def model(x, y, is_training):
# %% We'll convert our MNIST vector data to a 4-D tensor:
# N x W x H x C
x_tensor = tf.reshape(x, [-1, 28, 28, 1])
# %% We'll use a new method called batch normalization.
# This process attempts to "reduce internal covariate shift"
# which is a fancy way of saying that it will normalize updates for each
# batch using a smoothed version of the batch mean and variance
# The original paper proposes using this before any nonlinearities
h_1 = lrelu(batch_norm(conv2d(x_tensor, 32, name='conv1'),
is_training,
scope='bn1'),
name='lrelu1')
h_2 = lrelu(batch_norm(conv2d(h_1, 64, name='conv2'),
is_training,
scope='bn2'),
name='lrelu2')
h_3 = lrelu(batch_norm(conv2d(h_2, 64, name='conv3'),
is_training,
scope='bn3'),
name='lrelu3')
h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4])
h_4 = linear(h_3_flat, 10)
y_pred = tf.nn.softmax(h_4)
# %% Define loss/eval/training functions
cross_entropy = -tf.reduce_sum(y * tf.log(y_pred))
train_step = tf.train.AdamOptimizer().minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
return [train_step, accuracy]
def encoder(x,
phase_train,
dimensions=[],
filter_sizes=[],
convolutional=False,
activation=tf.nn.relu,
output_activation=tf.nn.sigmoid,
reuse=False):
if convolutional: # transforms 2D tensor
x_tensor = to_tensor(x) # into a 4D (BWHC) one
else:
x_tensor = tf.reshape(tensor=x, shape=[-1, dimensions[0]])
dimensions = dimensions[1:]
current_input = x_tensor
for layer_i, n_output in enumerate(dimensions):
with tf.variable_scope(str(layer_i), reuse=reuse):
if convolutional:
h, W = conv2d(
x=current_input,
n_output=n_output,
k_h=filter_sizes[layer_i], # height, width
k_w=filter_sizes[layer_i], # for conv filters
padding='SAME', # the size of the output remains the same
reuse=reuse)
else:
h, W = linear( # = fully connected
x=current_input,
n_output=n_output,
reuse=reuse)
# before activation, normalize the output
# (can be seen as a similar process as the activation,
# except it's there to make sure the data is 'smooth'
# throughout the network (in the same way as one
# normalizes the input data)
norm = bn.batch_norm(x=h,
phase_train=phase_train,
name='bn',
reuse=reuse)
output = activation(norm)
current_input = output
flattened = flatten(current_input, name='flatten', reuse=reuse)
if output_activation is None:
return flattened
else:
return output_activation(flattened)
def decoder(
z,
phase_train,
dimensions=[],
channels=[],
filter_sizes=[],
convolutional=False, # will be used convolutionally here
activation=tf.nn.relu, # why two different
output_activation=tf.nn.tanh, # activation functions?
reuse=None):
if convolutional:
with tf.variable_scope('fc', reuse=reuse):
z1, W = linear( # check out the generator for an idea of
x=z, # what channels and dimensions look like
n_output=channels[0] * dimensions[0][0] * dimensions[0][1],
reuse=reuse)
rsz = tf.reshape(
z1, [-1, dimensions[0][0], dimensions[0][1], channels[0]])
current_input = activation(features=bn.batch_norm(
name='bn', x=rsz, phase_train=phase_train, reuse=reuse))
dimensions = dimensions[1:]
channels = channels[1:]
filter_sizes = filter_sizes[1:]
else:
current_input = z
for layer_i, n_output in enumerate(dimensions):
with tf.variable_scope(str(layer_i), reuse=reuse):
if convolutional:
h, W = deconv2d(x=current_input,
n_output_h=n_output[0],
n_output_w=n_output[1],
n_output_ch=channels[layer_i],
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i],
padding='SAME',
reuse=reuse)
else:
h, W = linear(x=current_input, n_output=n_output, reuse=reuse)
# applying batch norm to all layers
if layer_i < len(dimensions) - 1:
norm = bn.batch_norm(x=h,
phase_train=phase_train,
name='bn',
reuse=reuse)
output = activation(norm)
else:
output = h
current_input = output
if output_activation is None:
return current_input
else:
return output_activation(current_input)
def VAE(input_shape=[None, 784],
output_shape=[None, 784],
n_filters=[64, 64, 64],
filter_sizes=[4, 4, 4],
n_hidden=32,
n_code=2,
activation=tf.nn.tanh,
dropout=False,
denoising=False,
convolutional=False,
variational=False,
softmax=False,
classifier='alexnet_v2'):
"""(Variational) (Convolutional) (Denoising) Autoencoder.
Uses tied weights.
Parameters
----------
input_shape : list, optional
Shape of the input to the network. e.g. for MNIST: [None, 784].
n_filters : list, optional
Number of filters for each layer.
If convolutional=True, this refers to the total number of output
filters to create for each layer, with each layer's number of output
filters as a list.
If convolutional=False, then this refers to the total number of neurons
for each layer in a fully connected network.
filter_sizes : list, optional
Only applied when convolutional=True. This refers to the ksize (height
and width) of each convolutional layer.
n_hidden : int, optional
Only applied when variational=True. This refers to the first fully
connected layer prior to the variational embedding, directly after
the encoding. After the variational embedding, another fully connected
layer is created with the same size prior to decoding. Set to 0 to
not use an additional hidden layer.
n_code : int, optional
Only applied when variational=True. This refers to the number of
latent Gaussians to sample for creating the inner most encoding.
activation : function, optional
Activation function to apply to each layer, e.g. tf.nn.relu
dropout : bool, optional
Whether or not to apply dropout. If using dropout, you must feed a
value for 'keep_prob', as returned in the dictionary. 1.0 means no
dropout is used. 0.0 means every connection is dropped. Sensible
values are between 0.5-0.8.
denoising : bool, optional
Whether or not to apply denoising. If using denoising, you must feed a
value for 'corrupt_rec', as returned in the dictionary. 1.0 means no
corruption is used. 0.0 means every feature is corrupted. Sensible
values are between 0.5-0.8.
convolutional : bool, optional
Whether or not to use a convolutional network or else a fully connected
network will be created. This effects the n_filters parameter's
meaning.
variational : bool, optional
Whether or not to create a variational embedding layer. This will
create a fully connected layer after the encoding, if `n_hidden` is
greater than 0, then will create a multivariate gaussian sampling
layer, then another fully connected layer. The size of the fully
connected layers are determined by `n_hidden`, and the size of the
sampling layer is determined by `n_code`.
Returns
-------
model : dict
{
'cost': Tensor to optimize.
'Ws': All weights of the encoder.
'x': Input Placeholder
'z': Inner most encoding Tensor (latent features)
'y': Reconstruction of the Decoder
'keep_prob': Amount to keep when using Dropout
'corrupt_rec': Amount to corrupt when using Denoising
'train': Set to True when training/Applies to Batch Normalization.
}
"""
# network input / placeholders for train (bn) and dropout
x = tf.placeholder(tf.float32, input_shape, 'x')
t = tf.placeholder(tf.float32, output_shape, 't')
label = tf.placeholder(tf.int32, [None], 'label')
phase_train = tf.placeholder(tf.bool, name='phase_train')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
corrupt_rec = tf.placeholder(tf.float32, name='corrupt_rec')
corrupt_cls = tf.placeholder(tf.float32, name='corrupt_cls')
# input of the reconstruction network
# np.tanh(2) = 0.964
current_input1 = utils.corrupt(x)*corrupt_rec + x*(1-corrupt_rec) \
if (denoising and phase_train is not None) else x
current_input1.set_shape(x.get_shape())
# 2d -> 4d if convolution
current_input1 = utils.to_tensor(current_input1) \
if convolutional else current_input1
Ws = []
shapes = []
# Build the encoder
for layer_i, n_output in enumerate(n_filters):
with tf.variable_scope('encoder/{}'.format(layer_i)):
shapes.append(current_input1.get_shape().as_list())
if convolutional:
h, W = utils.conv2d(x=current_input1,
n_output=n_output,
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i])
else:
h, W = utils.linear(x=current_input1, n_output=n_output)
h = activation(batch_norm(h, phase_train, 'bn' + str(layer_i)))
if dropout:
h = tf.nn.dropout(h, keep_prob)
Ws.append(W)
current_input1 = h
shapes.append(current_input1.get_shape().as_list())
with tf.variable_scope('variational'):
if variational:
dims = current_input1.get_shape().as_list()
flattened = utils.flatten(current_input1)
if n_hidden:
h = utils.linear(flattened, n_hidden, name='W_fc')[0]
h = activation(batch_norm(h, phase_train, 'fc/bn'))
if dropout:
h = tf.nn.dropout(h, keep_prob)
else:
h = flattened
z_mu = utils.linear(h, n_code, name='mu')[0]
z_log_sigma = 0.5 * utils.linear(h, n_code, name='log_sigma')[0]
# modified by yidawang
# s, u, v = tf.svd(z_log_sigma)
# z_log_sigma = tf.matmul(
# tf.matmul(u, tf.diag(s)), tf.transpose(v))
# end yidawang
# Sample from noise distribution p(eps) ~ N(0, 1)
epsilon = tf.random_normal(tf.stack([tf.shape(x)[0], n_code]))
# Sample from posterior
z = z_mu + tf.multiply(epsilon, tf.exp(z_log_sigma))
if n_hidden:
h = utils.linear(z, n_hidden, name='fc_t')[0]
h = activation(batch_norm(h, phase_train, 'fc_t/bn'))
if dropout:
h = tf.nn.dropout(h, keep_prob)
else:
h = z
size = dims[1] * dims[2] * dims[3] if convolutional else dims[1]
h = utils.linear(h, size, name='fc_t2')[0]
current_input1 = activation(batch_norm(h, phase_train, 'fc_t2/bn'))
if dropout:
current_input1 = tf.nn.dropout(current_input1, keep_prob)
if convolutional:
current_input1 = tf.reshape(
current_input1,
tf.stack([
tf.shape(current_input1)[0], dims[1], dims[2], dims[3]
]))
else:
z = current_input1
shapes.reverse()
n_filters.reverse()
Ws.reverse()
n_filters += [input_shape[-1]]
# %%
# Decoding layers
for layer_i, n_output in enumerate(n_filters[1:]):
with tf.variable_scope('decoder/{}'.format(layer_i)):
shape = shapes[layer_i + 1]
if convolutional:
h, W = utils.deconv2d(x=current_input1,
n_output_h=shape[1],
n_output_w=shape[2],
n_output_ch=shape[3],
n_input_ch=shapes[layer_i][3],
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i])
else:
h, W = utils.linear(x=current_input1, n_output=n_output)
h = activation(batch_norm(h, phase_train, 'dec/bn' + str(layer_i)))
if dropout:
h = tf.nn.dropout(h, keep_prob)
current_input1 = h
y = current_input1
t_flat = utils.flatten(t)
y_flat = utils.flatten(y)
# l2 loss
loss_x = tf.reduce_mean(
tf.reduce_sum(tf.squared_difference(t_flat, y_flat), 1))
loss_z = 0
if variational:
# Variational lower bound, kl-divergence
loss_z = tf.reduce_mean(-0.5 * tf.reduce_sum(
1.0 + 2.0 * z_log_sigma - tf.square(z_mu) -
tf.exp(2.0 * z_log_sigma), 1))
# Add l2 loss
cost_vae = tf.reduce_mean(loss_x + loss_z)
else:
# Just optimize l2 loss
cost_vae = tf.reduce_mean(loss_x)
# Alexnet for clasification based on softmax using TensorFlow slim
if softmax:
axis = list(range(len(x.get_shape())))
mean1, variance1 = tf.nn.moments(t, axis) \
if (phase_train is True) else tf.nn.moments(x, axis)
mean2, variance2 = tf.nn.moments(y, axis)
var_prob = variance2 / variance1
# Input of the classification network
current_input2 = utils.corrupt(x)*corrupt_cls + \
x*(1-corrupt_cls) \
if (denoising and phase_train is True) else x
current_input2.set_shape(x.get_shape())
current_input2 = utils.to_tensor(current_input2) \
if convolutional else current_input2
y_concat = tf.concat([current_input2, y], 3)
with tf.variable_scope('deconv/concat'):
shape = shapes[layer_i + 1]
if convolutional:
# Here we set the input of classification network is
# the twice of
# the input of the reconstruction network
# 112->224 for alexNet and 150->300 for inception v3 and v4
y_concat, W = utils.deconv2d(
x=y_concat,
n_output_h=y_concat.get_shape()[1] * 2,
n_output_w=y_concat.get_shape()[1] * 2,
n_output_ch=y_concat.get_shape()[3],
n_input_ch=y_concat.get_shape()[3],
k_h=3,
k_w=3)
Ws.append(W)
# The following are optional networks for classification network
if classifier == 'squeezenet':
predictions, net = squeezenet.squeezenet(y_concat, num_classes=13)
elif classifier == 'zigzagnet':
predictions, net = squeezenet.zigzagnet(y_concat, num_classes=13)
elif classifier == 'alexnet_v2':
predictions, end_points = alexnet.alexnet_v2(y_concat,
num_classes=13)
elif classifier == 'inception_v1':
predictions, end_points = inception.inception_v1(y_concat,
num_classes=13)
elif classifier == 'inception_v2':
predictions, end_points = inception.inception_v2(y_concat,
num_classes=13)
elif classifier == 'inception_v3':
predictions, end_points = inception.inception_v3(y_concat,
num_classes=13)
label_onehot = tf.one_hot(label, 13, axis=-1, dtype=tf.int32)
cost_s = tf.losses.softmax_cross_entropy(label_onehot, predictions)
cost_s = tf.reduce_mean(cost_s)
acc = tf.nn.in_top_k(predictions, label, 1)
else:
predictions = tf.one_hot(label, 13, 1, 0)
label_onehot = tf.one_hot(label, 13, 1, 0)
cost_s = 0
acc = 0
# Using Summaries for Tensorboard
tf.summary.scalar('cost_vae', cost_vae)
tf.summary.scalar('cost_s', cost_s)
tf.summary.scalar('loss_x', loss_x)
tf.summary.scalar('loss_z', loss_z)
tf.summary.scalar('corrupt_rec', corrupt_rec)
tf.summary.scalar('corrupt_cls', corrupt_cls)
tf.summary.scalar('var_prob', var_prob)
merged = tf.summary.merge_all()
return {
'cost_vae': cost_vae,
'cost_s': cost_s,
'loss_x': loss_x,
'loss_z': loss_z,
'Ws': Ws,
'x': x,
't': t,
'label': label,
'label_onehot': label_onehot,
'predictions': predictions,
'z': z,
'y': y,
'acc': acc,
'keep_prob': keep_prob,
'corrupt_rec': corrupt_rec,
'corrupt_cls': corrupt_cls,
'var_prob': var_prob,
'train': phase_train,
'merged': merged
}
def VAE(input_shape=[None, 784],
n_filters=[64, 64, 64],
filter_sizes=[4, 4, 4],
n_hidden=32,
n_code=2,
activation=tf.nn.tanh,
dropout=False,
denoising=False,
convolutional=False,
variational=False):
"""(Variational) (Convolutional) (Denoising) Autoencoder.
Uses tied weights.
Parameters
----------
input_shape : list, optional
Shape of the input to the network. e.g. for MNIST: [None, 784].
n_filters : list, optional
Number of filters for each layer.
If convolutional=True, this refers to the total number of output
filters to create for each layer, with each layer's number of output
filters as a list.
If convolutional=False, then this refers to the total number of neurons
for each layer in a fully connected network.
filter_sizes : list, optional
Only applied when convolutional=True. This refers to the ksize (height
and width) of each convolutional layer.
n_hidden : int, optional
Only applied when variational=True. This refers to the first fully
connected layer prior to the variational embedding, directly after
the encoding. After the variational embedding, another fully connected
layer is created with the same size prior to decoding. Set to 0 to
not use an additional hidden layer.
n_code : int, optional
Only applied when variational=True. This refers to the number of
latent Gaussians to sample for creating the inner most encoding.
activation : function, optional
Activation function to apply to each layer, e.g. tf.nn.relu
dropout : bool, optional
Whether or not to apply dropout. If using dropout, you must feed a
value for 'keep_prob', as returned in the dictionary. 1.0 means no
dropout is used. 0.0 means every connection is dropped. Sensible
values are between 0.5-0.8.
denoising : bool, optional
Whether or not to apply denoising. If using denoising, you must feed a
value for 'corrupt_prob', as returned in the dictionary. 1.0 means no
corruption is used. 0.0 means every feature is corrupted. Sensible
values are between 0.5-0.8.
convolutional : bool, optional
Whether or not to use a convolutional network or else a fully connected
network will be created. This effects the n_filters parameter's
meaning.
variational : bool, optional
Whether or not to create a variational embedding layer. This will
create a fully connected layer after the encoding, if `n_hidden` is
greater than 0, then will create a multivariate gaussian sampling
layer, then another fully connected layer. The size of the fully
connected layers are determined by `n_hidden`, and the size of the
sampling layer is determined by `n_code`.
Returns
-------
model : dict
{
'cost': Tensor to optimize.
'Ws': All weights of the encoder.
'x': Input Placeholder
'z': Inner most encoding Tensor (latent features)
'y': Reconstruction of the Decoder
'keep_prob': Amount to keep when using Dropout
'corrupt_prob': Amount to corrupt when using Denoising
'train': Set to True when training/Applies to Batch Normalization.
}
"""
# network input / placeholders for train (bn) and dropout
x = tf.placeholder(tf.float32, input_shape, 'x')
phase_train = tf.placeholder(tf.bool, name='phase_train')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
corrupt_prob = tf.placeholder(tf.float32, [1])
# apply noise if denoising
x_ = (utils.corrupt(x) * corrupt_prob + x *
(1 - corrupt_prob)) if denoising else x
# 2d -> 4d if convolution
x_tensor = utils.to_tensor(x_) if convolutional else x_
current_input = x_tensor
Ws = []
shapes = []
# Build the encoder
for layer_i, n_output in enumerate(n_filters):
with tf.variable_scope('encoder/{}'.format(layer_i)):
shapes.append(current_input.get_shape().as_list())
if convolutional:
h, W = utils.conv2d(x=current_input,
n_output=n_output,
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i])
else:
h, W = utils.linear(x=current_input, n_output=n_output)
h = activation(batch_norm(h, phase_train, 'bn' + str(layer_i)))
if dropout:
h = tf.nn.dropout(h, keep_prob)
Ws.append(W)
current_input = h
shapes.append(current_input.get_shape().as_list())
with tf.variable_scope('variational'):
if variational:
dims = current_input.get_shape().as_list()
flattened = utils.flatten(current_input)
if n_hidden:
h = utils.linear(flattened, n_hidden, name='W_fc')[0]
h = activation(batch_norm(h, phase_train, 'fc/bn'))
if dropout:
h = tf.nn.dropout(h, keep_prob)
else:
h = flattened
z_mu = utils.linear(h, n_code, name='mu')[0]
z_log_sigma = 0.5 * utils.linear(h, n_code, name='log_sigma')[0]
# Sample from noise distribution p(eps) ~ N(0, 1)
epsilon = tf.random_normal(tf.stack([tf.shape(x)[0], n_code]))
# Sample from posterior
z = z_mu + tf.multiply(epsilon, tf.exp(z_log_sigma))
if n_hidden:
h = utils.linear(z, n_hidden, name='fc_t')[0]
h = activation(batch_norm(h, phase_train, 'fc_t/bn'))
if dropout:
h = tf.nn.dropout(h, keep_prob)
else:
h = z
size = dims[1] * dims[2] * dims[3] if convolutional else dims[1]
h = utils.linear(h, size, name='fc_t2')[0]
current_input = activation(batch_norm(h, phase_train, 'fc_t2/bn'))
if dropout:
current_input = tf.nn.dropout(current_input, keep_prob)
if convolutional:
current_input = tf.reshape(
current_input,
tf.stack([
tf.shape(current_input)[0], dims[1], dims[2], dims[3]
]))
else:
z = current_input
shapes.reverse()
n_filters.reverse()
Ws.reverse()
n_filters += [input_shape[-1]]
# %%
# Decoding layers
for layer_i, n_output in enumerate(n_filters[1:]):
with tf.variable_scope('decoder/{}'.format(layer_i)):
shape = shapes[layer_i + 1]
if convolutional:
h, W = utils.deconv2d(x=current_input,
n_output_h=shape[1],
n_output_w=shape[2],
n_output_ch=shape[3],
n_input_ch=shapes[layer_i][3],
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i])
else:
h, W = utils.linear(x=current_input, n_output=n_output)
h = activation(batch_norm(h, phase_train, 'dec/bn' + str(layer_i)))
if dropout:
h = tf.nn.dropout(h, keep_prob)
current_input = h
y = current_input
x_flat = utils.flatten(x)
y_flat = utils.flatten(y)
# l2 loss
loss_x = tf.reduce_sum(tf.squared_difference(x_flat, y_flat), 1)
if variational:
# variational lower bound, kl-divergence
loss_z = -0.5 * tf.reduce_sum(
1.0 + 2.0 * z_log_sigma - tf.square(z_mu) -
tf.exp(2.0 * z_log_sigma), 1)
# add l2 loss
cost = tf.reduce_mean(loss_x + loss_z)
else:
# just optimize l2 loss
cost = tf.reduce_mean(loss_x)
return {
'cost': cost,
'Ws': Ws,
'x': x,
'z': z,
'y': y,
'keep_prob': keep_prob,
'corrupt_prob': corrupt_prob,
'train': phase_train
}
def decoder(z,
phase_train,
dimensions=[],
channels=[],
filter_sizes=[],
convolutional=False,
activation=tf.nn.relu,
output_activation=tf.nn.tanh,
reuse=None):
"""Decoder network codes input `x` to layers defined by dimensions.
In contrast with `encoder`, this requires information on the number of
output channels in each layer for convolution. Otherwise, it is mostly
the same.
Parameters
----------
z : tf.Tensor
Input to the decoder network, e.g. tf.Placeholder or tf.Variable
phase_train : tf.Placeholder
Placeholder defining whether the network is in train mode or not.
Used for changing the behavior of batch normalization which updates
its statistics during train mode.
dimensions : list, optional
List of the number of neurons in each layer (convolutional=False) -or-
List of the number of filters in each layer (convolutional=True), e.g.
[100, 100, 100, 100] for a 4-layer deep network with 100 in each layer.
channels : list, optional
For decoding when convolutional=True, require the number of output
channels in each layer.
filter_sizes : list, optional
List of the size of the kernel in each layer, e.g.:
[3, 3, 3, 3] is a 4-layer deep network w/ 3 x 3 kernels in every layer.
convolutional : bool, optional
Whether or not to use convolutional layers.
activation : fn, optional
Function for applying an activation, e.g. tf.nn.relu
output_activation : fn, optional
Function for applying an activation on the last layer, e.g. tf.nn.relu
reuse : bool, optional
For each layer's variable scope, whether to reuse existing variables.
Returns
-------
h : tf.Tensor
Output tensor of the decoder
"""
if convolutional:
with tf.variable_scope('fc', reuse=reuse):
z1, W = linear(
x=z,
n_output=channels[0] * dimensions[0][0] * dimensions[0][1],
reuse=reuse)
rsz = tf.reshape(
z1, [-1, dimensions[0][0], dimensions[0][1], channels[0]])
current_input = activation(
features=bn.batch_norm(
name='bn',
x=rsz,
phase_train=phase_train,
reuse=reuse))
dimensions = dimensions[1:]
channels = channels[1:]
filter_sizes = filter_sizes[1:]
else:
current_input = z
for layer_i, n_output in enumerate(dimensions):
with tf.variable_scope(str(layer_i), reuse=reuse):
if convolutional:
h, W = deconv2d(
x=current_input,
n_output_h=n_output[0],
n_output_w=n_output[1],
n_output_ch=channels[layer_i],
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i],
padding='SAME',
reuse=reuse)
else:
h, W = linear(
x=current_input,
n_output=n_output,
reuse=reuse)
if layer_i < len(dimensions) - 1:
norm = bn.batch_norm(
x=h,
phase_train=phase_train,
name='bn', reuse=reuse)
output = activation(norm)
else:
output = h
current_input = output
if output_activation is None:
return current_input
else:
return output_activation(current_input)
def encoder(x, phase_train, dimensions=[], filter_sizes=[],
convolutional=False, activation=tf.nn.relu,
output_activation=tf.nn.sigmoid, reuse=False):
"""Encoder network codes input `x` to layers defined by dimensions.
Parameters
----------
x : tf.Tensor
Input to the encoder network, e.g. tf.Placeholder or tf.Variable
phase_train : tf.Placeholder
Placeholder defining whether the network is in train mode or not.
Used for changing the behavior of batch normalization which updates
its statistics during train mode.
dimensions : list, optional
List of the number of neurons in each layer (convolutional=False) -or-
List of the number of filters in each layer (convolutional=True), e.g.
[100, 100, 100, 100] for a 4-layer deep network with 100 in each layer.
filter_sizes : list, optional
List of the size of the kernel in each layer, e.g.:
[3, 3, 3, 3] is a 4-layer deep network w/ 3 x 3 kernels in every layer.
convolutional : bool, optional
Whether or not to use convolutional layers.
activation : fn, optional
Function for applying an activation, e.g. tf.nn.relu
output_activation : fn, optional
Function for applying an activation on the last layer, e.g. tf.nn.relu
reuse : bool, optional
For each layer's variable scope, whether to reuse existing variables.
Returns
-------
h : tf.Tensor
Output tensor of the encoder
"""
# %%
# ensure 2-d is converted to square tensor.
if convolutional:
x_tensor = to_tensor(x)
else:
x_tensor = tf.reshape(
tensor=x,
shape=[-1, dimensions[0]])
dimensions = dimensions[1:]
current_input = x_tensor
for layer_i, n_output in enumerate(dimensions):
with tf.variable_scope(str(layer_i), reuse=reuse):
if convolutional:
h, W = conv2d(
x=current_input,
n_output=n_output,
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i],
padding='SAME',
reuse=reuse)
else:
h, W = linear(
x=current_input,
n_output=n_output,
reuse=reuse)
norm = bn.batch_norm(
x=h,
phase_train=phase_train,
name='bn',
reuse=reuse)
output = activation(norm)
current_input = output
flattened = flatten(current_input, name='flatten', reuse=reuse)
if output_activation is None:
return flattened
else:
return output_activation(flattened)
def decoder(z,
phase_train,
dimensions=[],
channels=[],
filter_sizes=[],
convolutional=False,
activation=tf.nn.relu,
output_activation=tf.nn.tanh,
reuse=None):
"""Decoder network codes input `x` to layers defined by dimensions.
In contrast with `encoder`, this requires information on the number of
output channels in each layer for convolution. Otherwise, it is mostly
the same.
Parameters
----------
z : tf.Tensor
Input to the decoder network, e.g. tf.Placeholder or tf.Variable
phase_train : tf.Placeholder
Placeholder defining whether the network is in train mode or not.
Used for changing the behavior of batch normalization which updates
its statistics during train mode.
dimensions : list, optional
List of the number of neurons in each layer (convolutional=False) -or-
List of the number of filters in each layer (convolutional=True), e.g.
[100, 100, 100, 100] for a 4-layer deep network with 100 in each layer.
channels : list, optional
For decoding when convolutional=True, require the number of output
channels in each layer.
filter_sizes : list, optional
List of the size of the kernel in each layer, e.g.:
[3, 3, 3, 3] is a 4-layer deep network w/ 3 x 3 kernels in every layer.
convolutional : bool, optional
Whether or not to use convolutional layers.
activation : fn, optional
Function for applying an activation, e.g. tf.nn.relu
output_activation : fn, optional
Function for applying an activation on the last layer, e.g. tf.nn.relu
reuse : bool, optional
For each layer's variable scope, whether to reuse existing variables.
Returns
-------
h : tf.Tensor
Output tensor of the decoder
"""
if convolutional:
with tf.variable_scope('fc', reuse=reuse):
z1, W = linear(x=z,
n_output=channels[0] * dimensions[0][0] *
dimensions[0][1],
reuse=reuse)
rsz = tf.reshape(
z1, [-1, dimensions[0][0], dimensions[0][1], channels[0]])
current_input = activation(features=bn.batch_norm(
name='bn', x=rsz, phase_train=phase_train, reuse=reuse))
dimensions = dimensions[1:]
channels = channels[1:]
filter_sizes = filter_sizes[1:]
else:
current_input = z
for layer_i, n_output in enumerate(dimensions):
with tf.variable_scope(str(layer_i), reuse=reuse):
if convolutional:
h, W = deconv2d(x=current_input,
n_output_h=n_output[0],
n_output_w=n_output[1],
n_output_ch=channels[layer_i],
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i],
padding='SAME',
reuse=reuse)
else:
h, W = linear(x=current_input, n_output=n_output, reuse=reuse)
if layer_i < len(dimensions) - 1:
norm = bn.batch_norm(x=h,
phase_train=phase_train,
name='bn',
reuse=reuse)
output = activation(norm)
else:
output = h
current_input = output
if output_activation is None:
return current_input
else:
return output_activation(current_input)
def encoder(x,
phase_train,
dimensions=[],
filter_sizes=[],
convolutional=False,
activation=tf.nn.relu,
output_activation=tf.nn.sigmoid,
reuse=False):
"""Encoder network codes input `x` to layers defined by dimensions.
Parameters
----------
x : tf.Tensor
Input to the encoder network, e.g. tf.Placeholder or tf.Variable
phase_train : tf.Placeholder
Placeholder defining whether the network is in train mode or not.
Used for changing the behavior of batch normalization which updates
its statistics during train mode.
dimensions : list, optional
List of the number of neurons in each layer (convolutional=False) -or-
List of the number of filters in each layer (convolutional=True), e.g.
[100, 100, 100, 100] for a 4-layer deep network with 100 in each layer.
filter_sizes : list, optional
List of the size of the kernel in each layer, e.g.:
[3, 3, 3, 3] is a 4-layer deep network w/ 3 x 3 kernels in every layer.
convolutional : bool, optional
Whether or not to use convolutional layers.
activation : fn, optional
Function for applying an activation, e.g. tf.nn.relu
output_activation : fn, optional
Function for applying an activation on the last layer, e.g. tf.nn.relu
reuse : bool, optional
For each layer's variable scope, whether to reuse existing variables.
Returns
-------
h : tf.Tensor
Output tensor of the encoder
"""
# %%
# ensure 2-d is converted to square tensor.
if convolutional:
x_tensor = to_tensor(x)
else:
x_tensor = tf.reshape(tensor=x, shape=[-1, dimensions[0]])
dimensions = dimensions[1:]
current_input = x_tensor
for layer_i, n_output in enumerate(dimensions):
with tf.variable_scope(str(layer_i), reuse=reuse):
if convolutional:
h, W = conv2d(x=current_input,
n_output=n_output,
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i],
padding='SAME',
reuse=reuse)
else:
h, W = linear(x=current_input, n_output=n_output, reuse=reuse)
norm = bn.batch_norm(x=h,
phase_train=phase_train,
name='bn',
reuse=reuse)
output = activation(norm)
current_input = output
flattened = flatten(current_input, name='flatten', reuse=reuse)
if output_activation is None:
return flattened
else:
return output_activation(flattened)
is_training = tf.placeholder(tf.bool, name='is_training')
# %% We'll convert our MNIST vector data to a 4-D tensor:
# N x W x H x C
x_tensor = tf.reshape(x, [-1, 28, 28, 1])
#ema.apply([batch_mean, batch_var])
# %% We'll use a new method called batch normalization.
# This process attempts to "reduce internal covariate shift"
# which is a fancy way of saying that it will normalize updates for each
# batch using a smoothed version of the batch mean and variance
'''
# The original paper proposes using this before any nonlinearities!!!!!!!!!!!!!!!
'''
# The original paper proposes using this before any nonlinearities!!!!!!!!!!!!!!!
h_1 = lrelu(batch_norm(conv2d(x_tensor, 32, name='conv1'),
is_training,
scope='bn1'),
name='lrelu1')
h_2 = lrelu(batch_norm(conv2d(h_1, 64, name='conv2'), is_training,
scope='bn2'),
name='lrelu2')
h_3 = lrelu(batch_norm(conv2d(h_2, 64, name='conv3'), is_training,
scope='bn3'),
name='lrelu3')
h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4])
h_4 = linear(h_3_flat, 10)
y_pred = tf.nn.softmax(h_4)
# %% Define loss/eval/training functions
cross_entropy = -tf.reduce_sum(y * tf.log(y_pred))
train_step = tf.train.AdamOptimizer().minimize(cross_entropy)
# %% We add a new type of placeholder to denote when we are training.
# This will be used to change the way we compute the network during
# training/testing.
is_training = tf.placeholder(tf.bool, name='is_training')
# %% We'll convert our MNIST vector data to a 4-D tensor:
# N x W x H x C
x_tensor = tf.reshape(x, [-1, 28, 28, 1])
# %% We'll use a new method called batch normalization.
# This process attempts to "reduce internal covariate shift"
# which is a fancy way of saying that it will normalize updates for each
# batch using a smoothed version of the batch mean and variance
# The original paper proposes using this before any nonlinearities
h_1 = lrelu(batch_norm(conv2d(x_tensor, 32, name='conv1'),
is_training, scope='bn1'), name='lrelu1')
h_2 = lrelu(batch_norm(conv2d(h_1, 64, name='conv2'),
is_training, scope='bn2'), name='lrelu2')
h_3 = lrelu(batch_norm(conv2d(h_2, 64, name='conv3'),
is_training, scope='bn3'), name='lrelu3')
h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4])
h_4 = linear(h_3_flat, 10)
y_pred = tf.nn.softmax(h_4)
# %% Define loss/eval/training functions
cross_entropy = -tf.reduce_sum(y * tf.log(y_pred))
train_step = tf.train.AdamOptimizer().minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
def VAE(input_shape=[None, 784],
n_filters=[64, 64, 64],
filter_sizes=[4, 4, 4],
n_hidden=32,
n_code=2,
activation=tf.nn.tanh,
dropout=False,
denoising=False,
convolutional=False,
variational=False):
"""(Variational) (Convolutional) (Denoising) Autoencoder.
Uses tied weights.
Parameters
----------
input_shape : list, optional
Shape of the input to the network. e.g. for MNIST: [None, 784].
n_filters : list, optional
Number of filters for each layer.
If convolutional=True, this refers to the total number of output
filters to create for each layer, with each layer's number of output
filters as a list.
If convolutional=False, then this refers to the total number of neurons
for each layer in a fully connected network.
filter_sizes : list, optional
Only applied when convolutional=True. This refers to the ksize (height
and width) of each convolutional layer.
n_hidden : int, optional
Only applied when variational=True. This refers to the first fully
connected layer prior to the variational embedding, directly after
the encoding. After the variational embedding, another fully connected
layer is created with the same size prior to decoding. Set to 0 to
not use an additional hidden layer.
n_code : int, optional
Only applied when variational=True. This refers to the number of
latent Gaussians to sample for creating the inner most encoding.
activation : function, optional
Activation function to apply to each layer, e.g. tf.nn.relu
dropout : bool, optional
Whether or not to apply dropout. If using dropout, you must feed a
value for 'keep_prob', as returned in the dictionary. 1.0 means no
dropout is used. 0.0 means every connection is dropped. Sensible
values are between 0.5-0.8.
denoising : bool, optional
Whether or not to apply denoising. If using denoising, you must feed a
value for 'corrupt_prob', as returned in the dictionary. 1.0 means no
corruption is used. 0.0 means every feature is corrupted. Sensible
values are between 0.5-0.8.
convolutional : bool, optional
Whether or not to use a convolutional network or else a fully connected
network will be created. This effects the n_filters parameter's
meaning.
variational : bool, optional
Whether or not to create a variational embedding layer. This will
create a fully connected layer after the encoding, if `n_hidden` is
greater than 0, then will create a multivariate gaussian sampling
layer, then another fully connected layer. The size of the fully
connected layers are determined by `n_hidden`, and the size of the
sampling layer is determined by `n_code`.
Returns
-------
model : dict
{
'cost': Tensor to optimize.
'Ws': All weights of the encoder.
'x': Input Placeholder
'z': Inner most encoding Tensor (latent features)
'y': Reconstruction of the Decoder
'keep_prob': Amount to keep when using Dropout
'corrupt_prob': Amount to corrupt when using Denoising
'train': Set to True when training/Applies to Batch Normalization.
}
"""
# network input / placeholders for train (bn) and dropout
x = tf.placeholder(tf.float32, input_shape, 'x')
phase_train = tf.placeholder(tf.bool, name='phase_train')
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
corrupt_prob = tf.placeholder(tf.float32, [1])
# apply noise if denoising
x_ = (utils.corrupt(x) * corrupt_prob + x * (1 - corrupt_prob)) if denoising else x
# 2d -> 4d if convolution
x_tensor = utils.to_tensor(x_) if convolutional else x_
current_input = x_tensor
Ws = []
shapes = []
# Build the encoder
for layer_i, n_output in enumerate(n_filters):
with tf.variable_scope('encoder/{}'.format(layer_i)):
shapes.append(current_input.get_shape().as_list())
if convolutional:
h, W = utils.conv2d(x=current_input,
n_output=n_output,
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i])
else:
h, W = utils.linear(x=current_input,
n_output=n_output)
h = activation(batch_norm(h, phase_train, 'bn' + str(layer_i)))
if dropout:
h = tf.nn.dropout(h, keep_prob)
Ws.append(W)
current_input = h
shapes.append(current_input.get_shape().as_list())
with tf.variable_scope('variational'):
if variational:
dims = current_input.get_shape().as_list()
flattened = utils.flatten(current_input)
if n_hidden:
h = utils.linear(flattened, n_hidden, name='W_fc')[0]
h = activation(batch_norm(h, phase_train, 'fc/bn'))
if dropout:
h = tf.nn.dropout(h, keep_prob)
else:
h = flattened
z_mu = utils.linear(h, n_code, name='mu')[0]
z_log_sigma = 0.5 * utils.linear(h, n_code, name='log_sigma')[0]
# Sample from noise distribution p(eps) ~ N(0, 1)
epsilon = tf.random_normal(
tf.stack([tf.shape(x)[0], n_code]))
# Sample from posterior
z = z_mu + tf.multiply(epsilon, tf.exp(z_log_sigma))
if n_hidden:
h = utils.linear(z, n_hidden, name='fc_t')[0]
h = activation(batch_norm(h, phase_train, 'fc_t/bn'))
if dropout:
h = tf.nn.dropout(h, keep_prob)
else:
h = z
size = dims[1] * dims[2] * dims[3] if convolutional else dims[1]
h = utils.linear(h, size, name='fc_t2')[0]
current_input = activation(batch_norm(h, phase_train, 'fc_t2/bn'))
if dropout:
current_input = tf.nn.dropout(current_input, keep_prob)
if convolutional:
current_input = tf.reshape(
current_input, tf.stack([
tf.shape(current_input)[0],
dims[1],
dims[2],
dims[3]]))
else:
z = current_input
shapes.reverse()
n_filters.reverse()
Ws.reverse()
n_filters += [input_shape[-1]]
# %%
# Decoding layers
for layer_i, n_output in enumerate(n_filters[1:]):
with tf.variable_scope('decoder/{}'.format(layer_i)):
shape = shapes[layer_i + 1]
if convolutional:
h, W = utils.deconv2d(x=current_input,
n_output_h=shape[1],
n_output_w=shape[2],
n_output_ch=shape[3],
n_input_ch=shapes[layer_i][3],
k_h=filter_sizes[layer_i],
k_w=filter_sizes[layer_i])
else:
h, W = utils.linear(x=current_input,
n_output=n_output)
h = activation(batch_norm(h, phase_train, 'dec/bn' + str(layer_i)))
if dropout:
h = tf.nn.dropout(h, keep_prob)
current_input = h
y = current_input
x_flat = utils.flatten(x)
y_flat = utils.flatten(y)
# l2 loss
loss_x = tf.reduce_sum(tf.squared_difference(x_flat, y_flat), 1)
if variational:
# variational lower bound, kl-divergence
loss_z = -0.5 * tf.reduce_sum(
1.0 + 2.0 * z_log_sigma -
tf.square(z_mu) - tf.exp(2.0 * z_log_sigma), 1)
# add l2 loss
cost = tf.reduce_mean(loss_x + loss_z)
else:
# just optimize l2 loss
cost = tf.reduce_mean(loss_x)
return {'cost': cost, 'Ws': Ws,
'x': x, 'z': z, 'y': y,
'keep_prob': keep_prob,
'corrupt_prob': corrupt_prob,
'train': phase_train}
def inference(x, y, keepProb, is_training, batch_size, learning_rate):
"""
Args:
images: Images returned from distorted_inputs() or inputs().
Returns:
Logits.
"""
with tf.variable_scope('conv1') as scope:
kernel = tf.Variable(tf.random_normal([11, 11, 3, 48], stddev=1e-4),
name='weights')
conv = tf.nn.conv2d(x, kernel, [1, 4, 4, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.001, dtype=tf.float32, shape=[48]),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv1 = batch_norm(lrelu(bias, name=scope.name),
is_training,
scope='bn1')
# pool1
pool1 = tf.nn.max_pool(conv1,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool1')
# conv2
with tf.variable_scope('conv2') as scope:
kernel = tf.Variable(tf.random_normal([5, 5, 48, 128], stddev=1e-4),
name='weights')
conv = tf.nn.conv2d(pool1, kernel, [1, 1, 1, 1],
padding='SAME') #use_cudnn_on_gpu=False,
biases = tf.Variable(tf.constant(0.001, dtype=tf.float32, shape=[128]),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv2 = batch_norm(lrelu(bias, name=scope.name),
is_training,
scope='bn2')
# pool2
pool2 = tf.nn.max_pool(conv2,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool2')
# conv3
with tf.variable_scope('conv3') as scope:
kernel = tf.Variable(tf.random_normal([3, 3, 128, 192], stddev=1e-2),
name='weights')
conv = tf.nn.conv2d(pool2, kernel, [1, 1, 1, 1], padding='SAME')
biases = tf.Variable(tf.constant(0.001, dtype=tf.float32, shape=[192]),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv3 = lrelu(bias, name=scope.name)
# conv4
with tf.variable_scope('conv4') as scope:
kernel = tf.Variable(tf.random_normal([3, 3, 192, 192], stddev=1e-2),
name='weights')
conv = tf.nn.conv2d(conv3, kernel, [1, 1, 1, 1],
padding='SAME') #use_cudnn_on_gpu=False,
biases = tf.Variable(tf.constant(0.001, dtype=tf.float32, shape=[192]),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv4 = lrelu(bias, name=scope.name)
# conv5
with tf.variable_scope('conv5') as scope:
kernel = tf.Variable(tf.random_normal([5, 5, 192, 128], stddev=1e-2),
name='weights')
conv = tf.nn.conv2d(conv4, kernel, [1, 1, 1, 1],
padding='SAME') #use_cudnn_on_gpu=False,
biases = tf.Variable(tf.constant(0.001, dtype=tf.float32, shape=[128]),
name='biases')
bias = tf.nn.bias_add(conv, biases)
conv5 = lrelu(bias, name=scope.name)
pool3 = tf.nn.max_pool(conv5,
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
# local1
with tf.variable_scope('local1') as scope:
# Move everything into depth so we can perform a single matrix multiply.
dim = 1
for d in pool3.get_shape()[1:].as_list():
dim *= d
reshape = tf.reshape(pool3, [batch_size, dim])
weights = tf.Variable(tf.random_normal([dim, 4096],
stddev=np.sqrt(1 / dim + 4096)),
name='weights')
biases = tf.Variable(tf.constant(0.0001,
dtype=tf.float32,
shape=[4096]),
name='biases')
local1 = tf.nn.relu_layer(reshape, weights, biases)
local1 = tf.nn.dropout(local1, keepProb)
# local2
with tf.variable_scope('local2') as scope:
weights = tf.Variable(tf.random_normal([4096, 4096],
stddev=np.sqrt(1 / 4096.0)),
name='weights')
biases = tf.Variable(tf.constant(0.0001,
dtype=tf.float32,
shape=[4096]),
name='biases')
local2 = tf.nn.relu_layer(local1, weights, biases)
local2 = tf.nn.dropout(local2, keepProb)
with tf.variable_scope('softmax_linear') as scope:
weights = tf.Variable(tf.random_normal([4096, 20],
stddev=np.sqrt(1 / 4096.0)),
name='weights')
biases = tf.Variable(tf.constant(0.01, dtype=tf.float32, shape=[20]),
name='biases')
softmax_linear = tf.nn.xw_plus_b(local2, weights, biases)
with tf.name_scope("loss"):
loss = tf.nn.l2_loss((y - softmax_linear), name="L2_loss")
with tf.name_scope("train_step"):
train_op = tf.train.RMSPropOptimizer(learning_rate).minimize(loss)
return train_op, softmax_linear, loss
# N x W x H x C
x_tensor = tf.reshape(x, [-1, 1, 26, 1]) # FOR MFCC-26 dim
#x_tensor = tf.reshape(x, [-1, 1, 40, 1]) # FOR CONVAE
# %% We'll use a new method called batch normalization.
# This process attempts to "reduce internal covariate shift"
# which is a fancy way of saying that it will normalize updates for each
# batch using a smoothed version of the batch mean and variance
# The original paper proposes using this before any nonlinearities
h_1 = lrelu(batch_norm(conv2d(x_tensor,
32,
name='conv1',
stride_h=1,
k_h=1,
k_w=3,
pool_size=[1, 1, 2, 1],
pool_stride=[1, 1, 1, 1]),
phase_train=is_training,
scope='bn1'),
name='lrelu1')
h_2 = lrelu(batch_norm(conv2d(h_1,
64,
name='conv2',
stride_h=1,
k_h=1,
k_w=3,
pool_size=[1, 1, 2, 1],
pool_stride=[1, 1, 1, 1]),
phase_train=is_training,
import tensorflow as tf from libs.batch_norm import batch_norm from libs.activations import lrelu from libs.connections import conv2d, linear from libs.datasets import MNIST # %% Setup input to the network and true output label. These are # simply placeholders which we'll fill in later. mnist = MNIST() x = tf.placeholder(tf.float32, [None, 784]) y = tf.placeholder(tf.float32, [None, 10]) x_tensor = tf.reshape(x, [-1, 28, 28, 1]) # %% Define the network: bn1 = batch_norm(-1, name='bn1') bn2 = batch_norm(-1, name='bn2') bn3 = batch_norm(-1, name='bn3') h_1 = lrelu(bn1(conv2d(x_tensor, 32, name='conv1')), name='lrelu1') h_2 = lrelu(bn2(conv2d(h_1, 64, name='conv2')), name='lrelu2') h_3 = lrelu(bn3(conv2d(h_2, 64, name='conv3')), name='lrelu3') h_3_flat = tf.reshape(h_3, [-1, 64 * 4 * 4]) h_4 = linear(h_3_flat, 10) y_pred = tf.nn.softmax(h_4) # %% Define loss/eval/training functions cross_entropy = -tf.reduce_sum(y * tf.log(y_pred)) train_step = tf.train.AdamOptimizer().minimize(cross_entropy) correct_prediction = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
