def chop_with_stride(x, y, size, stride, rand_shift=True):
assert isinstance(x, np.ndarray) and isinstance(y, np.ndarray)
checker.check_type([size, stride], int)
out_len = SignalSet.chop_with_stride_len_f(len(x), size, stride)
x_out = np.zeros(shape=(out_len, size))
y_out = np.zeros(shape=(out_len, *y.shape[1:]))
if rand_shift:
remain = len(x) - ((out_len - 1) * stride + size)
shift = np.random.randint(remain + 1)
else:
shift = 0
for i in range(out_len):
# Fill in x
x_out[i] = x[shift + stride * i:shift + stride * i + size]
# Fill in y if necessary
if len(x) == len(y):
y_out[i] = y[shift + stride * i:shift + stride * i + size]
if len(x) != len(y):
assert len(y) == 1
y = np.tile(y, (out_len, 1))
return x_out, y
def _check_signals(signals, responses):
# Check signals
if isinstance(signals, Signal): signals = [signals]
checker.check_type(signals, Signal)
signal_dict = {}
# Make sure all signals are sampled under a same fs
fs = signals[0].fs
for i in range(1, len(signals)):
if fs != signals[i].fs:
raise ValueError(
'!! All signals in SignalSet must have the same sampling frequency'
)
signal_dict[pedia.signals] = signals
# Check responses
if responses is not None:
if isinstance(responses, Signal): responses = [responses]
if len(signals) != len(responses):
raise ValueError(
'!! length of responses({}) does not match that of signals({})'
.format(len(responses), len(signals)))
checker.check_type(responses, Signal)
for r in responses:
if r.fs != fs:
raise ValueError(
'!! All responses must have the same sampling frequency with signals'
)
signal_dict[pedia.responses] = responses
# Return signal dict and fs
return signal_dict, fs
def __init__(self,
output_dim,
memory_units=None,
mem_config=None,
use_mem_wisely=False,
weight_regularizer=None,
**kwargs):
# Call parent's constructor
RNet.__init__(self, self.net_name)
# Attributes
self.output_dim = output_dim
self.memory_units = (self.MemoryUnit.parse_units(mem_config)
if memory_units is None else memory_units)
self.memory_units = [mu for mu in self.memory_units if mu.size > 0]
checker.check_type(self.memory_units, Ham.MemoryUnit)
self._state_size = sum([mu.size for mu in self.memory_units])
self._activation = activations.get('tanh', **kwargs)
self._use_mem_wisely = use_mem_wisely
self._truncate = kwargs.get('truncate', False)
self._weight_regularizer = regularizers.get(weight_regularizer,
**kwargs)
# self._use_global_reg = kwargs.get('global_reg', False)
self._kwargs = kwargs
def __init__(self,
kernel_key,
num_neurons,
input_,
suffix,
weight_initializer='glorot_normal',
prune_frac=0,
LN=False,
gain_initializer='ones',
etch=None,
weight_dropout=0.0,
**kwargs):
# Call parent's initializer
super().__init__(kernel_key,
num_neurons,
weight_initializer,
prune_frac,
etch=etch,
weight_dropout=weight_dropout,
**kwargs)
self.input_ = checker.check_type(input_, tf.Tensor)
self.suffix = checker.check_type(suffix, str)
self.LN = checker.check_type(LN, bool)
self.gain_initializer = initializers.get(gain_initializer)
def __init__(self,
configs,
activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
bias_initializer='zeros',
reverse=False,
use_reset_gate=False,
dropout=0.0,
layer_normalization=False,
**kwargs):
"""
:param configs: a list or tuple of tuples with format (size, num, delta)
or a string with format `S1xM1xD1+S2xM2xD2+...`
"""
# Call parent's constructor
CellBase.__init__(self, activation, weight_initializer, use_bias,
bias_initializer, layer_normalization, **kwargs)
# Specific attributes
self._reverse = checker.check_type(reverse, bool)
self._use_reset_gate = checker.check_type(use_reset_gate, bool)
self._groups = self._get_groups(configs)
self._state_size = self._get_total_size(self._groups)
self._dropout_rate = checker.check_type(dropout, float)
assert 0 <= dropout < 1
# matrices for SOG v1
self._D = None
self._S = None
def __init__(self,
configs,
activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
bias_initializer='zeros',
reverse=False,
use_reset_gate=False,
reset_who='s',
shunt_output=False,
gate_output=False,
**kwargs):
"""
:param configs: a list or tuple of tuples with format (size, num, delta)
or a string with format `S1xM1xD1+S2xM2xD2+...`
"""
# Call parent's constructor
CellBase.__init__(self, activation, weight_initializer, use_bias,
bias_initializer, **kwargs)
# Specific attributes
self._reverse = checker.check_type(reverse, bool)
self._use_reset_gate = checker.check_type(use_reset_gate, bool)
self._shunt_output = checker.check_type(shunt_output, bool)
self._gate_output = checker.check_type(gate_output, bool)
self._groups = self._get_groups(configs)
self._state_size = self._get_total_size(self._groups)
assert reset_who in ('a', 's')
self._reset_who = reset_who
def __init__(self,
state_size,
use_reset_gate=True,
activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
bias_initializer='zeros',
z_bias_initializer='zeros',
reset_who='s',
dropout=0.0,
zoneout=0.0,
**kwargs):
"""
:param reset_who: in ('x', 'y')
'x': a_h = W_h * (h_{t-1} \odot r_t)
'y': a_h = r_t \odot (W_h * h_{t-1})
\hat{h}_t = \varphi(Wx*x + a_h + b)
in which r_t is the reset gate at time step t,
\odot is the Hadamard product, W_h is the hidden-to-hidden matrix
"""
# Call parent's constructor
CellBase.__init__(self, activation, weight_initializer, use_bias,
bias_initializer, **kwargs)
# Specific attributes
self._state_size = checker.check_positive_integer(state_size)
self._use_reset_gate = checker.check_type(use_reset_gate, bool)
self._z_bias_initializer = initializers.get(z_bias_initializer)
self._dropout_rate = checker.check_type(dropout, float)
self._zoneout_rate = checker.check_type(zoneout, float)
assert reset_who in ('s', 'a')
self._reset_who = reset_who
def __init__(
self,
state_size,
activation='tanh',
use_bias=True,
weight_initializer='xavier_uniform',
bias_initializer='zeros',
input_gate=True,
output_gate=True,
forget_gate=True,
with_peepholes=False,
**kwargs):
"""
:param state_size: state size: positive int
:param activation: activation: string or callable
:param use_bias: whether to use bias
:param weight_initializer: weight initializer identifier
:param bias_initializer: bias initializer identifier
"""
# Call parent's constructor
RNet.__init__(self, BasicLSTMCell.net_name)
# Attributes
self._state_size = state_size
self._activation = activations.get(activation, **kwargs)
self._use_bias = checker.check_type(use_bias, bool)
self._weight_initializer = initializers.get(weight_initializer)
self._bias_initializer = initializers.get(bias_initializer)
self._input_gate = checker.check_type(input_gate, bool)
self._output_gate = checker.check_type(output_gate, bool)
self._forget_gate = checker.check_type(forget_gate, bool)
self._with_peepholes = checker.check_type(with_peepholes, bool)
self._output_scale = state_size
def _gen_rnn_batches(self, x, y, batch_size, num_steps):
checker.check_positive_integer(batch_size, 'batch size')
checker.check_type(num_steps, int)
# Get batch partitions
data_x, L = self._get_batch_partition(x, batch_size)
if y is not None:
if len(x) == len(y):
data_y, Ly = self._get_batch_partition(y, batch_size)
assert L == Ly
else:
assert len(y) == 1
data_y = y
# Chop data further
if num_steps < 0: num_steps = L
round_len = int(np.ceil(L / num_steps))
for i in range(round_len):
batch_x = data_x[:, i * num_steps:min((i + 1) * num_steps, L)]
batch_y = None
if y is not None:
if len(x) == len(y):
batch_y = data_y[:,
i * num_steps:min((i + 1) * num_steps, L)]
else:
assert isinstance(y, np.ndarray)
batch_y = np.tile(y,
[batch_x.shape[0], batch_x.shape[1], 1])
batch = DataSet(batch_x, batch_y, in_rnn_format=True)
# State should be reset at the beginning of a sequence
if i == 0: batch.should_reset_state = True
batch.name = self.name + '_{}'.format(i + 1)
yield batch
def get_data_batches(self,
data_set,
batch_size,
num_steps=None,
shuffle=False):
""" Get batch generator.
:param data_set: an instance of DataSet or BigData from which data batches
will be extracted
:param batch_size: if is None, default value will be assigned according to
the input type of this model
:param num_steps: step number for RNN data batches
:param shuffle: whether to shuffle
:return: a generator or a list
"""
# Data set must be an instance of DataSet or BigData
assert isinstance(data_set, (DataSet, BigData))
if self.input_type is InputTypes.BATCH:
# If model's input type is normal batch, num_steps will be ignored
# If batch size is not specified and data is a DataSet, feed it all at
# once into model
if batch_size is None and isinstance(data_set, DataSet):
return [data_set.stack]
checker.check_positive_integer(batch_size)
data_batches = data_set.gen_batches(batch_size, shuffle=shuffle)
elif self.input_type is InputTypes.RNN_BATCH:
if batch_size is None: batch_size = 1
if num_steps is None: num_steps = -1
checker.check_positive_integer(batch_size)
checker.check_type(num_steps, int)
data_batches = data_set.gen_rnn_batches(batch_size, num_steps,
shuffle)
else:
raise ValueError('!! Can not resolve input type of this model')
return data_batches
def get_round_length(self, batch_size, num_steps=None):
"""Get round length for training
:param batch_size: Batch size. For irregular sequences, this value should
be set to 1.
:param num_steps: Step number. If provided, round length will be calculated
for RNN model
:return: Round length for training
"""
# Make sure features exist
self._check_feature()
checker.check_positive_integer(batch_size, 'batch_size')
if num_steps is None:
# :: For feed-forward models
return int(np.ceil(self.stack.size / batch_size))
else:
# :: For recurrent models
checker.check_type(num_steps, int)
if self.is_regular_array: arrays = [self.features]
elif self.parallel_on:
return self._get_pe_round_length(batch_size, num_steps)
else:
arrays = self.features
len_f = lambda x: x if self.len_f is None else self.len_f
if num_steps < 0: return len(arrays)
else:
return int(
sum([
np.ceil(len_f(len(array)) // batch_size / num_steps)
for array in arrays
]))
def __init__(self,
fast_size,
fast_layers,
slow_size,
hyper_kernel,
activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
bias_initializer='zeros',
input_dropout=0.0,
output_dropout=0.0,
forget_bias=0,
**kwargs):
# Call parent's constructor
CellBase.__init__(self, activation, weight_initializer, use_bias,
bias_initializer, **kwargs)
self.kernel_key = checker.check_type(hyper_kernel, str)
# Specific attributes
self._fast_size = checker.check_positive_integer(fast_size)
self._fast_layers = checker.check_positive_integer(fast_layers)
self._slow_size = checker.check_positive_integer(slow_size)
self._hyper_kernel = self._get_hyper_kernel(hyper_kernel,
do=th.rec_dropout,
forget_bias=forget_bias)
self._input_do = checker.check_type(input_dropout, float)
self._output_do = checker.check_type(output_dropout, float)
def __init__(self,
configs,
factoring_dim=None,
psi_config=None,
activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
bias_initializer='zeros',
reverse=False,
**kwargs):
"""
:param psi_config: e.g. 's:xs+g;xs', 's:x'
"""
# Call parent's constructor
CellBase.__init__(self, activation, weight_initializer, use_bias,
bias_initializer, **kwargs)
# Specific attributes
self._reverse = checker.check_type(reverse, bool)
self._groups = self._get_groups(configs)
self._state_size = self._get_total_size(self._groups)
if factoring_dim is None: factoring_dim = self._state_size
self._fd = checker.check_positive_integer(factoring_dim)
if not psi_config: psi_config = 's:x'
self._psi_string = checker.check_type(psi_config, str)
self._psi_config = self._parse_psi_string()
def __init__(
self,
state_size,
activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
couple_fi=False,
cell_bias_initializer='zeros',
input_bias_initializer='zeros',
output_bias_initializer='zeros',
forget_bias_initializer='zeros',
use_output_activation=True,
**kwargs):
# Call parent's constructor
CellBase.__init__(self, activation, weight_initializer,
use_bias, cell_bias_initializer, **kwargs)
# Specific attributes
self._state_size = checker.check_positive_integer(state_size)
self._input_bias_initializer = initializers.get(input_bias_initializer)
self._output_bias_initializer = initializers.get(output_bias_initializer)
self._forget_bias_initializer = initializers.get(forget_bias_initializer)
self._couple_fi = checker.check_type(couple_fi, bool)
self._use_output_activation = checker.check_type(
use_output_activation, bool)
def load_data(path, memory_depth=1, validate_size=5000, test_size=88000):
data_sets = WHBM.load(path,
validate_size=validate_size,
test_size=test_size,
memory_depth=memory_depth,
skip_head=True)
checker.check_type(data_sets, SignalSet)
return data_sets
def __call__(self, *input_list):
"""Link neuron array to graph
:param input_list: inputs to be fully connected
:return: neuron outputs
"""
input_list = [x for x in input_list if x is not None]
# Add inputs
if input_list:
checker.check_type(input_list, tf.Tensor)
# Concatenation is forbidden when LN is on
if all([
len(input_list) > 1, self._layer_normalization,
self._normalize_each_psi
]):
for x in input_list:
self.add_kernel(x)
else:
# Concatenate to speed up calculation if necessary
if len(input_list) > 1: x = tf.concat(input_list, axis=-1)
else: x = input_list[0]
# Add kernel
self.add_kernel(x)
# Make sure psi_kernel is not empty
assert self.psi_kernels
# Link
with tf.variable_scope(self.scope):
# Calculate summed input a, ref: Ba, etc. Layer Normalization, 2016
a_list = [kernel() for kernel in self.psi_kernels]
a = a_list[0] if len(a_list) == 1 else tf.add_n(
a_list, 'summed_inputs')
# Do layer normalization here if necessary
if self._layer_normalization:
if not self._normalize_each_psi:
a = PsiKernel.layer_normalization(a,
self._gain_initializer,
False)
# If LN is on, use_bias option must be True
self._use_bias = True
# Add bias if necessary
if self._use_bias:
if self.bias_kernel is None: self.register_bias_kernel()
bias = self.bias_kernel()
# Some kernels may generate bias of shape [batch_size, num_neurons]
if len(bias.shape) == 1:
a = tf.nn.bias_add(a, self.bias_kernel())
else:
a = a + bias
# Activate if necessary
if self._activation: a = self._activation(a)
return a
def _check_data(self):
"""data_dict should be a non-empty dictionary containing equilength lists of
regular numpy arrays. Samples in the same sequence list must have the
same shape
summ_dict should be a dictionary which stores summaries of each sequence.
"""
# Check data_dict and summ_dict
if not isinstance(self.data_dict, dict) or len(self.data_dict) == 0:
raise TypeError('!! data_dict must be a non-empty dictionary')
if not isinstance(self.summ_dict, dict):
raise TypeError('!! summ_dict must be a dictionary')
list_length = len(list(self.data_dict.values())[0])
# Check each item in data_dict
for name, seq_list in self.data_dict.items():
checker.check_type(seq_list, np.ndarray)
# Check type and length
if not isinstance(seq_list, list) or len(seq_list) != list_length:
raise ValueError('!! {} should be a list with length {}'.format(
name, list_length))
# Check structure
if [len(s) for s in seq_list] != self.structure:
raise ValueError('!! sequence list structure inconformity: {} '.format(
name))
# Make sure len(sample_shape) > 0
if len(seq_list[0].shape) < 2:
seq_list = [s.reshape(-1, 1) for s in seq_list]
# Check sample shape
shapes = [s.shape[1:] for s in seq_list]
if shapes.count(shapes[0]) != len(shapes):
raise AssertionError('!! Sample shape in {} are inconformity'.format(
name))
self.data_dict[name] = seq_list
# Check each item in summ_dict
for name, summ_list in self.summ_dict.items():
# Check type and length
if not isinstance(summ_list, list) or len(summ_list) != list_length:
raise ValueError('!! {} should be a list of length {}'.format(
name, list_length))
if checker.check_scalar_list(summ_list): continue
checker.check_type(summ_list, np.ndarray)
# Check structure
for i, summ in enumerate(summ_list):
if summ.shape[0] > 1: summ_list[i] = np.reshape(summ, (1, *summ.shape))
# Check sample shape
shapes = [s.shape[1:] for s in summ_list]
if shapes.count(shapes[0]) != len(shapes):
raise AssertionError('!! Sample shape in {} are inconformity'.format(
name))
def __init__(
self,
temporal_configs,
output_size=None,
spatial_configs=None,
temporal_reverse=False,
spatial_reverse=False,
temporal_activation='tanh',
spatial_activation='tanh',
weight_initializer='xavier_normal',
use_bias=True,
bias_initializer='zeros',
**kwargs):
"""
:param output_size:
Denote y as cell output, s as state
(1) output_size is 0 or None
y = new_s
(2) output_size is not None
y = neuron(x, prev_s, ...)
"""
# Call parent's constructor
CellBase.__init__(self, temporal_activation, weight_initializer,
use_bias, bias_initializer, **kwargs)
self._temporal_activation = self._activation
self._spatial_activation = spatial_activation
# Specific attributes
self._temporal_groups = self._get_groups(temporal_configs)
self._state_size = self._get_total_size(self._temporal_groups)
# Set spatial groups
self._output_size = None if output_size == 0 else output_size
self._spatial_groups = []
if spatial_configs is not None:
output_dim = (self._output_size if self._output_size is not None
else self._state_size)
# Set spatial_groups
if spatial_configs == 'default':
# Check output size
num_groups = output_dim // 2
assert num_groups * 2 == output_dim
self._spatial_groups = [(2, num_groups, 1)]
else:
assert isinstance(spatial_configs, str) and len(spatial_configs) > 0
self._spatial_groups = self._get_groups(spatial_configs)
total_size = self._get_total_size(self._spatial_groups)
# Check output dim
assert output_dim == total_size
self._reverse_t = checker.check_type(temporal_reverse, bool)
self._reverse_s = checker.check_type(spatial_reverse, bool)
def _get_periods(self, periods, **kwargs):
# Get max groups
max_groups = kwargs.get('max_groups', 7)
if periods is None:
periods = []
i = 0
for _ in range(self._state_size):
periods.append(2**i)
i += 1
if i >= max_groups: i = 0
else: checker.check_type(periods, int)
assert len(periods) == self._state_size
return sorted(periods)
def _brutal_chop_len_f(self, bs, ns, sz):
assert isinstance(self, BigData)
round_len = 0
assert ns is not None
for len_list in self.structure:
checker.check_type(len_list, int)
# For RNN models
if ns < 0: round_len += len(len_list)
else:
round_len += int(
sum([np.ceil(size // sz // bs / ns) for size in len_list]))
# Return round length
return round_len
def _get_file_name(cls, train_size, test_size, unique_, cheat,
local_binary, multiple, rule):
checker.check_positive_integer(train_size)
checker.check_positive_integer(test_size)
checker.check_positive_integer(multiple)
checker.check_type(unique_, bool)
if rule is not None: tail = rule
elif unique_: tail = 'U'
else: tail = 'NU'
file_name = '{}{}_{}+{}_{}_{}_{}.tfds'.format(
cls.DATA_NAME, '' if multiple == 1 else '(x{})'.format(multiple),
train_size, test_size, tail, 'C' if cheat else 'NC',
'LB' if local_binary else 'P')
if multiple > 1: file_name = 'm' + file_name
return file_name
def _pad_sequences(sequences, max_steps):
"""Receive a list of irregular sequences and output a regular numpy array"""
assert isinstance(sequences, list)
checker.check_positive_integer(max_steps)
checker.check_type(sequences, np.ndarray)
if all([s.shape[0] == sequences[0].shape[0] for s in sequences]):
return np.stack(sequences, axis=0)
sample_shape = sequences[0].shape[1:]
assert len(sample_shape) > 0
stack = np.zeros(shape=(len(sequences), max_steps, *sample_shape))
for i, s in enumerate(sequences): stack[i, :len(s)] = s
return stack
def _check_data(self):
"""Features and data_dict should not be empty at the same time.
All data array or list provided must have the same length.
If features (or targets) are provided as a list (or a tuple),
its elements must be numpy arrays with exactly the same shape (except
for the first dimension)."""
# Make sure data_dict is a dictionary
if not isinstance(self.data_dict, dict):
raise TypeError('!! data_dict provided must be a dict')
# Put all data arrays to a single dict for later check
data_dict = self.data_dict.copy()
if self.features is not None:
data_dict[pedia.features] = self.features
if self.targets is not None:
# TODO
# if type(self.features) != type(self.targets):
# raise TypeError('!! features and targets must be of the same type')
data_dict[pedia.targets] = self.targets
# Make sure at least one data array is provided
if len(data_dict) == 0:
raise AssertionError('!! data not found')
# Make sure all data array have the same size
size = -1
for key, val in data_dict.items():
# Make sure all data arrays are instances of list or ndarray or sth.
if not hasattr(val, '__len__'):
raise AttributeError(
'!! {} data must have __len__ attribute'.format(key))
if size == -1: size = len(val)
elif size != len(val):
raise ValueError('!! all data array must have the same size')
# Make sure features and targets are (lists of) numpy arrays
if key in (pedia.features, pedia.targets):
checker.check_type(val, np.ndarray)
# If features and targets are stored in a list (or a tuple), check
# .. the shape of each numpy array
if not isinstance(val, np.ndarray):
assert isinstance(val, (list, tuple))
shape = None
for array in val:
assert isinstance(array, np.ndarray)
if shape is None: shape = array.shape[1:]
elif shape != array.shape[1:]:
raise ValueError(
'!! samples in {} list should have the same shape'
.format(key))
def __init__(
self,
state_size,
activation='tanh',
use_bias=True,
weight_initializer='xavier_normal',
bias_initializer='zeros',
**kwargs):
"""
:param state_size: state size: positive int
:param activation: activation: string or callable
:param use_bias: whether to use bias
:param weight_initializer: weight initializer identifier
:param bias_initializer: bias initializer identifier
"""
# Call parent's constructor
RNet.__init__(self, self.net_name)
# Attributes
self._state_size = state_size
self._activation = activations.get(activation, **kwargs)
self._use_bias = checker.check_type(use_bias, bool)
self._weight_initializer = initializers.get(weight_initializer)
self._bias_initializer = initializers.get(bias_initializer)
self._output_scale = state_size
def __init__(self,
state_size,
periods=None,
activation='tanh',
use_bias=True,
weight_initializer='xavier_uniform',
bias_initializer='zeros',
**kwargs):
"""
:param state_size: State size
:param periods: a list of integers. If not provided, periods will be set
to a default exponential series {2^{i-1}}_{i=0}^{state_size}
"""
# Call parent's constructor
RNet.__init__(self, ClockworkRNN.net_name)
# Attributes
self._state_size = checker.check_positive_integer(state_size)
self._periods = self._get_periods(periods, **kwargs)
self._activation = activations.get(activation, **kwargs)
self._use_bias = checker.check_type(use_bias, bool)
self._weight_initializer = initializers.get(weight_initializer)
self._bias_initializer = initializers.get(bias_initializer)
# modules = [(start_index, size, period)+]
self._modules = []
self._init_modules(**kwargs)
def __init__(self,
num_neurons,
group_size,
head_size=-1,
activation=None,
use_bias=True,
weight_initializer='xavier_normal',
bias_initializer='zeros',
**kwargs):
"""
Softmax over groups applied to neurons.
Case 1: head_size < 0: does not use extra neurons
Case 2: head_size = 0: use extra neurons without a head
Case 3: head_size > 0: use extra neurons with a head
"""
# Call parent's constructor
super().__init__(activation, weight_initializer, use_bias,
bias_initializer, **kwargs)
# Specific attributes
self._num_neurons = checker.check_positive_integer(num_neurons)
self._group_size = checker.check_positive_integer(group_size)
self._head_size = checker.check_type(head_size, int)
# Developer options
options = th.developer_options
def __init__(
self,
num_filter_list,
num_classes,
kernel_initializer='glorot_uniform',
left_repeats=2,
right_repeats=2,
activation='relu',
dropout_rate=0.5,
name='unet',
level=1,
**kwargs):
# Sanity check
assert isinstance(num_filter_list, (tuple, list)) and num_filter_list
# Call parent's constructor
# TODO: the level logic is not elegant
super().__init__(name, level=level, **kwargs)
# Specific attributes
self.num_filter_list = num_filter_list
self.num_classes = checker.check_positive_integer(num_classes)
# self.kernel_sizes = kernel_sizes
self.kernel_initializer = kernel_initializer
self.activation = activation
self.dropout_rate = checker.check_type(dropout_rate, float)
self.left_repeats = checker.check_positive_integer(left_repeats)
self.right_repeats = checker.check_positive_integer(right_repeats)
# Add layers
self._add_layers()
def __init__(self,
output_dim=None,
spatial_configs=None,
reverse=False,
activation='relu',
use_bias=True,
weight_initializer='xavier_normal',
bias_initializer='zeros',
**kwargs):
assert isinstance(activation, str)
self.activation_string = activation
# Call parent's constructor
LayerWithNeurons.__init__(self, activation, weight_initializer,
use_bias, bias_initializer, **kwargs)
assert not (output_dim is None and spatial_configs is None)
self._spatial_groups = []
if spatial_configs is not None:
self._spatial_groups = self._get_groups(spatial_configs)
total_size = self._get_total_size(self._spatial_groups)
if output_dim is None: output_dim = total_size
assert output_dim == total_size
self._output_dim = checker.check_positive_integer(output_dim)
self._reverse = checker.check_type(reverse, bool)
self.neuron_scale = [output_dim]
def saturate_loss(tensor, mu=0.5, encourage_saturation=True):
# TODO: this method is still being developed. DO NOT USE WITHOUT GUIDE
# Each entry in tensor should be in range [0, 1], which should be guaranteed
# ... by users
checker.check_type(tensor, tf.Tensor)
assert 0 < mu < 1
# Encourage saturation
if encourage_saturation:
# Calculate distance to saturation
left = tensor[tf.less(tensor, mu)]
right = tensor[tf.greater(tensor, 0.5)]
return tf.norm(left) - tf.norm(right)
else:
# Calculate distance to unsaturation
degree = tf.abs(tensor - mu)
# Calculate loss using reduce mean
return tf.reduce_mean(degree)
def evaluate(f, data_set, plot=False):
if not callable(f): raise AssertionError('!! Input f must be callable')
checker.check_type(data_set, SignalSet)
assert isinstance(data_set, SignalSet)
if data_set.targets is None:
raise ValueError('!! Responses not found in SignalSet')
u, y = data_set.features, np.ravel(data_set.targets)
assert isinstance(y, Signal)
# Show status
console.show_status('Evaluating {} ...'.format(data_set.name))
# In evaluation, the sum of each metric is started at t = 1000 instead of
# t = 0 to eliminate the influence of transient errors at the beginning of
# the simulation
start_at = 1000
model_output = Signal(f(u), fs=y.fs)
delta = y - model_output
err = delta[start_at:]
assert isinstance(err, Signal)
ratio = lambda val: 100.0 * val / y.rms
# The mean value of the simulation error in time domain
val = err.average
console.supplement('E[err] = {:.4f}mV ({:.3f}%)'.format(
val * 1000, ratio(val)))
# The standard deviation of the error in time domain
val = float(np.std(err))
console.supplement('STD[err] = {:.4f}mV ({:.3f}%)'.format(
val * 1000, ratio(val)))
# The root mean square value of the error in time domain
val = err.rms
console.supplement('RMS[err] = {:.4f}mV ({:.3f}%)'.format(
val * 1000, ratio(val)))
# Plot
if not plot: return
from tframe.data.sequences.signals.figure import Figure, Subplot
fig = Figure('Simulation Error')
# Add ground truth
prefix = 'System Output, $||y|| = {:.4f}$'.format(y.norm)
fig.add(Subplot.PowerSpectrum(y, prefix=prefix))
# Add model output
prefix = 'Model Output, RMS($\Delta$) = ${:.4f}mV$'.format(1000 * err.rms)
fig.add(Subplot.PowerSpectrum(model_output, prefix=prefix, Error=delta))
# Plot
fig.plot()