters_size, n_neigh):
"""Build tensorflow graph with input and two output layers
Parameters
-----------
x_tf: tf.tensors (n_observations, n_channels)
placeholder of recording for running tensorflow
channel_index: np.array (n_channels, n_neigh)
Each row indexes its neighboring channels. For example,
channel_index[c] is the index of neighboring channels (including
itself) If any value is equal to n_channels, it is nothing but
a placeholder in a case that a channel has less than n_neigh
neighboring channels
threshold: int
threshold on a probability to determine
location of spikes
Returns
-------
spike_index_tf: tf tensor (n_spikes, 2)
tensorflow tensor that produces spike_index
"""
######################
# Loading parameters #
######################
# TODO: need to ask why we are sending different channel indexes
# save neighbor channel index
small_channel_index = channel_index[:, :n_neigh]
# placeholder for input recording
x_tf = tf.placeholder("float", [None, None])
# Temporal shape of input
T = tf.shape(x_tf)[0]
####################
# Building network #
####################
# input tensor into CNN - add one dimension at the beginning and
# at the end
x_cnn_tf = tf.expand_dims(tf.expand_dims(x_tf, -1), 0)
vars_dict, layer11 = cls._make_network(x_cnn_tf,
waveform_length,
filters_size,
n_neigh,
padding='SAME')
W2 = vars_dict['W2']
b2 = vars_dict['b2']
K1, K2 = filters_size
# first spatial layer
zero_added_layer11 = tf.concat((tf.transpose(layer11, [2, 0, 1, 3]),
tf.zeros((1, 1, T, K2))),
axis=0)
temp = tf.transpose(tf.gather(zero_added_layer11, small_channel_index),
[0, 2, 3, 1, 4])
temp2 = conv2d_VALID(tf.reshape(temp, [-1, T, n_neigh, K2]), W2) + b2
o_layer = tf.transpose(temp2, [2, 1, 0, 3])
################################
# Output layer transformations #
################################
o_layer_val = tf.squeeze(o_layer)
# probability output - just sigmoid of output layer
probability_tf = tf.sigmoid(o_layer_val)
# spike index output (local maximum crossing a threshold)
temporal_max = tf.squeeze(max_pool(o_layer, [1, 3, 1, 1]) - 1e-8)
higher_than_max_pool = o_layer_val >= temporal_max
higher_than_threshold = (o_layer_val >
np.log(threshold / (1 - threshold)))
both_higher = tf.logical_and(higher_than_max_pool,
higher_than_threshold)
index_all = tf.cast(tf.where(both_higher), 'int32')
spike_index_tf = cls._remove_edge_spikes(x_tf, index_all,
waveform_length)
# waveform output from spike index output
waveform_tf = cls._make_waveform_tf(x_tf, spike_index_tf,
channel_index, waveform_length)
return x_tf, spike_index_tf, probability_tf, waveform_tf, vars_dict
def train_triage(x_train, y_train, n_filters, n_iter, n_batch, l2_reg_scale,
train_step_size, nn_name):
"""
Trains the triage network
Parameters:
-----------
x_train: np.array
[number of data, temporal length, number of channels] training data
for the triage network.
y_train: np.array
[number of data] training label for the triage network.
nn_name: string
name of the .ckpt to be saved.
"""
# get parameters
ndata, R, C = x_train.shape
K1, K2 = n_filters
# x and y input tensors
x_tf = tf.placeholder("float", [n_batch, R, C])
y_tf = tf.placeholder("float", [n_batch])
# first layer: temporal feature
W1 = weight_variable([R, 1, 1, K1])
b1 = bias_variable([K1])
layer1 = tf.nn.relu(conv2d_VALID(tf.expand_dims(x_tf, -1), W1) + b1)
# second layer: feataure mapping
W11 = weight_variable([1, 1, K1, K2])
b11 = bias_variable([K2])
layer11 = tf.nn.relu(conv2d(layer1, W11) + b11)
# third layer: spatial convolution
W2 = weight_variable([1, C, K2, 1])
b2 = bias_variable([1])
o_layer = tf.squeeze(conv2d_VALID(layer11, W2) + b2)
# cross entropy
cross_entropy = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=o_layer, labels=y_tf))
# regularization term
weights = tf.trainable_variables()
l2_regularizer = tf.contrib.layers.l2_regularizer(scale=l2_reg_scale)
regularization_penalty = tf.contrib.layers.apply_regularization(
l2_regularizer, weights)
regularized_loss = cross_entropy + regularization_penalty
# train step
train_step = tf.train.AdamOptimizer(train_step_size).minimize(
regularized_loss)
# saver
saver = tf.train.Saver({
"W1": W1,
"W11": W11,
"W2": W2,
"b1": b1,
"b11": b11,
"b2": b2
})
############
# training #
############
bar = progressbar.ProgressBar(maxval=n_iter)
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
for i in range(0, n_iter):
idx_batch = np.random.choice(ndata, n_batch, replace=False)
sess.run(train_step,
feed_dict={
x_tf: x_train[idx_batch],
y_tf: y_train[idx_batch]
})
bar.update(i + 1)
saver.save(sess, nn_name)
idx_batch = np.random.choice(ndata, n_batch, replace=False)
output = sess.run(o_layer, feed_dict={x_tf: x_train[idx_batch]})
y_test = y_train[idx_batch]
tp = np.mean(output[y_test == 1] > 0)
fp = np.mean(output[y_test == 0] > 0)
print('Approximate training true positive rate: ' + str(tp) +
', false positive rate: ' + str(fp))
bar.finish()
