def anneal_learning_rate(initial_learning_rate, normaliser, iteration):
return sfloatX(initial_learning_rate / (1 + (iteration / normaliser)))
from neuralnilm.objectives import mdn_nll
from neuralnilm.utils import sfloatX
from neuralnilm.updates import anneal_learning_rate
from neuralnilm.plot import MDNPlotter
# Number of units in the hidden (recurrent) layer
N_HIDDEN_LAYERS = 2
N_UNITS_PER_LAYER = 25
N_COMPONENTS = 2
# Number of training sequences in each batch
N_SEQ_PER_BATCH = 16
SEQ_LENGTH = 256
SHAPE = (N_SEQ_PER_BATCH, SEQ_LENGTH, 1)
# SGD learning rate
INITIAL_LEARNING_RATE = sfloatX(5e-4)
# LEARNING_RATE_NORMALISER = 5
learning_rate = theano.shared(INITIAL_LEARNING_RATE, name='learning_rate')
LEARNING_RATE_CHANGES = {
500: 1e-04,
1000: 5e-05,
2000: 1e-05,
3000: 5e-06,
4000: 1e-06,
10000: 5e-07,
50000: 1e-07
}
# Number of iterations to train the net
N_ITERATIONS = 100
VALIDATE = False
n_seq_per_batch = error.shape[1]
# n_seq_per_batch = 16 # CHANGE THIS!
for seq_i in range(n_seq_per_batch):
elements_above_thresh = t[seq_i, :, 0] > THRESHOLD
n_above_thresh = elements_above_thresh.sum()
if n_above_thresh == 0:
error[seq_i, :, :] = 0
else:
error[seq_i, elements_above_thresh, 0] *= 0.5 / n_above_thresh
n_below_thresh = seq_length - n_above_thresh
error[seq_i, -elements_above_thresh, 0] *= 0.5 / n_below_thresh
return error.sum()
TWO_PI = sfloatX(2 * np.pi)
def mdn_nll(theta, t):
"""Computes the mean of negative log likelihood for P(t|theta) for
Mixture Density Network output layers.
:parameters:
- theta : Output of the net. Contains mu, sigma, mixing
- t : targets. Shape = (minibatch_size, output_size)
:returns:
- NLL per output
"""
# Adapted from NLL() in
# github.com/aalmah/ift6266amjad/blob/master/experiments/mdn.py
def anneal_learning_rate(initial_learning_rate, normaliser, iteration):
return sfloatX(initial_learning_rate / (1 + (iteration / normaliser)))
from neuralnilm.layers import MixtureDensityLayer
from neuralnilm.objectives import mdn_nll
from neuralnilm.utils import sfloatX
from neuralnilm.updates import anneal_learning_rate
from neuralnilm.plot import MDNPlotter
# Number of units in the hidden (recurrent) layer
N_HIDDEN_LAYERS = 2
N_UNITS_PER_LAYER = 25
N_COMPONENTS = 2
# Number of training sequences in each batch
N_SEQ_PER_BATCH = 16
SEQ_LENGTH = 256
SHAPE = (N_SEQ_PER_BATCH, SEQ_LENGTH, 1)
# SGD learning rate
INITIAL_LEARNING_RATE = sfloatX(5e-4)
# LEARNING_RATE_NORMALISER = 5
learning_rate = theano.shared(INITIAL_LEARNING_RATE, name='learning_rate')
LEARNING_RATE_CHANGES = {
500: 1e-04,
1000: 5e-05,
2000: 1e-05,
3000: 5e-06,
4000: 1e-06,
10000: 5e-07,
50000: 1e-07
}
# Number of iterations to train the net
N_ITERATIONS = 100
VALIDATE = False