def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None, outdir='outputs/lstm/',
input_size=None, hidden_size=None, output_size=None,
activation='sigmoid', hidden_activation='relu', inner_hidden_activation='sigmoid',
mrg=RNG_MRG.MRG_RandomStreams(1),
weights_init='uniform', weights_interval='montreal', weights_mean=0, weights_std=5e-3,
bias_init=0.0,
r_weights_init='identity', r_weights_interval='montreal', r_weights_mean=0, r_weights_std=5e-3,
r_bias_init=0.0,
cost_function='mse', cost_args=None,
noise='dropout', noise_level=None, noise_decay=False, noise_decay_amount=.99,
direction='forward',
clip_recurrent_grads=False):
"""
Initialize a simple recurrent network.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere. For recurrent nets,
this will be the initial starting value for hidden layers.
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters.
outdir : str
The location to produce outputs from training or running the :class:`RNN`. If None, nothing will be saved.
input_size : int
The size (dimensionality) of the input. If shape is provided in `inputs_hook`, this is optional.
hidden_size : int
The size (dimensionality) of the hidden layers. If shape is provided in `hiddens_hook`, this is optional.
output_size : int
The size (dimensionality) of the output.
activation : str or callable
The nonlinear (or linear) activation to perform after the dot product from hiddens -> output layer.
This activation function should be appropriate for the output unit types, i.e. 'sigmoid' for binary.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The activation to perform for the hidden units.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
inner_hidden_activation : str or callable
The activation to perform for the hidden gates.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
mrg : random
A random number generator that is used when adding noise.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
r_weights_init : str
Determines the method for initializing recurrent model weights. See opendeep.utils.nnet for options.
r_weights_interval : str or float
If Uniform `r_weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
r_weights_mean : float
If Gaussian `r_weights_init`, the mean value to use.
r_weights_std : float
If Gaussian `r_weights_init`, the standard deviation to use.
r_bias_init : float
The initial value to use for the recurrent bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the output cost of the model.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
noise : str
What type of noise to use for the hidden layers and outputs. See opendeep.utils.noise
for options. This should be appropriate for the unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
noise_decay : str or False
Whether to use `noise` scheduling (decay `noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the model learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_decay_amount : float
The amount to reduce the `noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
direction : str
The direction this recurrent model should go over its inputs.
Can be 'forward', 'backward', or 'bidirectional'.
clip_recurrent_grads : False or float, optional
Whether to clip the gradients for the parameters that unroll over timesteps (such as the weights
connecting previous hidden states to the current hidden state, and not the weights from current
input to hiddens). If it is a float, the gradients for the weights will be hard clipped to the range
`+-clip_recurrent_grads`.
"""
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(LSTM, self).__init__(**initial_parameters)
##################
# specifications #
##################
backward = direction.lower() == 'backward'
bidirectional = direction.lower() == 'bidirectional'
#########################################
# activation, cost, and noise functions #
#########################################
# recurrent hidden activation function!
self.hidden_activation_func = get_activation_function(hidden_activation)
self.inner_hidden_activation_func = get_activation_function(inner_hidden_activation)
# output activation function!
activation_func = get_activation_function(activation)
# Cost function
cost_function = get_cost_function(cost_function)
cost_args = cost_args or dict()
# Now deal with noise if we added it:
if noise:
log.debug('Adding %s noise switch.' % str(noise))
if noise_level is not None:
noise_level = sharedX(value=noise_level)
noise_func = get_noise(noise, noise_level=noise_level, mrg=mrg)
else:
noise_func = get_noise(noise, mrg=mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.noise_switch = sharedX(value=1, name="basiclayer_noise_switch")
# noise scheduling
if noise_decay and noise_level is not None:
self.noise_schedule = get_decay_function(noise_decay,
noise_level,
noise_level.get_value(),
noise_decay_amount)
###############
# inputs hook #
###############
# grab info from the inputs_hook
# in the case of an inputs_hook, recurrent will always work with the leading tensor dimension
# being the temporal dimension.
# input is 3D tensor of (timesteps, batch_size, data_dim)
# if input is 2D tensor, assume it is of the form (timesteps, data_dim) i.e. batch_size is 1. Convert to 3D.
# if input is > 3D tensor, assume it is of form (timesteps, batch_size, data...) and flatten to 3D.
if self.inputs_hook is not None:
self.input = self.inputs_hook[1]
if self.input.ndim == 1:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 'x'), [1, 2])
self.input_size = 1
elif self.input.ndim == 2:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 1), 1)
elif self.input.ndim > 3:
self.input = self.input.flatten(3)
self.input_size = sum(self.input_size)
else:
raise NotImplementedError("Recurrent input with %d dimensions not supported!" % self.input.ndim)
xs = self.input
else:
# Assume input coming from optimizer is (batches, timesteps, data)
# so, we need to reshape to (timesteps, batches, data)
self.input = T.tensor3("Xs")
xs = self.input.dimshuffle(1, 0, 2)
# The target outputs for supervised training - in the form of (batches, timesteps, output) which is
# the same dimension ordering as the expected input from optimizer.
# therefore, we need to swap it like we did to input xs.
self.target = T.tensor3("Ys")
ys = self.target.dimshuffle(1, 0, 2)
################
# hiddens hook #
################
# set an initial value for the recurrent hiddens from hook
if self.hiddens_hook is not None:
h_init = self.hiddens_hook[1]
self.hidden_size = self.hiddens_hook[0]
else:
# deal with h_init after parameters are made (have to make the same size as hiddens that are computed)
self.hidden_size = hidden_size
##################
# for generating #
##################
# symbolic scalar for how many recurrent steps to use during generation from the model
self.n_steps = T.iscalar("generate_n_steps")
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if self.params_hook is not None:
if not bidirectional:
(W_x_c, W_x_i, W_x_f, W_x_o,
U_h_c, U_h_i, U_h_f, U_h_o,
W_h_y, b_c, b_i, b_f, b_o,
b_y) = self.params_hook
recurrent_params = [U_h_c, U_h_i, U_h_f, U_h_o]
else:
(W_x_c, W_x_i, W_x_f, W_x_o,
U_h_c, U_h_i, U_h_f, U_h_o,
U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b,
W_h_y, b_c, b_i, b_f, b_o,
b_y) = self.params_hook
recurrent_params = [U_h_c, U_h_i, U_h_f, U_h_o, U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b]
# otherwise, construct our params
else:
# all input-to-hidden weights
W_x_c, W_x_i, W_x_f, W_x_o = [
get_weights(weights_init=weights_init,
shape=(self.input_size, self.hidden_size),
name="W_x_%s" % sub,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for sub in ['c', 'i', 'f', 'o']
]
# all hidden-to-hidden weights
U_h_c, U_h_i, U_h_f, U_h_o = [
get_weights(weights_init=r_weights_init,
shape=(self.hidden_size, self.hidden_size),
name="U_h_%s" % sub,
# if gaussian
mean=r_weights_mean,
std=r_weights_std,
# if uniform
interval=r_weights_interval)
for sub in ['c', 'i', 'f', 'o']
]
# hidden-to-output weights
W_h_y = get_weights(weights_init=weights_init,
shape=(self.hidden_size, self.output_size),
name="W_h_y",
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
# biases
b_c, b_i, b_f, b_o = [
get_bias(shape=(self.hidden_size,),
name="b_%s" % sub,
init_values=r_bias_init)
for sub in ['c', 'i', 'f', 'o']
]
# output bias
b_y = get_bias(shape=(self.output_size,),
name="b_y",
init_values=bias_init)
# clip gradients if we are doing that
recurrent_params = [U_h_c, U_h_i, U_h_f, U_h_o]
if clip_recurrent_grads:
clip = abs(clip_recurrent_grads)
U_h_c, U_h_i, U_h_f, U_h_o = [theano.gradient.grad_clip(p, -clip, clip) for p in recurrent_params]
# bidirectional params
if bidirectional:
# all hidden-to-hidden weights
U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b = [
get_weights(weights_init=r_weights_init,
shape=(self.hidden_size, self.hidden_size),
name="U_h_%s_b" % sub,
# if gaussian
mean=r_weights_mean,
std=r_weights_std,
# if uniform
interval=r_weights_interval)
for sub in ['c', 'i', 'f', 'o']
]
recurrent_params += [U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b]
if clip_recurrent_grads:
clip = abs(clip_recurrent_grads)
U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b = [theano.gradient.grad_clip(p, -clip, clip) for p in
[U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b]]
# put all the parameters into our list, and make sure it is in the same order as when we try to load
# them from a params_hook!!!
self.params = [W_x_c, W_x_i, W_x_f, W_x_o] + recurrent_params + [W_h_y, b_c, b_i, b_f, b_o, b_y]
# make h_init the right sized tensor
if not self.hiddens_hook:
h_init = T.zeros_like(T.dot(xs[0], W_x_c))
c_init = T.zeros_like(T.dot(xs[0], W_x_c))
###############
# computation #
###############
# move some computation outside of scan to speed it up!
x_c = T.dot(xs, W_x_c) + b_c
x_i = T.dot(xs, W_x_i) + b_i
x_f = T.dot(xs, W_x_f) + b_f
x_o = T.dot(xs, W_x_o) + b_o
# now do the recurrent stuff
(self.hiddens, _), self.updates = theano.scan(
fn=self.recurrent_step,
sequences=[x_c, x_i, x_f, x_o],
outputs_info=[h_init, c_init],
non_sequences=[U_h_c, U_h_i, U_h_f, U_h_o],
go_backwards=backward,
name="lstm_scan",
strict=True
)
# if bidirectional, do the same in reverse!
if bidirectional:
(hiddens_b, _), updates_b = theano.scan(
fn=self.recurrent_step,
sequences=[x_c, x_i, x_f, x_o],
outputs_info=[h_init, c_init],
non_sequences=[U_h_c_b, U_h_i_b, U_h_f_b, U_h_o_b],
go_backwards=not backward,
name="lstm_scan_back",
strict=True
)
# flip the hiddens to be the right direction
hiddens_b = hiddens_b[::-1]
# update stuff
self.updates.update(updates_b)
self.hiddens += hiddens_b
# add noise (like dropout) if we wanted it!
if noise:
self.hiddens = T.switch(self.noise_switch,
noise_func(input=self.hiddens),
self.hiddens)
# now compute the outputs from the leftover (top level) hiddens
self.output = activation_func(
T.dot(self.hiddens, W_h_y) + b_y
)
# now to define the cost of the model - use the cost function to compare our output with the target value.
self.cost = cost_function(output=self.output, target=ys, **cost_args)
log.info("Initialized an LSTM!")
def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None,
input_size=28*28, hidden_size=1000, noise_level=0.4,
hidden_activation='tanh', visible_activation='sigmoid', cost_function='binary_crossentropy'):
# initialize the Model superclass
super(DenoisingAutoencoder, self).__init__(
**{arg: val for (arg, val) in locals().iteritems() if arg is not 'self'}
)
# Define model hyperparameters
# deal with the inputs_hook and hiddens_hook for the size parameters!
# if the hook exists, grab the size from the first element of the tuple.
if self.inputs_hook is not None:
assert len(self.inputs_hook) == 2, "Was expecting inputs_hook to be a tuple."
self.input_size = inputs_hook[0]
if self.hiddens_hook is not None:
assert len(self.hiddens_hook) == 2, "was expecting hiddens_hook to be a tuple."
hidden_size = hiddens_hook[0]
# use the helper methods to grab appropriate activation functions from names!
hidden_activation = get_activation_function(hidden_activation)
visible_activation = get_activation_function(visible_activation)
# do the same for the cost function
cost_function = get_cost_function(cost_function)
# Now, define the symbolic input to the model (Theano)
# We use a matrix rather than a vector so that minibatch processing can be done in parallel.
# Make sure to deal with 'inputs_hook' if it exists!
if self.inputs_hook is not None:
# grab the new input variable from the inputs_hook tuple
x = self.inputs_hook[1]
else:
x = T.matrix("X")
self.inputs = [x]
# Build the model's parameters - a weight matrix and two bias vectors
# Make sure to deal with 'params_hook' if it exists!
if self.params_hook:
# check to see if it contains the three necessary variables
assert len(self.params_hook) == 3, "Not correct number of params to DAE, needs 3!"
W, b0, b1 = self.params_hook
else:
W = get_weights_uniform(shape=(self.input_size, hidden_size), name="W")
b0 = get_bias(shape=self.input_size, name="b0")
b1 = get_bias(shape=hidden_size, name="b1")
self.params = [W, b0, b1]
# Perform the computation for a denoising autoencoder!
# first, add noise (corrupt) the input
corrupted_input = salt_and_pepper(input=x, noise_level=noise_level)
# next, run the hidden layer given the inputs (the encoding function)
# We don't need to worry about hiddens_hook during training, because we can't
# run a cost without having the input!
# hiddens_hook is more for the run function and linking methods below.
hiddens = hidden_activation(T.dot(corrupted_input, W) + b1)
# finally, create the reconstruction from the hidden layer (we tie the weights with W.T)
reconstruction = visible_activation(T.dot(hiddens, W.T) + b0)
# the training cost is reconstruction error
self.train_cost = cost_function(output=reconstruction, target=x)
# Compile everything into a Theano function for prediction!
# When using real-world data in predictions, we wouldn't corrupt the input first.
# Therefore, create another version of the hiddens and reconstruction without adding the noise.
# Here is where we would handle hiddens_hook because this is a generative model!
# For the run function, it would take in the hiddens instead of the input variable x.
if self.hiddens_hook is not None:
self.hiddens = self.hiddens_hook[1]
else:
self.hiddens = hidden_activation(T.dot(x, W) + b1)
# make the reconstruction (generated) from the hiddens
self.recon_predict = visible_activation(T.dot(self.hiddens, W.T) + b0)
# now compile the run function accordingly - if it used x or hiddens as the input.
if self.hiddens_hook is not None:
self.f_run = function(inputs=[self.hiddens], outputs=self.recon_predict)
else:
self.f_run = function(inputs=[x], outputs=self.recon_predict)
def __init__(self,
config=None,
defaults=_defaults,
inputs_hook=None,
hiddens_hook=None,
params_hook=None,
input_size=None,
hidden_size=None,
corruption_level=None,
hidden_activation=None,
visible_activation=None,
cost_function=None):
# Now, initialize with Model class to combine config and defaults!
# Here, defaults is defined via a dictionary. However, you could also
# pass a filename to a JSON or YAML file with the same format.
super(DenoisingAutoencoder, self).__init__(config=config,
defaults=defaults)
# Any parameter from the 'config' will overwrite the 'defaults' dictionary.
# These parameters are now accessible from the 'self.args' variable!
# When accessing model parameters, it is best practice to try to find the parameters
# explicitly passed first, and then go to the 'self.args' configuration.
# Define model hyperparameters
# deal with the inputs_hook and hiddens_hook for the size parameters!
# if the hook exists, grab the size from the first element of the tuple.
if inputs_hook:
input_size = inputs_hook[0]
# otherwise, grab the size from the configurations.
else:
input_size = input_size or self.args.get('input_size')
if hiddens_hook:
hidden_size = hiddens_hook[0]
else:
hidden_size = hidden_size or self.args.get('hidden_size')
corruption_level = corruption_level or self.args.get(
'corruption_level')
# use the helper methods to grab appropriate activation functions from names!
hidden_act_name = hidden_activation or self.args.get(
'hidden_activation')
hidden_activation = get_activation_function(hidden_act_name)
visible_act_name = visible_activation or self.args.get(
'visible_activation')
visible_activation = get_activation_function(visible_act_name)
# do the same for the cost function
cost_func_name = cost_function or self.args.get('cost_function')
cost_function = get_cost_function(cost_func_name)
# Now, define the symbolic input to the model (Theano)
# We use a matrix rather than a vector so that minibatch processing can be done in parallel.
# Make sure to deal with 'inputs_hook' if it exists!
if inputs_hook:
# grab the new input variable from the inputs_hook tuple
x = inputs_hook[1]
else:
x = T.fmatrix("X")
self.inputs = [x]
# Build the model's parameters - a weight matrix and two bias vectors
# Make sure to deal with 'params_hook' if it exists!
if params_hook:
# check to see if it contains the three necessary variables
assert len(params_hook
) == 3, "Not correct number of params to DAE, needs 3!"
W, b0, b1 = params_hook
else:
W = get_weights_uniform(shape=(input_size, hidden_size), name="W")
b0 = get_bias(shape=input_size, name="b0")
b1 = get_bias(shape=hidden_size, name="b1")
self.params = [W, b0, b1]
# Perform the computation for a denoising autoencoder!
# first, add noise (corrupt) the input
corrupted_input = salt_and_pepper(input=x,
corruption_level=corruption_level)
# next, compute the hidden layer given the inputs (the encoding function)
# We don't need to worry about hiddens_hook during training, because we can't
# compute a cost without having the input!
# hiddens_hook is more for the predict function and linking methods below.
hiddens = hidden_activation(T.dot(corrupted_input, W) + b1)
# finally, create the reconstruction from the hidden layer (we tie the weights with W.T)
reconstruction = visible_activation(T.dot(hiddens, W.T) + b0)
# the training cost is reconstruction error
self.train_cost = cost_function(output=reconstruction, target=x)
# Compile everything into a Theano function for prediction!
# When using real-world data in predictions, we wouldn't corrupt the input first.
# Therefore, create another version of the hiddens and reconstruction without adding the noise.
# Here is where we would handle hiddens_hook because this is a generative model!
# For the predict function, it would take in the hiddens instead of the input variable x.
if hiddens_hook:
self.hiddens = hiddens_hook[1]
else:
self.hiddens = hidden_activation(T.dot(x, W) + b1)
# make the reconstruction (generated) from the hiddens
self.recon_predict = visible_activation(T.dot(self.hiddens, W.T) + b0)
# now compile the predict function accordingly - if it used x or hiddens as the input.
if hiddens_hook:
self.f_predict = function(inputs=[self.hiddens],
outputs=self.recon_predict)
else:
self.f_predict = function(inputs=[x], outputs=self.recon_predict)
def __init__(
self,
inputs_hook=None,
params_hook=None,
outdir="outputs/basic",
input_size=None,
output_size=None,
activation="rectifier",
cost="mse",
cost_args=None,
weights_init="uniform",
weights_mean=0,
weights_std=5e-3,
weights_interval="montreal",
bias_init=0.0,
noise=None,
noise_level=None,
mrg=RNG_MRG.MRG_RandomStreams(1),
**kwargs
):
"""
Initialize a basic layer.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together. For now, it needs to include the shape information (normally the
dimensionality of the input i.e. input_size).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the layer. If shape is provided in `inputs_hook`,
this is optional.
output_size : int
The size (dimensionality) of the output from the layer.
activation : str or callable
The activation function to use after the dot product going from input -> output. This can be a string
representing an option from opendeep.utils.activation, or your own function as long as it is callable.
cost : str or callable
The cost function to use when training the layer. This should be appropriate for the output type, i.e.
mse for real-valued outputs, binary cross-entropy for binary outputs, etc.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
weights_init : str
Determines the method for initializing input -> output weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
noise : str
What type of noise to use for corrupting the output (if not None). See opendeep.utils.noise
for options. This should be appropriate for the output activation, i.e. Gaussian for tanh or other
real-valued activations, etc. Often, you will use 'dropout' here as a regularization in BasicLayers.
noise_level : float
The amount of noise to use for the noise function specified by `noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
mrg : random
A random number generator that is used when adding noise.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
"""
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals().copy()
initial_parameters.pop("self")
super(Dense, self).__init__(**initial_parameters)
##################
# specifications #
##################
# grab info from the inputs_hook, or from parameters
if inputs_hook is not None: # inputs_hook is a tuple of (Shape, Input)
assert len(inputs_hook) == 2, "Expected inputs_hook to be tuple!" # make sure inputs_hook is a tuple
self.input = inputs_hook[1]
else:
# make the input a symbolic matrix
self.input = T.matrix("X")
# now that we have the input specs, define the output 'target' variable to be used in supervised training!
if kwargs.get("out_as_probs") == False:
self.target = T.vector("Y", dtype="int64")
else:
self.target = T.matrix("Y")
# either grab the output's desired size from the parameter directly, or copy input_size
self.output_size = self.output_size or self.input_size
# other specifications
# activation function!
activation_func = get_activation_function(activation)
# cost function!
cost_func = get_cost_function(cost)
cost_args = cost_args or dict()
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if params_hook is not None:
# make sure the params_hook has W (weights matrix) and b (bias vector)
assert len(params_hook) == 2, "Expected 2 params (W and b) for Dense, found {0!s}!".format(len(params_hook))
W, b = params_hook
else:
W = get_weights(
weights_init=weights_init,
shape=(self.input_size, self.output_size),
name="W",
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval,
)
# grab the bias vector
b = get_bias(shape=output_size, name="b", init_values=bias_init)
# Finally have the two parameters - weights matrix W and bias vector b. That is all!
self.params = [W, b]
###############
# computation #
###############
# Here is the meat of the computation transforming input -> output
# It simply involves a matrix multiplication of inputs*weights, adding the bias vector, and then passing
# the result through our activation function (normally something nonlinear such as: max(0, output))
self.output = activation_func(T.dot(self.input, W) + b)
# Now deal with noise if we added it:
if noise:
log.debug("Adding noise switch.")
if noise_level is not None:
noise_func = get_noise(noise, noise_level=noise_level, mrg=mrg)
else:
noise_func = get_noise(noise, mrg=mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.switch = sharedX(value=1, name="basiclayer_noise_switch")
self.output = T.switch(self.switch, noise_func(input=self.output), self.output)
# now to define the cost of the model - use the cost function to compare our output with the target value.
self.cost = cost_func(output=self.output, target=self.target, **cost_args)
log.debug(
"Initialized a basic fully-connected layer with shape %s and activation: %s",
str((self.input_size, self.output_size)),
str(activation),
)
def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None, outdir=None,
input_size=None, hidden_size=None,
layers=2, walkbacks=4,
visible_activation='sigmoid', hidden_activation='tanh',
input_sampling=True, mrg=RNG_MRG.MRG_RandomStreams(1),
tied_weights=True,
weights_init='uniform', weights_interval='montreal', weights_mean=0, weights_std=5e-3,
bias_init=0,
cost_function='binary_crossentropy', cost_args=None,
add_noise=True, noiseless_h1=True,
hidden_noise='gaussian', hidden_noise_level=2, input_noise='salt_and_pepper', input_noise_level=0.4,
noise_decay='exponential', noise_annealing=1,
image_width=None, image_height=None,
rnn_hidden_size=None, rnn_hidden_activation='rectifier',
rnn_weights_init='identity',
rnn_weights_mean=0, rnn_weights_std=5e-3, rnn_weights_interval='montreal',
rnn_bias_init=0,
generate_n_steps=200):
"""
Initialize an RNN-GSN.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere.
This is used for linking different models together (e.g. setting the DAE model's hidden layers to the RNN's
output layer gives a generative recurrent model.) For now, it needs to include the shape
information (normally the dimensionality of the hiddens i.e. n_hidden).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the DAE. If shape is provided in `inputs_hook`, this is optional.
The :class:`Model` requires an `output_size`, which gets set to this value because the DAE is an
unsupervised model. The output is a reconstruction of the input.
hidden_size : int
The size (dimensionality) of the hidden layer for the DAE. Generally, you want it to be larger than
`input_size`, which is known as *overcomplete*.
layers : int
The number of hidden layers to use.
walkbacks : int
The number of walkbacks to perform (the variable K in Bengio's paper above). A walkback is a Gibbs sample
from the DAE, which means the model generates inputs in sequence, where each generated input is compared
to the original input to create the reconstruction cost for training. For running the model, the very last
generated input in the Gibbs chain is used as the output.
visible_activation : str or callable
The nonlinear (or linear) visible activation to perform after the dot product from hiddens -> visible layer.
This activation function should be appropriate for the input unit types, i.e. 'sigmoid' for binary inputs.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The nonlinear (or linear) hidden activation to perform after the dot product from visible -> hiddens layer.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
input_sampling : bool
During walkbacks, whether to sample from the generated input to create a new starting point for the next
walkback (next step in the Gibbs chain). This generally makes walkbacks more effective by making the
process more stochastic - more likely to find spurious modes in the model's representation.
mrg : random
A random number generator that is used when adding noise into the network and for sampling from the input.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
tied_weights : bool
DAE has two weight matrices - W from input -> hiddens and V from hiddens -> input. This boolean
determines if V = W.T, which 'ties' V to W and reduces the number of parameters necessary during training.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the reconstruction cost of the model. This should be appropriate
for the type of input, i.e. use 'binary_crossentropy' for binary inputs, or 'mse' for real-valued inputs.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
add_noise : bool
Whether to add noise (corrupt) the input before passing it through the computation graph during training.
This should most likely be set to the default of True, because this is a *denoising* autoencoder after all.
noiseless_h1 : bool
Whether to not add noise (corrupt) the hidden layer during computation.
hidden_noise : str
What type of noise to use for corrupting the hidden layer (if not `noiseless_h1`). See opendeep.utils.noise
for options. This should be appropriate for the hidden unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
hidden_noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
input_noise : str
What type of noise to use for corrupting the input before computation (if `add_noise`).
See opendeep.utils.noise for options. This should be appropriate for the input units, i.e. salt-and-pepper
for binary units, etc.
input_noise_level : float
The amount of noise used to corrupt the input. This could be the masking probability for salt-and-pepper,
standard deviation for Gaussian, interval for Uniform, etc.
noise_decay : str or False
Whether to use `input_noise` scheduling (decay `input_noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the DAE learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_annealing : float
The amount to reduce the `input_noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
image_width : int
If the input should be represented as an image, the width of the input image. If not specified, it will be
close to the square factor of the `input_size`.
image_height : int
If the input should be represented as an image, the height of the input image. If not specified, it will be
close to the square factor of the `input_size`.
rnn_hidden_size : int
The number of hidden units (dimensionality) to use in the recurrent layer.
rnn_hidden_activation : str or Callable
The activation function to apply to recurrent units. See opendeep.utils.activation for options.
rnn_weights_init : str
Determines the method for initializing recurrent weights. See opendeep.utils.nnet for options. 'Identity'
works well with 'rectifier' `rnn_hidden_activation`.
rnn_weights_mean : float
If Gaussian `rnn_weights_init`, the mean value to use.
rnn_weights_std : float
If Gaussian `rnn_weights_init`, the standard deviation to use.
rnn_weights_interval : str or float
If Uniform `rnn_weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
rnn_bias_init : float
The initial value to use for the recurrent bias parameter. Most often, the default of 0.0 is preferred.
generate_n_steps : int
When generating from the model, how many steps to generate.
"""
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(RNN_GSN, self).__init__(**initial_parameters)
##################
# specifications #
##################
self.input_size = input_size
self.layers = layers
self.walkbacks = walkbacks
self.input_sampling = input_sampling
self.mrg = mrg
self.tied_weights = tied_weights
self.noise_decay = noise_decay
self.noise_annealing = noise_annealing
self.add_noise = add_noise
self.noiseless_h1 = noiseless_h1
self.hidden_noise = hidden_noise
self.hidden_noise_level = hidden_noise_level
self.input_noise = input_noise
self.input_noise_level = input_noise_level
self.image_width = image_width
self.image_height = image_height
# grab info from the inputs_hook, hiddens_hook, or from parameters
if self.inputs_hook is not None: # inputs_hook is a tuple of (Shape, Input)
raise NotImplementedError("Inputs_hook not implemented yet for RNN-GSN")
else:
# make the input a symbolic matrix - a sequence of inputs
self.input = T.matrix('Xs')
# set an initial value for the recurrent hiddens
self.u0 = T.zeros((rnn_hidden_size,))
# make a symbolic vector for the initial recurrent hiddens value to use during generation for the model
self.generate_u0 = T.vector("generate_u0")
# either grab the hidden's desired size from the parameter directly, or copy n_in
self.hidden_size = hidden_size or self.input_size
# deal with hiddens_hook
if self.hiddens_hook is not None:
raise NotImplementedError("Hiddens_hook not implemented yet for RNN-GSN")
# other specifications
# visible activation function!
self.visible_activation_func = get_activation_function(visible_activation)
# make sure the sampling functions are appropriate for the activation functions.
if is_binary(self.visible_activation_func):
self.visible_sampling = mrg.binomial
else:
# TODO: implement non-binary activation
log.error("Non-binary visible activation not supported yet!")
raise NotImplementedError("Non-binary visible activation not supported yet!")
# hidden activation function!
self.hidden_activation_func = get_activation_function(hidden_activation)
# recurrent hidden activation function!
self.rnn_hidden_activation_func = get_activation_function(rnn_hidden_activation)
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args
# symbolic scalar for how many recurrent steps to use during generation from the model
self.n_steps = T.iscalar("generate_n_steps")
# determine the sizes of each layer in a list.
# layer sizes, from h0 to hK (h0 is the visible layer)
self.layer_sizes = [self.input_size] + [self.hidden_size] * self.layers
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if self.params_hook is not None:
# if tied weights, expect (layers*2 + 1) params for GSN and (int(layers+1)/int(2) + 3) for RNN
if self.tied_weights:
expected_num = (2*self.layers + 1) + (int(self.layers+1)/2 + 3)
assert len(self.params_hook) == expected_num, \
"Tied weights: expected {0!s} params, found {1!s}!".format(expected_num, len(self.params_hook))
gsn_len = (2*self.layers + 1)
self.weights_list = self.params_hook[:self.layers]
self.bias_list = self.params_hook[self.layers:gsn_len]
# if untied weights, expect layers*3 + 1 params
else:
expected_num = (3*self.layers + 1) + (int(self.layers + 1)/2 + 3)
assert len(self.params_hook) == expected_num, \
"Untied weights: expected {0!s} params, found {1!s}!".format(expected_num, len(self.params_hook))
gsn_len = (3*self.layers + 1)
self.weights_list = self.params_hook[:2*self.layers]
self.bias_list = self.params_hook[2*self.layers:gsn_len]
rnn_len = gsn_len + int(self.layers + 1) / 2
self.recurrent_to_gsn_weights_list = self.params_hook[gsn_len:rnn_len]
self.W_u_u = self.params_hook[rnn_len:rnn_len + 1]
self.W_x_u = self.params_hook[rnn_len + 1:rnn_len + 2]
self.recurrent_bias = self.params_hook[rnn_len + 2:rnn_len + 3]
# otherwise, construct our params
else:
# initialize a list of weights and biases based on layer_sizes for the GSN
self.weights_list = [get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i], self.layer_sizes[i + 1]),
name="W_{0!s}_{1!s}".format(i, i + 1),
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in range(self.layers)]
# add more weights if we aren't tying weights between layers (need to add for higher-lower layers now)
if not self.tied_weights:
self.weights_list.extend(
[get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i + 1], self.layer_sizes[i]),
name="W_{0!s}_{1!s}".format(i + 1, i),
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in reversed(range(self.layers))]
)
# initialize each layer bias to 0's.
self.bias_list = [get_bias(shape=(self.layer_sizes[i],),
name='b_' + str(i),
init_values=bias_init)
for i in range(self.layers + 1)]
self.recurrent_to_gsn_weights_list = [
get_weights(weights_init=rnn_weights_init,
shape=(rnn_hidden_size, self.layer_sizes[layer]),
name="W_u_h{0!s}".format(layer),
# if gaussian
mean=rnn_weights_mean,
std=rnn_weights_std,
# if uniform
interval=rnn_weights_interval)
for layer in range(self.layers + 1) if layer % 2 != 0
]
self.W_u_u = get_weights(weights_init=rnn_weights_init,
shape=(rnn_hidden_size, rnn_hidden_size),
name="W_u_u",
# if gaussian
mean=rnn_weights_mean,
std=rnn_weights_std,
#if uniform
interval=rnn_weights_interval)
self.W_x_u = get_weights(weights_init=rnn_weights_init,
shape=(self.input_size, rnn_hidden_size),
name="W_x_u",
# if gaussian
mean=rnn_weights_mean,
std=rnn_weights_std,
# if uniform
interval=rnn_weights_interval)
self.recurrent_bias = get_bias(shape=(rnn_hidden_size,),
name="b_u",
init_values=rnn_bias_init)
# build the params of the model into a list
self.gsn_params = self.weights_list + self.bias_list
self.params = self.gsn_params + \
self.recurrent_to_gsn_weights_list + \
[self.W_u_u, self.W_x_u, self.recurrent_bias]
log.debug("rnn-gsn params: %s", str(self.params))
# Create the RNN-GSN graph!
self.x_sample, self.cost, self.monitors, self.updates_train, self.x_ts, self.updates_generate, self.u_t = \
self._build_rnngsn()
log.info("Initialized an RNN-GSN!")
def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None, outdir='outputs/gsn/',
input_size=None, hidden_size=1000,
layers=2, walkbacks=4,
visible_activation='sigmoid', hidden_activation='tanh',
input_sampling=True, mrg=RNG_MRG.MRG_RandomStreams(1),
tied_weights=True,
weights_init='uniform', weights_interval='montreal', weights_mean=0, weights_std=5e-3,
bias_init=0.0,
cost_function='binary_crossentropy', cost_args=None,
add_noise=True, noiseless_h1=True,
hidden_noise='gaussian', hidden_noise_level=2, input_noise='salt_and_pepper', input_noise_level=0.4,
noise_decay='exponential', noise_annealing=1,
image_width=None, image_height=None,
**kwargs):
"""
Initialize a GSN.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere.
This is used for linking different models together (e.g. setting the DAE model's hidden layers to the RNN's
output layer gives a generative recurrent model.) For now, it needs to include the shape
information (normally the dimensionality of the hiddens i.e. n_hidden).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the DAE. If shape is provided in `inputs_hook`, this is optional.
The :class:`Model` requires an `output_size`, which gets set to this value because the DAE is an
unsupervised model. The output is a reconstruction of the input.
hidden_size : int
The size (dimensionality) of the hidden layer for the DAE. Generally, you want it to be larger than
`input_size`, which is known as *overcomplete*.
visible_activation : str or callable
The nonlinear (or linear) visible activation to perform after the dot product from hiddens -> visible layer.
This activation function should be appropriate for the input unit types, i.e. 'sigmoid' for binary inputs.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The nonlinear (or linear) hidden activation to perform after the dot product from visible -> hiddens layer.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
layers : int
The number of hidden layers to use.
walkbacks : int
The number of walkbacks to perform (the variable K in Bengio's paper above). A walkback is a Gibbs sample
from the DAE, which means the model generates inputs in sequence, where each generated input is compared
to the original input to create the reconstruction cost for training. For running the model, the very last
generated input in the Gibbs chain is used as the output.
input_sampling : bool
During walkbacks, whether to sample from the generated input to create a new starting point for the next
walkback (next step in the Gibbs chain). This generally makes walkbacks more effective by making the
process more stochastic - more likely to find spurious modes in the model's representation.
mrg : random
A random number generator that is used when adding noise into the network and for sampling from the input.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
tied_weights : bool
DAE has two weight matrices - W from input -> hiddens and V from hiddens -> input. This boolean
determines if V = W.T, which 'ties' V to W and reduces the number of parameters necessary during training.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the reconstruction cost of the model. This should be appropriate
for the type of input, i.e. use 'binary_crossentropy' for binary inputs, or 'mse' for real-valued inputs.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
add_noise : bool
Whether to add noise (corrupt) the input before passing it through the computation graph during training.
This should most likely be set to the default of True, because this is a *denoising* autoencoder after all.
noiseless_h1 : bool
Whether to not add noise (corrupt) the hidden layer during computation.
hidden_noise : str
What type of noise to use for corrupting the hidden layer (if not `noiseless_h1`). See opendeep.utils.noise
for options. This should be appropriate for the hidden unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
hidden_noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
input_noise : str
What type of noise to use for corrupting the input before computation (if `add_noise`).
See opendeep.utils.noise for options. This should be appropriate for the input units, i.e. salt-and-pepper
for binary units, etc.
input_noise_level : float
The amount of noise used to corrupt the input. This could be the masking probability for salt-and-pepper,
standard deviation for Gaussian, interval for Uniform, etc.
noise_decay : str or False
Whether to use `input_noise` scheduling (decay `input_noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the DAE learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_annealing : float
The amount to reduce the `input_noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
image_width : int
If the input should be represented as an image, the width of the input image. If not specified, it will be
close to the square factor of the `input_size`.
image_height : int
If the input should be represented as an image, the height of the input image. If not specified, it will be
close to the square factor of the `input_size`.
"""
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(GSN, self).__init__(**initial_parameters)
# when the input should be thought of as an image, either use the specified width and height,
# or try to make as square as possible.
if image_height is None and image_width is None:
(_h, _w) = closest_to_square_factors(self.input_size)
self.image_width = _w
self.image_height = _h
else:
self.image_height = image_height
self.image_width = image_width
############################
# Theano variables and RNG #
############################
if self.inputs_hook is None:
self.X = T.matrix('X')
else:
# inputs_hook is a (shape, input) tuple
self.X = self.inputs_hook[1]
##########################
# Network specifications #
##########################
# generally, walkbacks should be at least 2*layers
if layers % 2 == 0:
if walkbacks < 2*layers:
log.warning('Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers',
str(layers), str(walkbacks))
else:
if walkbacks < 2*layers-1:
log.warning('Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers',
str(layers), str(walkbacks))
self.add_noise = add_noise
self.noise_annealing = as_floatX(noise_annealing) # noise schedule parameter
self.hidden_noise_level = sharedX(hidden_noise_level, dtype=theano.config.floatX)
self.hidden_noise = get_noise(name=hidden_noise, noise_level=self.hidden_noise_level, mrg=mrg)
self.input_noise_level = sharedX(input_noise_level, dtype=theano.config.floatX)
self.input_noise = get_noise(name=input_noise, noise_level=self.input_noise_level, mrg=mrg)
self.walkbacks = walkbacks
self.tied_weights = tied_weights
self.layers = layers
self.noiseless_h1 = noiseless_h1
self.input_sampling = input_sampling
self.noise_decay = noise_decay
# if there was a hiddens_hook, unpack the hidden layers in the tensor
if self.hiddens_hook is not None:
hidden_size = self.hiddens_hook[0]
self.hiddens_flag = True
else:
self.hiddens_flag = False
# determine the sizes of each layer in a list.
# layer sizes, from h0 to hK (h0 is the visible layer)
hidden_size = list(raise_to_list(hidden_size))
if len(hidden_size) == 1:
self.layer_sizes = [self.input_size] + hidden_size * self.layers
else:
assert len(hidden_size) == self.layers, "Hiddens sizes and number of hidden layers mismatch." + \
"Hiddens %d and layers %d" % (len(hidden_size), self.layers)
self.layer_sizes = [self.input_size] + hidden_size
if self.hiddens_hook is not None:
self.hiddens = self.unpack_hiddens(self.hiddens_hook[1])
#########################
# Activation functions! #
#########################
# hidden unit activation
self.hidden_activation = get_activation_function(hidden_activation)
# Visible layer activation
self.visible_activation = get_activation_function(visible_activation)
# make sure the sampling functions are appropriate for the activation functions.
if is_binary(self.visible_activation):
self.visible_sampling = mrg.binomial
else:
# TODO: implement non-binary activation
log.error("Non-binary visible activation not supported yet!")
raise NotImplementedError("Non-binary visible activation not supported yet!")
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args or dict()
###############
# Parameters! #
###############
# make sure to deal with params_hook!
if self.params_hook is not None:
# if tied weights, expect layers*2 + 1 params
if self.tied_weights:
assert len(self.params_hook) == 2*layers + 1, \
"Tied weights: expected {0!s} params, found {1!s}!".format(2*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:layers]
self.bias_list = self.params_hook[layers:]
# if untied weights, expect layers*3 + 1 params
else:
assert len(self.params_hook) == 3*layers + 1, \
"Untied weights: expected {0!s} params, found {1!s}!".format(3*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:2*layers]
self.bias_list = self.params_hook[2*layers:]
# otherwise, construct our params
else:
# initialize a list of weights and biases based on layer_sizes for the GSN
self.weights_list = [get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i], self.layer_sizes[i+1]),
name="W_{0!s}_{1!s}".format(i, i+1),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in range(layers)]
# add more weights if we aren't tying weights between layers (need to add for higher-lower layers now)
if not tied_weights:
self.weights_list.extend(
[get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i+1], self.layer_sizes[i]),
name="W_{0!s}_{1!s}".format(i+1, i),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in reversed(range(layers))]
)
# initialize each layer bias to 0's.
self.bias_list = [get_bias(shape=(self.layer_sizes[i],),
name='b_' + str(i),
init_values=bias_init)
for i in range(layers+1)]
# build the params of the model into a list
self.params = self.weights_list + self.bias_list
log.debug("gsn params: %s", str(self.params))
# using the properties, build the computational graph
self.cost, self.monitors, self.output, self.hiddens = self.build_computation_graph()
def __init__(self,
inputs_hook=None,
hiddens_hook=None,
params_hook=None,
outdir='outputs/rnn/',
input_size=None,
hidden_size=None,
output_size=None,
layers=1,
activation='sigmoid',
hidden_activation='relu',
mrg=RNG_MRG.MRG_RandomStreams(1),
weights_init='uniform',
weights_interval='montreal',
weights_mean=0,
weights_std=5e-3,
bias_init=0.0,
r_weights_init='identity',
r_weights_interval='montreal',
r_weights_mean=0,
r_weights_std=5e-3,
r_bias_init=0.0,
cost_function='mse',
cost_args=None,
noise='dropout',
noise_level=None,
noise_decay=False,
noise_decay_amount=.99,
direction='forward',
clip_recurrent_grads=False):
"""
Initialize a simple recurrent network.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere. For recurrent nets,
this will be the initial starting value for hidden layers.
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters.
outdir : str
The location to produce outputs from training or running the :class:`RNN`. If None, nothing will be saved.
input_size : int
The size (dimensionality) of the input. If shape is provided in `inputs_hook`, this is optional.
hidden_size : int
The size (dimensionality) of the hidden layers. If shape is provided in `hiddens_hook`, this is optional.
output_size : int
The size (dimensionality) of the output.
layers : int
The number of stacked hidden layers to use.
activation : str or callable
The nonlinear (or linear) activation to perform after the dot product from hiddens -> output layer.
This activation function should be appropriate for the output unit types, i.e. 'sigmoid' for binary.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The activation to perform for the hidden layers.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
mrg : random
A random number generator that is used when adding noise.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
r_weights_init : str
Determines the method for initializing recurrent model weights. See opendeep.utils.nnet for options.
r_weights_interval : str or float
If Uniform `r_weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
r_weights_mean : float
If Gaussian `r_weights_init`, the mean value to use.
r_weights_std : float
If Gaussian `r_weights_init`, the standard deviation to use.
r_bias_init : float
The initial value to use for the recurrent bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the output cost of the model.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
noise : str
What type of noise to use for the hidden layers and outputs. See opendeep.utils.noise
for options. This should be appropriate for the unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
noise_decay : str or False
Whether to use `noise` scheduling (decay `noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the model learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_decay_amount : float
The amount to reduce the `noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
direction : str
The direction this recurrent model should go over its inputs. Can be 'forward', 'backward', or
'bidirectional'. In the case of 'bidirectional', it will make two passes over the sequence,
computing two sets of hiddens and merging them before running through the final decoder.
clip_recurrent_grads : False or float, optional
Whether to clip the gradients for the parameters that unroll over timesteps (such as the weights
connecting previous hidden states to the current hidden state, and not the weights from current
input to hiddens). If it is a float, the gradients for the weights will be hard clipped to the range
`+-clip_recurrent_grads`.
Raises
------
AssertionError
When asserting various properties of input parameters. See error messages.
"""
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(RNN, self).__init__(**initial_parameters)
##################
# specifications #
##################
self.direction = direction
self.bidirectional = (direction == "bidirectional")
self.backward = (direction == "backward")
self.layers = layers
self.noise = noise
self.weights_init = weights_init
self.weights_mean = weights_mean
self.weights_std = weights_std
self.weights_interval = weights_interval
self.r_weights_init = r_weights_init
self.r_weights_mean = r_weights_mean
self.r_weights_std = r_weights_std
self.r_weights_interval = r_weights_interval
self.bias_init = bias_init
self.r_bias_init = r_bias_init
#########################################
# activation, cost, and noise functions #
#########################################
# recurrent hidden activation function!
self.hidden_activation_func = get_activation_function(
hidden_activation)
# output activation function!
self.activation_func = get_activation_function(activation)
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args or dict()
# Now deal with noise if we added it:
if self.noise:
log.debug('Adding %s noise switch.' % str(noise))
if noise_level is not None:
noise_level = sharedX(value=noise_level)
self.noise_func = get_noise(noise,
noise_level=noise_level,
mrg=mrg)
else:
self.noise_func = get_noise(noise, mrg=mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.noise_switch = sharedX(value=1,
name="basiclayer_noise_switch")
# noise scheduling
if noise_decay and noise_level is not None:
self.noise_schedule = get_decay_function(
noise_decay, noise_level, noise_level.get_value(),
noise_decay_amount)
###############
# inputs hook #
###############
# grab info from the inputs_hook
# in the case of an inputs_hook, recurrent will always work with the leading tensor dimension
# being the temporal dimension.
# input is 3D tensor of (timesteps, batch_size, data_dim)
# if input is 2D tensor, assume it is of the form (timesteps, data_dim) i.e. batch_size is 1. Convert to 3D.
# if input is > 3D tensor, assume it is of form (timesteps, batch_size, data...) and flatten to 3D.
if self.inputs_hook is not None:
self.input = self.inputs_hook[1]
if self.input.ndim == 1:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 'x'),
[1, 2])
self.input_size = 1
elif self.input.ndim == 2:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 1), 1)
elif self.input.ndim == 3:
pass
elif self.input.ndim > 3:
self.input = self.input.flatten(3)
self.input_size = sum(self.input_size)
else:
raise NotImplementedError(
"Recurrent input with %d dimensions not supported!" %
self.input.ndim)
else:
# Assume input coming from optimizer is (batches, timesteps, data)
# so, we need to reshape to (timesteps, batches, data)
xs = T.tensor3("Xs")
xs = xs.dimshuffle(1, 0, 2)
self.input = xs
# The target outputs for supervised training - in the form of (batches, timesteps, output) which is
# the same dimension ordering as the expected input from optimizer.
# therefore, we need to swap it like we did to input xs.
ys = T.tensor3("Ys")
ys = ys.dimshuffle(1, 0, 2)
self.target = ys
################
# hiddens hook #
################
# set an initial value for the recurrent hiddens from hook
if self.hiddens_hook is not None:
self.h_init = self.hiddens_hook[1]
self.hidden_size = self.hiddens_hook[0]
else:
# deal with h_init after parameters are made (have to make the same size as hiddens that are computed)
self.hidden_size = hidden_size
##################
# for generating #
##################
# symbolic scalar for how many recurrent steps to use during generation from the model
self.n_steps = T.iscalar("generate_n_steps")
self.output, self.hiddens, self.updates, self.cost, self.params = self.build_computation_graph(
)
def __init__(self,
inputs_hook=None,
params_hook=None,
outdir='outputs/basic',
input_size=None,
output_size=None,
activation='rectifier',
cost='mse',
cost_args=None,
weights_init='uniform',
weights_mean=0,
weights_std=5e-3,
weights_interval='montreal',
bias_init=0.0,
noise=None,
noise_level=None,
mrg=RNG_MRG.MRG_RandomStreams(1),
**kwargs):
"""
Initialize a basic layer.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together. For now, it needs to include the shape information (normally the
dimensionality of the input i.e. input_size).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the layer. If shape is provided in `inputs_hook`,
this is optional.
output_size : int
The size (dimensionality) of the output from the layer.
activation : str or callable
The activation function to use after the dot product going from input -> output. This can be a string
representing an option from opendeep.utils.activation, or your own function as long as it is callable.
cost : str or callable
The cost function to use when training the layer. This should be appropriate for the output type, i.e.
mse for real-valued outputs, binary cross-entropy for binary outputs, etc.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
weights_init : str
Determines the method for initializing input -> output weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
noise : str
What type of noise to use for corrupting the output (if not None). See opendeep.utils.noise
for options. This should be appropriate for the output activation, i.e. Gaussian for tanh or other
real-valued activations, etc. Often, you will use 'dropout' here as a regularization in BasicLayers.
noise_level : float
The amount of noise to use for the noise function specified by `noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
mrg : random
A random number generator that is used when adding noise.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
"""
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(Dense, self).__init__(**initial_parameters)
##################
# specifications #
##################
# grab info from the inputs_hook, or from parameters
if inputs_hook is not None: # inputs_hook is a tuple of (Shape, Input)
assert len(
inputs_hook
) == 2, 'Expected inputs_hook to be tuple!' # make sure inputs_hook is a tuple
self.input = inputs_hook[1]
else:
# make the input a symbolic matrix
self.input = T.matrix('X')
# now that we have the input specs, define the output 'target' variable to be used in supervised training!
if kwargs.get('out_as_probs') == False:
self.target = T.vector('Y', dtype='int64')
else:
self.target = T.matrix('Y')
# either grab the output's desired size from the parameter directly, or copy input_size
self.output_size = self.output_size or self.input_size
# other specifications
# activation function!
activation_func = get_activation_function(activation)
# cost function!
cost_func = get_cost_function(cost)
cost_args = cost_args or dict()
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if params_hook is not None:
# make sure the params_hook has W (weights matrix) and b (bias vector)
assert len(params_hook) == 2, \
"Expected 2 params (W and b) for Dense, found {0!s}!".format(len(params_hook))
W, b = params_hook
else:
W = get_weights(
weights_init=weights_init,
shape=(self.input_size, self.output_size),
name="W",
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
# grab the bias vector
b = get_bias(shape=output_size, name="b", init_values=bias_init)
# Finally have the two parameters - weights matrix W and bias vector b. That is all!
self.params = [W, b]
###############
# computation #
###############
# Here is the meat of the computation transforming input -> output
# It simply involves a matrix multiplication of inputs*weights, adding the bias vector, and then passing
# the result through our activation function (normally something nonlinear such as: max(0, output))
self.output = activation_func(T.dot(self.input, W) + b)
# Now deal with noise if we added it:
if noise:
log.debug('Adding noise switch.')
if noise_level is not None:
noise_func = get_noise(noise, noise_level=noise_level, mrg=mrg)
else:
noise_func = get_noise(noise, mrg=mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.switch = sharedX(value=1, name="basiclayer_noise_switch")
self.output = T.switch(self.switch, noise_func(input=self.output),
self.output)
# now to define the cost of the model - use the cost function to compare our output with the target value.
self.cost = cost_func(output=self.output,
target=self.target,
**cost_args)
log.debug(
"Initialized a basic fully-connected layer with shape %s and activation: %s",
str((self.input_size, self.output_size)), str(activation))
def __init__(self,
inputs_hook=None,
hiddens_hook=None,
params_hook=None,
outdir='outputs/gsn/',
input_size=None,
hidden_size=1000,
layers=2,
walkbacks=4,
visible_activation='sigmoid',
hidden_activation='tanh',
input_sampling=True,
mrg=RNG_MRG.MRG_RandomStreams(1),
tied_weights=True,
weights_init='uniform',
weights_interval='montreal',
weights_mean=0,
weights_std=5e-3,
bias_init=0.0,
cost_function='binary_crossentropy',
cost_args=None,
add_noise=True,
noiseless_h1=True,
hidden_noise='gaussian',
hidden_noise_level=2,
input_noise='salt_and_pepper',
input_noise_level=0.4,
noise_decay='exponential',
noise_annealing=1,
image_width=None,
image_height=None,
**kwargs):
"""
Initialize a GSN.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere.
This is used for linking different models together (e.g. setting the DAE model's hidden layers to the RNN's
output layer gives a generative recurrent model.) For now, it needs to include the shape
information (normally the dimensionality of the hiddens i.e. n_hidden).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the DAE. If shape is provided in `inputs_hook`, this is optional.
The :class:`Model` requires an `output_size`, which gets set to this value because the DAE is an
unsupervised model. The output is a reconstruction of the input.
hidden_size : int
The size (dimensionality) of the hidden layer for the DAE. Generally, you want it to be larger than
`input_size`, which is known as *overcomplete*.
visible_activation : str or callable
The nonlinear (or linear) visible activation to perform after the dot product from hiddens -> visible layer.
This activation function should be appropriate for the input unit types, i.e. 'sigmoid' for binary inputs.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The nonlinear (or linear) hidden activation to perform after the dot product from visible -> hiddens layer.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
layers : int
The number of hidden layers to use.
walkbacks : int
The number of walkbacks to perform (the variable K in Bengio's paper above). A walkback is a Gibbs sample
from the DAE, which means the model generates inputs in sequence, where each generated input is compared
to the original input to create the reconstruction cost for training. For running the model, the very last
generated input in the Gibbs chain is used as the output.
input_sampling : bool
During walkbacks, whether to sample from the generated input to create a new starting point for the next
walkback (next step in the Gibbs chain). This generally makes walkbacks more effective by making the
process more stochastic - more likely to find spurious modes in the model's representation.
mrg : random
A random number generator that is used when adding noise into the network and for sampling from the input.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
tied_weights : bool
DAE has two weight matrices - W from input -> hiddens and V from hiddens -> input. This boolean
determines if V = W.T, which 'ties' V to W and reduces the number of parameters necessary during training.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the reconstruction cost of the model. This should be appropriate
for the type of input, i.e. use 'binary_crossentropy' for binary inputs, or 'mse' for real-valued inputs.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
add_noise : bool
Whether to add noise (corrupt) the input before passing it through the computation graph during training.
This should most likely be set to the default of True, because this is a *denoising* autoencoder after all.
noiseless_h1 : bool
Whether to not add noise (corrupt) the hidden layer during computation.
hidden_noise : str
What type of noise to use for corrupting the hidden layer (if not `noiseless_h1`). See opendeep.utils.noise
for options. This should be appropriate for the hidden unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
hidden_noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
input_noise : str
What type of noise to use for corrupting the input before computation (if `add_noise`).
See opendeep.utils.noise for options. This should be appropriate for the input units, i.e. salt-and-pepper
for binary units, etc.
input_noise_level : float
The amount of noise used to corrupt the input. This could be the masking probability for salt-and-pepper,
standard deviation for Gaussian, interval for Uniform, etc.
noise_decay : str or False
Whether to use `input_noise` scheduling (decay `input_noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the DAE learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_annealing : float
The amount to reduce the `input_noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
image_width : int
If the input should be represented as an image, the width of the input image. If not specified, it will be
close to the square factor of the `input_size`.
image_height : int
If the input should be represented as an image, the height of the input image. If not specified, it will be
close to the square factor of the `input_size`.
"""
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(GSN, self).__init__(**initial_parameters)
# when the input should be thought of as an image, either use the specified width and height,
# or try to make as square as possible.
if image_height is None and image_width is None:
(_h, _w) = closest_to_square_factors(self.input_size)
self.image_width = _w
self.image_height = _h
else:
self.image_height = image_height
self.image_width = image_width
############################
# Theano variables and RNG #
############################
if self.inputs_hook is None:
self.X = T.matrix('X')
else:
# inputs_hook is a (shape, input) tuple
self.X = self.inputs_hook[1]
##########################
# Network specifications #
##########################
# generally, walkbacks should be at least 2*layers
if layers % 2 == 0:
if walkbacks < 2 * layers:
log.warning(
'Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers', str(layers),
str(walkbacks))
else:
if walkbacks < 2 * layers - 1:
log.warning(
'Not enough walkbacks for the layers! Layers is %s and walkbacks is %s. '
'Generaly want 2X walkbacks to layers', str(layers),
str(walkbacks))
self.add_noise = add_noise
self.noise_annealing = as_floatX(
noise_annealing) # noise schedule parameter
self.hidden_noise_level = sharedX(hidden_noise_level,
dtype=theano.config.floatX)
self.hidden_noise = get_noise(name=hidden_noise,
noise_level=self.hidden_noise_level,
mrg=mrg)
self.input_noise_level = sharedX(input_noise_level,
dtype=theano.config.floatX)
self.input_noise = get_noise(name=input_noise,
noise_level=self.input_noise_level,
mrg=mrg)
self.walkbacks = walkbacks
self.tied_weights = tied_weights
self.layers = layers
self.noiseless_h1 = noiseless_h1
self.input_sampling = input_sampling
self.noise_decay = noise_decay
# if there was a hiddens_hook, unpack the hidden layers in the tensor
if self.hiddens_hook is not None:
hidden_size = self.hiddens_hook[0]
self.hiddens_flag = True
else:
self.hiddens_flag = False
# determine the sizes of each layer in a list.
# layer sizes, from h0 to hK (h0 is the visible layer)
hidden_size = list(raise_to_list(hidden_size))
if len(hidden_size) == 1:
self.layer_sizes = [self.input_size] + hidden_size * self.layers
else:
assert len(hidden_size) == self.layers, "Hiddens sizes and number of hidden layers mismatch." + \
"Hiddens %d and layers %d" % (len(hidden_size), self.layers)
self.layer_sizes = [self.input_size] + hidden_size
if self.hiddens_hook is not None:
self.hiddens = self.unpack_hiddens(self.hiddens_hook[1])
#########################
# Activation functions! #
#########################
# hidden unit activation
self.hidden_activation = get_activation_function(hidden_activation)
# Visible layer activation
self.visible_activation = get_activation_function(visible_activation)
# make sure the sampling functions are appropriate for the activation functions.
if is_binary(self.visible_activation):
self.visible_sampling = mrg.binomial
else:
# TODO: implement non-binary activation
log.error("Non-binary visible activation not supported yet!")
raise NotImplementedError(
"Non-binary visible activation not supported yet!")
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args or dict()
###############
# Parameters! #
###############
# make sure to deal with params_hook!
if self.params_hook is not None:
# if tied weights, expect layers*2 + 1 params
if self.tied_weights:
assert len(self.params_hook) == 2*layers + 1, \
"Tied weights: expected {0!s} params, found {1!s}!".format(2*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:layers]
self.bias_list = self.params_hook[layers:]
# if untied weights, expect layers*3 + 1 params
else:
assert len(self.params_hook) == 3*layers + 1, \
"Untied weights: expected {0!s} params, found {1!s}!".format(3*layers+1, len(self.params_hook))
self.weights_list = self.params_hook[:2 * layers]
self.bias_list = self.params_hook[2 * layers:]
# otherwise, construct our params
else:
# initialize a list of weights and biases based on layer_sizes for the GSN
self.weights_list = [
get_weights(
weights_init=weights_init,
shape=(self.layer_sizes[i], self.layer_sizes[i + 1]),
name="W_{0!s}_{1!s}".format(i, i + 1),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval) for i in range(layers)
]
# add more weights if we aren't tying weights between layers (need to add for higher-lower layers now)
if not tied_weights:
self.weights_list.extend([
get_weights(
weights_init=weights_init,
shape=(self.layer_sizes[i + 1], self.layer_sizes[i]),
name="W_{0!s}_{1!s}".format(i + 1, i),
rng=mrg,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in reversed(range(layers))
])
# initialize each layer bias to 0's.
self.bias_list = [
get_bias(shape=(self.layer_sizes[i], ),
name='b_' + str(i),
init_values=bias_init) for i in range(layers + 1)
]
# build the params of the model into a list
self.params = self.weights_list + self.bias_list
log.debug("gsn params: %s", str(self.params))
# using the properties, build the computational graph
self.cost, self.monitors, self.output, self.hiddens = self.build_computation_graph(
)
def __init__(self,
inputs_hook=None,
hiddens_hook=None,
params_hook=None,
input_size=28 * 28,
hidden_size=1000,
noise_level=0.4,
hidden_activation='tanh',
visible_activation='sigmoid',
cost_function='binary_crossentropy'):
# initialize the Model superclass
super(DenoisingAutoencoder, self).__init__(**{
arg: val
for (arg, val) in locals().iteritems() if arg is not 'self'
})
# Define model hyperparameters
# deal with the inputs_hook and hiddens_hook for the size parameters!
# if the hook exists, grab the size from the first element of the tuple.
if self.inputs_hook is not None:
assert len(self.inputs_hook
) == 2, "Was expecting inputs_hook to be a tuple."
self.input_size = inputs_hook[0]
if self.hiddens_hook is not None:
assert len(self.hiddens_hook
) == 2, "was expecting hiddens_hook to be a tuple."
hidden_size = hiddens_hook[0]
# use the helper methods to grab appropriate activation functions from names!
hidden_activation = get_activation_function(hidden_activation)
visible_activation = get_activation_function(visible_activation)
# do the same for the cost function
cost_function = get_cost_function(cost_function)
# Now, define the symbolic input to the model (Theano)
# We use a matrix rather than a vector so that minibatch processing can be done in parallel.
# Make sure to deal with 'inputs_hook' if it exists!
if self.inputs_hook is not None:
# grab the new input variable from the inputs_hook tuple
x = self.inputs_hook[1]
else:
x = T.matrix("X")
self.inputs = [x]
# Build the model's parameters - a weight matrix and two bias vectors
# Make sure to deal with 'params_hook' if it exists!
if self.params_hook:
# check to see if it contains the three necessary variables
assert len(self.params_hook
) == 3, "Not correct number of params to DAE, needs 3!"
W, b0, b1 = self.params_hook
else:
W = get_weights_uniform(shape=(self.input_size, hidden_size),
name="W")
b0 = get_bias(shape=self.input_size, name="b0")
b1 = get_bias(shape=hidden_size, name="b1")
self.params = [W, b0, b1]
# Perform the computation for a denoising autoencoder!
# first, add noise (corrupt) the input
corrupted_input = salt_and_pepper(input=x, noise_level=noise_level)
# next, run the hidden layer given the inputs (the encoding function)
# We don't need to worry about hiddens_hook during training, because we can't
# run a cost without having the input!
# hiddens_hook is more for the run function and linking methods below.
hiddens = hidden_activation(T.dot(corrupted_input, W) + b1)
# finally, create the reconstruction from the hidden layer (we tie the weights with W.T)
reconstruction = visible_activation(T.dot(hiddens, W.T) + b0)
# the training cost is reconstruction error
self.train_cost = cost_function(output=reconstruction, target=x)
# Compile everything into a Theano function for prediction!
# When using real-world data in predictions, we wouldn't corrupt the input first.
# Therefore, create another version of the hiddens and reconstruction without adding the noise.
# Here is where we would handle hiddens_hook because this is a generative model!
# For the run function, it would take in the hiddens instead of the input variable x.
if self.hiddens_hook is not None:
self.hiddens = self.hiddens_hook[1]
else:
self.hiddens = hidden_activation(T.dot(x, W) + b1)
# make the reconstruction (generated) from the hiddens
self.recon_predict = visible_activation(T.dot(self.hiddens, W.T) + b0)
# now compile the run function accordingly - if it used x or hiddens as the input.
if self.hiddens_hook is not None:
self.f_run = function(inputs=[self.hiddens],
outputs=self.recon_predict)
else:
self.f_run = function(inputs=[x], outputs=self.recon_predict)
def __init__(self,
inputs_hook=None,
hiddens_hook=None,
params_hook=None,
outdir='outputs/gru/',
input_size=None,
hidden_size=None,
output_size=None,
activation='sigmoid',
hidden_activation='relu',
inner_hidden_activation='sigmoid',
mrg=RNG_MRG.MRG_RandomStreams(1),
weights_init='uniform',
weights_interval='montreal',
weights_mean=0,
weights_std=5e-3,
bias_init=0.0,
r_weights_init='identity',
r_weights_interval='montreal',
r_weights_mean=0,
r_weights_std=5e-3,
r_bias_init=0.0,
cost_function='mse',
cost_args=None,
noise='dropout',
noise_level=None,
noise_decay=False,
noise_decay_amount=.99,
forward=True,
clip_recurrent_grads=False):
"""
Initialize a simple recurrent network.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere. For recurrent nets,
this will be the initial starting value for hidden layers.
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters.
outdir : str
The location to produce outputs from training or running the :class:`RNN`. If None, nothing will be saved.
input_size : int
The size (dimensionality) of the input. If shape is provided in `inputs_hook`, this is optional.
hidden_size : int
The size (dimensionality) of the hidden layers. If shape is provided in `hiddens_hook`, this is optional.
output_size : int
The size (dimensionality) of the output.
activation : str or callable
The nonlinear (or linear) activation to perform after the dot product from hiddens -> output layer.
This activation function should be appropriate for the output unit types, i.e. 'sigmoid' for binary.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The activation to perform for the hidden units.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
inner_hidden_activation : str or callable
The activation to perform for the hidden gates.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
mrg : random
A random number generator that is used when adding noise.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
r_weights_init : str
Determines the method for initializing recurrent model weights. See opendeep.utils.nnet for options.
r_weights_interval : str or float
If Uniform `r_weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
r_weights_mean : float
If Gaussian `r_weights_init`, the mean value to use.
r_weights_std : float
If Gaussian `r_weights_init`, the standard deviation to use.
r_bias_init : float
The initial value to use for the recurrent bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the output cost of the model.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
noise : str
What type of noise to use for the hidden layers and outputs. See opendeep.utils.noise
for options. This should be appropriate for the unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
noise_decay : str or False
Whether to use `noise` scheduling (decay `noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the model learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_decay_amount : float
The amount to reduce the `noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
forward : bool
The direction this recurrent model should go over its inputs. True means forward, False mean backward.
clip_recurrent_grads : False or float, optional
Whether to clip the gradients for the parameters that unroll over timesteps (such as the weights
connecting previous hidden states to the current hidden state, and not the weights from current
input to hiddens). If it is a float, the gradients for the weights will be hard clipped to the range
`+-clip_recurrent_grads`.
"""
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(GRU, self).__init__(**initial_parameters)
##################
# specifications #
##################
#########################################
# activation, cost, and noise functions #
#########################################
# recurrent hidden activation function!
self.hidden_activation_func = get_activation_function(
hidden_activation)
self.inner_hidden_activation_func = get_activation_function(
inner_hidden_activation)
# output activation function!
activation_func = get_activation_function(activation)
# Cost function
cost_function = get_cost_function(cost_function)
cost_args = cost_args or dict()
# Now deal with noise if we added it:
if noise:
log.debug('Adding %s noise switch.' % str(noise))
if noise_level is not None:
noise_level = sharedX(value=noise_level)
noise_func = get_noise(noise, noise_level=noise_level, mrg=mrg)
else:
noise_func = get_noise(noise, mrg=mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.noise_switch = sharedX(value=1, name="gru_noise_switch")
# noise scheduling
if noise_decay and noise_level is not None:
self.noise_schedule = get_decay_function(
noise_decay, noise_level, noise_level.get_value(),
noise_decay_amount)
###############
# inputs hook #
###############
# grab info from the inputs_hook
# in the case of an inputs_hook, recurrent will always work with the leading tensor dimension
# being the temporal dimension.
# input is 3D tensor of (timesteps, batch_size, data_dim)
# if input is 2D tensor, assume it is of the form (timesteps, data_dim) i.e. batch_size is 1. Convert to 3D.
# if input is > 3D tensor, assume it is of form (timesteps, batch_size, data...) and flatten to 3D.
if self.inputs_hook is not None:
self.input = self.inputs_hook[1]
if self.input.ndim == 1:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 'x'),
[1, 2])
self.input_size = 1
elif self.input.ndim == 2:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 1), 1)
elif self.input.ndim > 3:
self.input = self.input.flatten(3)
self.input_size = sum(self.input_size)
else:
raise NotImplementedError(
"Recurrent input with %d dimensions not supported!" %
self.input.ndim)
xs = self.input
else:
# Assume input coming from optimizer is (batches, timesteps, data)
# so, we need to reshape to (timesteps, batches, data)
self.input = T.tensor3("Xs")
xs = self.input.dimshuffle(1, 0, 2)
# The target outputs for supervised training - in the form of (batches, timesteps, output) which is
# the same dimension ordering as the expected input from optimizer.
# therefore, we need to swap it like we did to input xs.
self.target = T.tensor3("Ys")
ys = self.target.dimshuffle(1, 0, 2)
################
# hiddens hook #
################
# set an initial value for the recurrent hiddens from hook
if self.hiddens_hook is not None:
h_init = self.hiddens_hook[1]
self.hidden_size = self.hiddens_hook[0]
else:
# deal with h_init after parameters are made (have to make the same size as hiddens that are computed)
self.hidden_size = hidden_size
##################
# for generating #
##################
# symbolic scalar for how many recurrent steps to use during generation from the model
self.n_steps = T.iscalar("generate_n_steps")
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if self.params_hook is not None:
(W_x_z, W_x_r, W_x_h, U_h_z, U_h_r, U_h_h, W_h_y, b_z, b_r, b_h,
b_y) = self.params_hook
recurrent_params = [U_h_z, U_h_r, U_h_h]
# otherwise, construct our params
else:
# all input-to-hidden weights
W_x_z, W_x_r, W_x_h = [
get_weights(
weights_init=weights_init,
shape=(self.input_size, self.hidden_size),
name="W_x_%s" % sub,
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval) for sub in ['z', 'r', 'h']
]
# all hidden-to-hidden weights
U_h_z, U_h_r, U_h_h = [
get_weights(
weights_init=r_weights_init,
shape=(self.hidden_size, self.hidden_size),
name="U_h_%s" % sub,
# if gaussian
mean=r_weights_mean,
std=r_weights_std,
# if uniform
interval=r_weights_interval) for sub in ['z', 'r', 'h']
]
# hidden-to-output weights
W_h_y = get_weights(
weights_init=weights_init,
shape=(self.hidden_size, self.output_size),
name="W_h_y",
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
# biases
b_z, b_r, b_h = [
get_bias(shape=(self.hidden_size, ),
name="b_%s" % sub,
init_values=r_bias_init) for sub in ['z', 'r', 'h']
]
# output bias
b_y = get_bias(shape=(self.output_size, ),
name="b_y",
init_values=bias_init)
# clip gradients if we are doing that
recurrent_params = [U_h_z, U_h_r, U_h_h]
if clip_recurrent_grads:
clip = abs(clip_recurrent_grads)
U_h_z, U_h_r, U_h_h = [
theano.gradient.grad_clip(p, -clip, clip)
for p in recurrent_params
]
# put all the parameters into our list, and make sure it is in the same order as when we try to load
# them from a params_hook!!!
self.params = [W_x_z, W_x_r, W_x_h
] + recurrent_params + [W_h_y, b_z, b_r, b_h, b_y]
# make h_init the right sized tensor
if not self.hiddens_hook:
h_init = T.zeros_like(T.dot(xs[0], W_x_h))
###############
# computation #
###############
# move some computation outside of scan to speed it up!
x_z = T.dot(xs, W_x_z) + b_z
x_r = T.dot(xs, W_x_r) + b_r
x_h = T.dot(xs, W_x_h) + b_h
# now do the recurrent stuff
self.hiddens, self.updates = theano.scan(
fn=self.recurrent_step,
sequences=[x_z, x_r, x_h],
outputs_info=[h_init],
non_sequences=[U_h_z, U_h_r, U_h_h],
go_backwards=not forward,
name="gru_scan",
strict=True)
# add noise (like dropout) if we wanted it!
if noise:
self.hiddens = T.switch(self.noise_switch,
noise_func(input=self.hiddens),
self.hiddens)
# now compute the outputs from the leftover (top level) hiddens
self.output = activation_func(T.dot(self.hiddens, W_h_y) + b_y)
# now to define the cost of the model - use the cost function to compare our output with the target value.
self.cost = cost_function(output=self.output, target=ys, **cost_args)
log.info("Initialized a GRU!")
def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None, outdir='outputs/rnn/',
input_size=None, hidden_size=None, output_size=None,
layers=1,
activation='sigmoid', hidden_activation='relu',
mrg=RNG_MRG.MRG_RandomStreams(1),
weights_init='uniform', weights_interval='montreal', weights_mean=0, weights_std=5e-3,
bias_init=0.0,
r_weights_init='identity', r_weights_interval='montreal', r_weights_mean=0, r_weights_std=5e-3,
r_bias_init=0.0,
cost_function='mse', cost_args=None,
noise='dropout', noise_level=None, noise_decay=False, noise_decay_amount=.99,
direction='forward',
clip_recurrent_grads=False):
"""
Initialize a simple recurrent network.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere. For recurrent nets,
this will be the initial starting value for hidden layers.
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters.
outdir : str
The location to produce outputs from training or running the :class:`RNN`. If None, nothing will be saved.
input_size : int
The size (dimensionality) of the input. If shape is provided in `inputs_hook`, this is optional.
hidden_size : int
The size (dimensionality) of the hidden layers. If shape is provided in `hiddens_hook`, this is optional.
output_size : int
The size (dimensionality) of the output.
layers : int
The number of stacked hidden layers to use.
activation : str or callable
The nonlinear (or linear) activation to perform after the dot product from hiddens -> output layer.
This activation function should be appropriate for the output unit types, i.e. 'sigmoid' for binary.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The activation to perform for the hidden layers.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
mrg : random
A random number generator that is used when adding noise.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
r_weights_init : str
Determines the method for initializing recurrent model weights. See opendeep.utils.nnet for options.
r_weights_interval : str or float
If Uniform `r_weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
r_weights_mean : float
If Gaussian `r_weights_init`, the mean value to use.
r_weights_std : float
If Gaussian `r_weights_init`, the standard deviation to use.
r_bias_init : float
The initial value to use for the recurrent bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the output cost of the model.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
noise : str
What type of noise to use for the hidden layers and outputs. See opendeep.utils.noise
for options. This should be appropriate for the unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
noise_decay : str or False
Whether to use `noise` scheduling (decay `noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the model learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_decay_amount : float
The amount to reduce the `noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
direction : str
The direction this recurrent model should go over its inputs. Can be 'forward', 'backward', or
'bidirectional'. In the case of 'bidirectional', it will make two passes over the sequence,
computing two sets of hiddens and merging them before running through the final decoder.
clip_recurrent_grads : False or float, optional
Whether to clip the gradients for the parameters that unroll over timesteps (such as the weights
connecting previous hidden states to the current hidden state, and not the weights from current
input to hiddens). If it is a float, the gradients for the weights will be hard clipped to the range
`+-clip_recurrent_grads`.
Raises
------
AssertionError
When asserting various properties of input parameters. See error messages.
"""
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(RNN, self).__init__(**initial_parameters)
##################
# specifications #
##################
self.direction = direction
self.bidirectional = (direction == "bidirectional")
self.backward = (direction == "backward")
self.layers = layers
self.noise = noise
self.weights_init = weights_init
self.weights_mean = weights_mean
self.weights_std = weights_std
self.weights_interval = weights_interval
self.r_weights_init = r_weights_init
self.r_weights_mean = r_weights_mean
self.r_weights_std = r_weights_std
self.r_weights_interval = r_weights_interval
self.bias_init = bias_init
self.r_bias_init = r_bias_init
#########################################
# activation, cost, and noise functions #
#########################################
# recurrent hidden activation function!
self.hidden_activation_func = get_activation_function(hidden_activation)
# output activation function!
self.activation_func = get_activation_function(activation)
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args or dict()
# Now deal with noise if we added it:
if self.noise:
log.debug('Adding %s noise switch.' % str(noise))
if noise_level is not None:
noise_level = sharedX(value=noise_level)
self.noise_func = get_noise(noise, noise_level=noise_level, mrg=mrg)
else:
self.noise_func = get_noise(noise, mrg=mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.noise_switch = sharedX(value=1, name="basiclayer_noise_switch")
# noise scheduling
if noise_decay and noise_level is not None:
self.noise_schedule = get_decay_function(noise_decay,
noise_level,
noise_level.get_value(),
noise_decay_amount)
###############
# inputs hook #
###############
# grab info from the inputs_hook
# in the case of an inputs_hook, recurrent will always work with the leading tensor dimension
# being the temporal dimension.
# input is 3D tensor of (timesteps, batch_size, data_dim)
# if input is 2D tensor, assume it is of the form (timesteps, data_dim) i.e. batch_size is 1. Convert to 3D.
# if input is > 3D tensor, assume it is of form (timesteps, batch_size, data...) and flatten to 3D.
if self.inputs_hook is not None:
self.input = self.inputs_hook[1]
if self.input.ndim == 1:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 'x'), [1, 2])
self.input_size = 1
elif self.input.ndim == 2:
self.input = T.unbroadcast(self.input.dimshuffle(0, 'x', 1), 1)
elif self.input.ndim == 3:
pass
elif self.input.ndim > 3:
self.input = self.input.flatten(3)
self.input_size = sum(self.input_size)
else:
raise NotImplementedError("Recurrent input with %d dimensions not supported!" % self.input.ndim)
else:
# Assume input coming from optimizer is (batches, timesteps, data)
# so, we need to reshape to (timesteps, batches, data)
xs = T.tensor3("Xs")
xs = xs.dimshuffle(1, 0, 2)
self.input = xs
# The target outputs for supervised training - in the form of (batches, timesteps, output) which is
# the same dimension ordering as the expected input from optimizer.
# therefore, we need to swap it like we did to input xs.
ys = T.tensor3("Ys")
ys = ys.dimshuffle(1, 0, 2)
self.target = ys
################
# hiddens hook #
################
# set an initial value for the recurrent hiddens from hook
if self.hiddens_hook is not None:
self.h_init = self.hiddens_hook[1]
self.hidden_size = self.hiddens_hook[0]
else:
# deal with h_init after parameters are made (have to make the same size as hiddens that are computed)
self.hidden_size = hidden_size
##################
# for generating #
##################
# symbolic scalar for how many recurrent steps to use during generation from the model
self.n_steps = T.iscalar("generate_n_steps")
self.output, self.hiddens, self.updates, self.cost, self.params = self.build_computation_graph()
def __init__(self, inputs_hook=None, hiddens_hook=None, params_hook=None, outdir=None,
input_size=None, hidden_size=None,
layers=2, walkbacks=4,
visible_activation='sigmoid', hidden_activation='tanh',
input_sampling=True, mrg=RNG_MRG.MRG_RandomStreams(1),
tied_weights=True,
weights_init='uniform', weights_interval='montreal', weights_mean=0, weights_std=5e-3,
bias_init=0,
cost_function='binary_crossentropy', cost_args=None,
add_noise=True, noiseless_h1=True,
hidden_noise='gaussian', hidden_noise_level=2, input_noise='salt_and_pepper', input_noise_level=0.4,
noise_decay='exponential', noise_annealing=1,
image_width=None, image_height=None,
rnn_hidden_size=None, rnn_hidden_activation='rectifier',
rnn_weights_init='identity',
rnn_weights_mean=0, rnn_weights_std=5e-3, rnn_weights_interval='montreal',
rnn_bias_init=0,
generate_n_steps=200):
"""
Initialize an RNN-GSN.
Parameters
----------
inputs_hook : Tuple of (shape, variable)
Routing information for the model to accept inputs from elsewhere. This is used for linking
different models together (e.g. setting the Softmax model's input layer to the DAE's hidden layer gives a
newly supervised classification model). For now, it needs to include the shape information (normally the
dimensionality of the input i.e. n_in).
hiddens_hook : Tuple of (shape, variable)
Routing information for the model to accept its hidden representation from elsewhere.
This is used for linking different models together (e.g. setting the DAE model's hidden layers to the RNN's
output layer gives a generative recurrent model.) For now, it needs to include the shape
information (normally the dimensionality of the hiddens i.e. n_hidden).
params_hook : List(theano shared variable)
A list of model parameters (shared theano variables) that you should use when constructing
this model (instead of initializing your own shared variables). This parameter is useful when you want to
have two versions of the model that use the same parameters - such as a training model with dropout applied
to layers and one without for testing, where the parameters are shared between the two.
outdir : str
The directory you want outputs (parameters, images, etc.) to save to. If None, nothing will
be saved.
input_size : int
The size (dimensionality) of the input to the DAE. If shape is provided in `inputs_hook`, this is optional.
The :class:`Model` requires an `output_size`, which gets set to this value because the DAE is an
unsupervised model. The output is a reconstruction of the input.
hidden_size : int
The size (dimensionality) of the hidden layer for the DAE. Generally, you want it to be larger than
`input_size`, which is known as *overcomplete*.
layers : int
The number of hidden layers to use.
walkbacks : int
The number of walkbacks to perform (the variable K in Bengio's paper above). A walkback is a Gibbs sample
from the DAE, which means the model generates inputs in sequence, where each generated input is compared
to the original input to create the reconstruction cost for training. For running the model, the very last
generated input in the Gibbs chain is used as the output.
visible_activation : str or callable
The nonlinear (or linear) visible activation to perform after the dot product from hiddens -> visible layer.
This activation function should be appropriate for the input unit types, i.e. 'sigmoid' for binary inputs.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
hidden_activation : str or callable
The nonlinear (or linear) hidden activation to perform after the dot product from visible -> hiddens layer.
See opendeep.utils.activation for a list of available activation functions. Alternatively, you can pass
your own function to be used as long as it is callable.
input_sampling : bool
During walkbacks, whether to sample from the generated input to create a new starting point for the next
walkback (next step in the Gibbs chain). This generally makes walkbacks more effective by making the
process more stochastic - more likely to find spurious modes in the model's representation.
mrg : random
A random number generator that is used when adding noise into the network and for sampling from the input.
I recommend using Theano's sandbox.rng_mrg.MRG_RandomStreams.
tied_weights : bool
DAE has two weight matrices - W from input -> hiddens and V from hiddens -> input. This boolean
determines if V = W.T, which 'ties' V to W and reduces the number of parameters necessary during training.
weights_init : str
Determines the method for initializing model weights. See opendeep.utils.nnet for options.
weights_interval : str or float
If Uniform `weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
weights_mean : float
If Gaussian `weights_init`, the mean value to use.
weights_std : float
If Gaussian `weights_init`, the standard deviation to use.
bias_init : float
The initial value to use for the bias parameter. Most often, the default of 0.0 is preferred.
cost_function : str or callable
The function to use when calculating the reconstruction cost of the model. This should be appropriate
for the type of input, i.e. use 'binary_crossentropy' for binary inputs, or 'mse' for real-valued inputs.
See opendeep.utils.cost for options. You can also specify your own function, which needs to be callable.
cost_args : dict
Any additional named keyword arguments to pass to the specified `cost_function`.
add_noise : bool
Whether to add noise (corrupt) the input before passing it through the computation graph during training.
This should most likely be set to the default of True, because this is a *denoising* autoencoder after all.
noiseless_h1 : bool
Whether to not add noise (corrupt) the hidden layer during computation.
hidden_noise : str
What type of noise to use for corrupting the hidden layer (if not `noiseless_h1`). See opendeep.utils.noise
for options. This should be appropriate for the hidden unit activation, i.e. Gaussian for tanh or other
real-valued activations, etc.
hidden_noise_level : float
The amount of noise to use for the noise function specified by `hidden_noise`. This could be the
standard deviation for gaussian noise, the interval for uniform noise, the dropout amount, etc.
input_noise : str
What type of noise to use for corrupting the input before computation (if `add_noise`).
See opendeep.utils.noise for options. This should be appropriate for the input units, i.e. salt-and-pepper
for binary units, etc.
input_noise_level : float
The amount of noise used to corrupt the input. This could be the masking probability for salt-and-pepper,
standard deviation for Gaussian, interval for Uniform, etc.
noise_decay : str or False
Whether to use `input_noise` scheduling (decay `input_noise_level` during the course of training),
and if so, the string input specifies what type of decay to use. See opendeep.utils.decay for options.
Noise decay (known as noise scheduling) effectively helps the DAE learn larger variance features first,
and then smaller ones later (almost as a kind of curriculum learning). May help it converge faster.
noise_annealing : float
The amount to reduce the `input_noise_level` after each training epoch based on the decay function specified
in `noise_decay`.
image_width : int
If the input should be represented as an image, the width of the input image. If not specified, it will be
close to the square factor of the `input_size`.
image_height : int
If the input should be represented as an image, the height of the input image. If not specified, it will be
close to the square factor of the `input_size`.
rnn_hidden_size : int
The number of hidden units (dimensionality) to use in the recurrent layer.
rnn_hidden_activation : str or Callable
The activation function to apply to recurrent units. See opendeep.utils.activation for options.
rnn_weights_init : str
Determines the method for initializing recurrent weights. See opendeep.utils.nnet for options. 'Identity'
works well with 'rectifier' `rnn_hidden_activation`.
rnn_weights_mean : float
If Gaussian `rnn_weights_init`, the mean value to use.
rnn_weights_std : float
If Gaussian `rnn_weights_init`, the standard deviation to use.
rnn_weights_interval : str or float
If Uniform `rnn_weights_init`, the +- interval to use. See opendeep.utils.nnet for options.
rnn_bias_init : float
The initial value to use for the recurrent bias parameter. Most often, the default of 0.0 is preferred.
generate_n_steps : int
When generating from the model, how many steps to generate.
"""
initial_parameters = locals().copy()
initial_parameters.pop('self')
super(RNN_GSN, self).__init__(**initial_parameters)
##################
# specifications #
##################
self.layers = layers
self.walkbacks = walkbacks
self.input_sampling = input_sampling
self.mrg = mrg
self.tied_weights = tied_weights
self.noise_decay = noise_decay
self.noise_annealing = noise_annealing
self.add_noise = add_noise
self.noiseless_h1 = noiseless_h1
self.hidden_noise = hidden_noise
self.hidden_noise_level = hidden_noise_level
self.input_noise = input_noise
self.input_noise_level = input_noise_level
self.image_width = image_width
self.image_height = image_height
# grab info from the inputs_hook, hiddens_hook, or from parameters
if self.inputs_hook is not None: # inputs_hook is a tuple of (Shape, Input)
raise NotImplementedError("Inputs_hook not implemented yet for RNN-GSN")
else:
# make the input a symbolic matrix - a sequence of inputs
self.input = T.matrix('Xs')
# set an initial value for the recurrent hiddens
self.u0 = T.zeros((rnn_hidden_size,))
# make a symbolic vector for the initial recurrent hiddens value to use during generation for the model
self.generate_u0 = T.vector("generate_u0")
# either grab the hidden's desired size from the parameter directly, or copy n_in
self.hidden_size = hidden_size or self.input_size
# deal with hiddens_hook
if self.hiddens_hook is not None:
raise NotImplementedError("Hiddens_hook not implemented yet for RNN-GSN")
# other specifications
# visible activation function!
self.visible_activation_func = get_activation_function(visible_activation)
# make sure the sampling functions are appropriate for the activation functions.
if is_binary(self.visible_activation_func):
self.visible_sampling = mrg.binomial
else:
# TODO: implement non-binary activation
log.error("Non-binary visible activation not supported yet!")
raise NotImplementedError("Non-binary visible activation not supported yet!")
# hidden activation function!
self.hidden_activation_func = get_activation_function(hidden_activation)
# recurrent hidden activation function!
self.rnn_hidden_activation_func = get_activation_function(rnn_hidden_activation)
# Cost function
self.cost_function = get_cost_function(cost_function)
self.cost_args = cost_args
# symbolic scalar for how many recurrent steps to use during generation from the model
self.n_steps = T.iscalar("generate_n_steps")
# determine the sizes of each layer in a list.
# layer sizes, from h0 to hK (h0 is the visible layer)
self.layer_sizes = [self.input_size] + [self.hidden_size] * self.layers
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if self.params_hook is not None:
# if tied weights, expect (layers*2 + 1) params for GSN and (int(layers+1)/int(2) + 3) for RNN
if self.tied_weights:
expected_num = (2*self.layers + 1) + (int(self.layers+1)/2 + 3)
assert len(self.params_hook) == expected_num, \
"Tied weights: expected {0!s} params, found {1!s}!".format(expected_num, len(self.params_hook))
gsn_len = (2*self.layers + 1)
self.weights_list = self.params_hook[:self.layers]
self.bias_list = self.params_hook[self.layers:gsn_len]
# if untied weights, expect layers*3 + 1 params
else:
expected_num = (3*self.layers + 1) + (int(self.layers + 1)/2 + 3)
assert len(self.params_hook) == expected_num, \
"Untied weights: expected {0!s} params, found {1!s}!".format(expected_num, len(self.params_hook))
gsn_len = (3*self.layers + 1)
self.weights_list = self.params_hook[:2*self.layers]
self.bias_list = self.params_hook[2*self.layers:gsn_len]
rnn_len = gsn_len + int(self.layers + 1) / 2
self.recurrent_to_gsn_weights_list = self.params_hook[gsn_len:rnn_len]
self.W_u_u = self.params_hook[rnn_len:rnn_len + 1]
self.W_x_u = self.params_hook[rnn_len + 1:rnn_len + 2]
self.recurrent_bias = self.params_hook[rnn_len + 2:rnn_len + 3]
# otherwise, construct our params
else:
# initialize a list of weights and biases based on layer_sizes for the GSN
self.weights_list = [get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i], self.layer_sizes[i + 1]),
name="W_{0!s}_{1!s}".format(i, i + 1),
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in range(self.layers)]
# add more weights if we aren't tying weights between layers (need to add for higher-lower layers now)
if not self.tied_weights:
self.weights_list.extend(
[get_weights(weights_init=weights_init,
shape=(self.layer_sizes[i + 1], self.layer_sizes[i]),
name="W_{0!s}_{1!s}".format(i + 1, i),
# if gaussian
mean=weights_mean,
std=weights_std,
# if uniform
interval=weights_interval)
for i in reversed(range(self.layers))]
)
# initialize each layer bias to 0's.
self.bias_list = [get_bias(shape=(self.layer_sizes[i],),
name='b_' + str(i),
init_values=bias_init)
for i in range(self.layers + 1)]
self.recurrent_to_gsn_weights_list = [
get_weights(weights_init=rnn_weights_init,
shape=(rnn_hidden_size, self.layer_sizes[layer]),
name="W_u_h{0!s}".format(layer),
# if gaussian
mean=rnn_weights_mean,
std=rnn_weights_std,
# if uniform
interval=rnn_weights_interval)
for layer in range(self.layers + 1) if layer % 2 != 0
]
self.W_u_u = get_weights(weights_init=rnn_weights_init,
shape=(rnn_hidden_size, rnn_hidden_size),
name="W_u_u",
# if gaussian
mean=rnn_weights_mean,
std=rnn_weights_std,
#if uniform
interval=rnn_weights_interval)
self.W_x_u = get_weights(weights_init=rnn_weights_init,
shape=(self.input_size, rnn_hidden_size),
name="W_x_u",
# if gaussian
mean=rnn_weights_mean,
std=rnn_weights_std,
# if uniform
interval=rnn_weights_interval)
self.recurrent_bias = get_bias(shape=(rnn_hidden_size,),
name="b_u",
init_values=rnn_bias_init)
# build the params of the model into a list
self.gsn_params = self.weights_list + self.bias_list
self.params = self.gsn_params + \
self.recurrent_to_gsn_weights_list + \
[self.W_u_u, self.W_x_u, self.recurrent_bias]
log.debug("rnn-gsn params: %s", str(self.params))
# Create the RNN-GSN graph!
self.x_sample, self.cost, self.monitors, self.updates_train, self.x_ts, self.updates_generate, self.u_t = \
self._build_rnngsn()
log.info("Initialized an RNN-GSN!")
def __init__(self, config=None, defaults=_default,
inputs_hook=None, params_hook=None,
input_size=None, output_size=None,
activation=None,
cost=None, cost_args=None,
weights_init=None, weights_mean=None, weights_std=None, weights_interval=None,
bias_init=None,
noise=None, noise_level=None, mrg=None,
outdir=None,
**kwargs):
# init Model to combine the defaults and config dictionaries with the initial parameters.
initial_parameters = locals()
initial_parameters.pop('self')
super(BasicLayer, self).__init__(**initial_parameters)
# all configuration parameters are now in self!
##################
# specifications #
##################
# grab info from the inputs_hook, or from parameters
if self.inputs_hook is not None: # inputs_hook is a tuple of (Shape, Input)
assert len(self.inputs_hook) == 2, 'Expected inputs_hook to be tuple!' # make sure inputs_hook is a tuple
self.input_size = self.inputs_hook[0] or self.input_size
self.input = self.inputs_hook[1]
else:
# make the input a symbolic matrix
self.input = T.fmatrix('X')
# now that we have the input specs, define the output 'target' variable to be used in supervised training!
self.target = T.fmatrix('Y')
# either grab the output's desired size from the parameter directly, or copy n_in
self.output_size = self.output_size or self.input_size
# other specifications
# activation function!
activation_func = get_activation_function(self.activation)
# cost function!
cost_func = get_cost_function(self.cost)
####################################################
# parameters - make sure to deal with params_hook! #
####################################################
if self.params_hook is not None:
# make sure the params_hook has W (weights matrix) and b (bias vector)
assert len(self.params_hook) == 2, \
"Expected 2 params (W and b) for BasicLayer, found {0!s}!".format(len(self.params_hook))
W, b = self.params_hook
else:
W = get_weights(weights_init=self.weights_init,
shape=(self.input_size, self.output_size),
name="W",
# if gaussian
mean=self.weights_mean,
std=self.weights_std,
# if uniform
interval=self.weights_interval)
# grab the bias vector
b = get_bias(shape=self.output_size, name="b", init_values=self.bias_init)
# Finally have the two parameters - weights matrix W and bias vector b. That is all!
self.params = [W, b]
###############
# computation #
###############
# Here is the meat of the computation transforming input -> output
# It simply involves a matrix multiplication of inputs*weights, adding the bias vector, and then passing
# the result through our activation function (normally something nonlinear such as: max(0, output))
self.output = activation_func(T.dot(self.input, W) + b)
# Now deal with noise if we added it:
if self.noise:
log.debug('Adding noise switch.')
if self.noise_level is not None:
noise_func = get_noise(self.noise, self.noise_level, self.mrg)
else:
noise_func = get_noise(self.noise, mrg=self.mrg)
# apply the noise as a switch!
# default to apply noise. this is for the cost and gradient functions to be computed later
# (not sure if the above statement is accurate such that gradient depends on initial value of switch)
self.switch = sharedX(value=1, name="basiclayer_noise_switch")
self.output = T.switch(self.switch,
noise_func(input=self.output),
self.output)
# now to define the cost of the model - use the cost function to compare our output with the target value.
self.cost = cost_func(output=self.output, target=self.target, **self.cost_args)
log.debug("Initialized a basic fully-connected layer with shape %s and activation: %s",
str((self.input_size, self.output_size)), str(self.activation))
