def test_max_norm(self):
array = get_example_array()
for m in get_test_values():
norm_instance = constraints.max_norm(m)
normed = norm_instance(backend.variable(array))
assert np.all(backend.eval(normed) < m)
# a more explicit example
norm_instance = constraints.max_norm(2.0)
x = np.array([[0, 0, 0], [1.0, 0, 0], [3, 0, 0], [3, 3, 3]]).T
x_normed_target = np.array(
[[0, 0, 0], [1.0, 0, 0], [2.0, 0, 0],
[2. / np.sqrt(3), 2. / np.sqrt(3), 2. / np.sqrt(3)]]).T
x_normed_actual = backend.eval(norm_instance(backend.variable(x)))
self.assertAllClose(x_normed_actual, x_normed_target, rtol=1e-05)
def load(architecture,
activation_fn,
optimizer,
learning_rate,
input_size,
output_size,
max_norm_weights=False):
input_layer = keras.layers.Input((input_size, ))
clayer = input_layer
for n in architecture:
clayer = keras.layers.Dense(
n,
activation=activation_fn_map[activation_fn],
kernel_initializer=keras.initializers.TruncatedNormal(
mean=0.0, stddev=1 / np.sqrt(float(n)), seed=None),
bias_initializer='zeros',
kernel_constraint=(max_norm(max_value=float(max_norm_weights))
if max_norm_weights else None))(clayer)
output_layer = keras.layers.Dense(output_size,
activation='softmax')(clayer)
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
optimizer = optimizer_map[optimizer](learning_rate=learning_rate)
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
return model
def _add_generator_block(old_model, config):
# get the end of the last block
block_end = old_model.layers[-2].output
# weights init
w_init = RandomNormal(stddev=0.02)
w_const = max_norm(1.0)
# upsample, and define new block
upsampling = UpSampling2D()(block_end)
# conv layers
x = upsampling
for i, strides in enumerate([3, 3]):
x = Conv2D(config['filters'],
strides,
padding='same',
kernel_initializer=w_init,
kernel_constraint=w_const)(x)
x = PixelNormalization()(x)
x = LeakyReLU(alpha=0.2)(x)
# add new output layer
out_image = Conv2D(config['n_channels'], 1, padding='same')(x)
# define model
model1 = Model(old_model.input, out_image)
# get the output layer from old model
out_old = old_model.layers[-1]
# connect the upsampling to the old output layer
out_image2 = out_old(upsampling)
# define new output image as the weighted sum of the old and new models
merged = WeightedSum()([out_image2, out_image])
# define model
model2 = Model(old_model.input, merged)
return [model1, model2]
def EEGNet_SSVEP(nb_classes = 12, Chans = 8, Samples = 256,
dropoutRate = 0.5, kernLength = 256, F1 = 96,
D = 1, F2 = 96, dropoutType = 'Dropout'):
""" SSVEP Variant of EEGNet, as used in [1].
Inputs:
nb_classes : int, number of classes to classify
Chans, Samples : number of channels and time points in the EEG data
dropoutRate : dropout fraction
kernLength : length of temporal convolution in first layer
F1, F2 : number of temporal filters (F1) and number of pointwise
filters (F2) to learn.
D : number of spatial filters to learn within each temporal
convolution.
dropoutType : Either SpatialDropout2D or Dropout, passed as a string.
[1]. Waytowich, N. et. al. (2018). Compact Convolutional Neural Networks
for Classification of Asynchronous Steady-State Visual Evoked Potentials.
Journal of Neural Engineering vol. 15(6).
http://iopscience.iop.org/article/10.1088/1741-2552/aae5d8
"""
if dropoutType == 'SpatialDropout2D':
dropoutType = SpatialDropout2D
elif dropoutType == 'Dropout':
dropoutType = Dropout
else:
raise ValueError('dropoutType must be one of SpatialDropout2D '
'or Dropout, passed as a string.')
input1 = Input(shape = (1, Chans, Samples))
##################################################################
block1 = Conv2D(F1, (1, kernLength), padding = 'same',
input_shape = (1, Chans, Samples),
use_bias = False)(input1)
block1 = BatchNormalization(axis = 1)(block1)
block1 = DepthwiseConv2D((Chans, 1), use_bias = False,
depth_multiplier = D,
depthwise_constraint = max_norm(1.))(block1)
block1 = BatchNormalization(axis = 1)(block1)
block1 = Activation('elu')(block1)
block1 = AveragePooling2D((1, 4))(block1)
block1 = dropoutType(dropoutRate)(block1)
block2 = SeparableConv2D(F2, (1, 16),
use_bias = False, padding = 'same')(block1)
block2 = BatchNormalization(axis = 1)(block2)
block2 = Activation('elu')(block2)
block2 = AveragePooling2D((1, 8))(block2)
block2 = dropoutType(dropoutRate)(block2)
flatten = Flatten(name = 'flatten')(block2)
dense = Dense(nb_classes, name = 'dense')(flatten)
softmax = Activation('softmax', name = 'softmax')(dense)
return Model(inputs=input1, outputs=softmax)
def _add_discriminator_block(old_model, config):
# new shape is double the size of previous one
old_input_shape = list(old_model.input.shape)
new_input_shape = (old_input_shape[-2] * 2, old_input_shape[-2] * 2,
old_input_shape[-1])
model_input = Input(shape=new_input_shape, name="doubled_dis_input")
# weights init
w_init = RandomNormal(stddev=0.02)
w_const = max_norm(1.0)
# conv layers
x = model_input
for strides in [1, 3, 3]:
x = Conv2D(config['filters'],
strides,
padding='same',
kernel_initializer=w_init,
kernel_constraint=w_const)(x)
x = LeakyReLU()(x)
x = AveragePooling2D()(x)
new_block = x
# skip the input, 1x1 and activation for the old model
for i in range(config['num_input_layers'], len(old_model.layers)):
x = old_model.layers[i](x)
# define straight-through model
model1 = Model(model_input, x)
# compile model
model1.compile(loss=wasserstein_loss,
optimizer=ProGan.get_optimizer(config))
# downsample the new larger image
downsample = AveragePooling2D()(model_input)
# connect old input processing to downsampled new input
old_block = old_model.layers[1](downsample)
old_block = old_model.layers[2](old_block)
# fade in output of old model input layer with new input
x = WeightedSum()([old_block, new_block])
# skip the input, 1x1 and activation for the old model
for i in range(config['num_input_layers'], len(old_model.layers)):
x = old_model.layers[i](x)
# define fade-in model
model2 = Model(model_input, x)
# compile model
model2.compile(loss=wasserstein_loss,
optimizer=ProGan.get_optimizer(config))
return [model1, model2]
def ShallowConvNet(nb_classes, Chans=64, Samples=128, dropoutRate=0.5):
""" Keras implementation of the Shallow Convolutional Network as described
in Schirrmeister et. al. (2017), Human Brain Mapping.
Assumes the input is a 2-second EEG signal sampled at 128Hz. Note that in
the original paper, they do temporal convolutions of length 25 for EEG
data sampled at 250Hz. We instead use length 13 since the sampling rate is
roughly half of the 250Hz which the paper used. The pool_size and stride
in later layers is also approximately half of what is used in the paper.
Note that we use the max_norm constraint on all convolutional layers, as
well as the classification layer. We also change the defaults for the
BatchNormalization layer. We used this based on a personal communication
with the original authors.
ours original paper
pool_size 1, 35 1, 75
strides 1, 7 1, 15
conv filters 1, 13 1, 25
Note that this implementation has not been verified by the original
authors. We do note that this implementation reproduces the results in the
original paper with minor deviations.
"""
# start the model
input_main = Input((1, Chans, Samples))
block1 = Conv2D(40, (1, 13),
input_shape=(1, Chans, Samples),
kernel_constraint=max_norm(2., axis=(0, 1, 2)))(input_main)
block1 = Conv2D(40, (Chans, 1),
use_bias=False,
kernel_constraint=max_norm(2., axis=(0, 1, 2)))(block1)
block1 = BatchNormalization(axis=1, epsilon=1e-05, momentum=0.1)(block1)
block1 = Activation(square)(block1)
block1 = AveragePooling2D(pool_size=(1, 35), strides=(1, 7))(block1)
block1 = Activation(log)(block1)
block1 = Dropout(dropoutRate)(block1)
flatten = Flatten()(block1)
dense = Dense(nb_classes, kernel_constraint=max_norm(0.5))(flatten)
softmax = Activation('softmax')(dense)
return Model(inputs=input_main, outputs=softmax)
def _build_discriminator(config):
model_input = Input(shape=tuple(config['input_shape']),
name="base_dis_input")
# weights init
w_init = RandomNormal(stddev=0.02)
w_const = max_norm(1.0)
# conv layers
x = model_input
for i, strides in enumerate([1, 3, 4]):
# ?? why minibatch layer only after 1x1 conv
if i == 1:
x = MinibatchStdev()(x)
x = Conv2D(config['filters'],
strides,
padding='same',
kernel_initializer=w_init,
kernel_constraint=w_const)(x)
x = LeakyReLU()(x)
# dense output
features = Flatten()(x)
model_output = Dense(1)(features)
# compile model
model = Model(model_input, model_output)
model.compile(loss=wasserstein_loss,
optimizer=Adam(lr=config['learning_rate'],
beta_1=config['beta_1'],
beta_2=config['beta_2'],
epsilon=config['epsilon']))
# store model
model_list = [[model, model]]
# create submodels
for i in range(1, config['num_blocks']):
# get prior model without the fade-on
old_model = model_list[i - 1][0]
# create new model for next resolution
models = ProGan._add_discriminator_block(old_model, config)
# store model
model_list.append(models)
return model_list
def _build_generator(init_shape, config):
latent_vector = Input(config['z_size'], name='generator_input')
x = Dense(np.prod(init_shape))(latent_vector)
x = Reshape(init_shape)(x)
# weights init
w_init = RandomNormal(stddev=0.02)
w_const = max_norm(1.0)
# conv layers
for i, strides in enumerate([4, 3]):
x = Conv2D(config['filters'],
strides,
padding='same',
kernel_initializer=w_init,
kernel_constraint=w_const)(x)
x = PixelNormalization()(x)
x = LeakyReLU()(x)
# conv 1x1
out_image = Conv2D(config['n_channels'], 1, padding='same')(x)
model = Model(latent_vector, out_image)
# store model
model_list = [[model, model]]
# create submodels
for i in range(1, config['num_blocks']):
# get prior model without the fade-on
old_model = model_list[i - 1][0]
# create new model for next resolution
models = ProGan._add_generator_block(old_model, config)
# store model
model_list.append(models)
return model_list
def attention_model_new_arch(src_vocab,
target_vocab,
src_timesteps,
target_timesteps,
units,
epochs=30):
encoder_inputs = Input(shape=(src_timesteps, ), name='encoder_inputs')
decoder_inputs = Input(shape=(target_timesteps - 1, target_vocab),
name='decoder_inputs')
embedding = Embedding(src_vocab,
units,
input_length=src_timesteps,
name='enc_embedding',
mask_zero=True)
embedding2 = Dropout(0.5)(embedding(encoder_inputs))
encoder_lstm = Bidirectional(LSTM(units,
return_sequences=True,
return_state=True,
kernel_constraint=max_norm(3.0),
recurrent_constraint=max_norm(3.0),
name='encoder_lstm'),
name='bidirectional_encoder')
encoder_out, forward_h, forward_c, backward_h, backward_c = encoder_lstm(
embedding2)
state_h = Concatenate()([forward_h, backward_h])
state_c = Concatenate()([forward_c, backward_c])
enc_states = [state_h, state_c]
# Decoder
decoder_lstm = LSTM(units * 2,
return_sequences=True,
return_state=True,
name='decoder_lstm')
decoder_out, _, _ = decoder_lstm(decoder_inputs, initial_state=enc_states)
# Attention
attn_layer = AttentionLayer(name='attention_layer')
attn_out, attn_states = attn_layer([encoder_out, decoder_out])
# concat attention and decoder output
decoder_output_concat = Concatenate(axis=-1)([decoder_out, attn_out])
# FC layer
tst = Dropout(0.5)(decoder_output_concat)
decoder_dense = Dense(target_vocab, activation='softmax')
decoder_pred = decoder_dense(tst)
model = Model(inputs=[encoder_inputs, decoder_inputs],
outputs=decoder_pred)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# Inference models
# Encoder Inference model
encoder_inf_inputs = Input(batch_shape=(
1,
src_timesteps,
),
name='encoder_inf_inputs')
encoder_inf_out, encoder_inf_fwd_h, encoder_inf_fwd_c, encoder_inf_back_h, encoder_inf_back_c = encoder_lstm(
embedding(encoder_inf_inputs))
encoder_inf_h = Concatenate()([encoder_inf_fwd_h, encoder_inf_back_h])
encoder_inf_c = Concatenate()([encoder_inf_fwd_c, encoder_inf_back_c])
encoder_model = Model(
inputs=encoder_inf_inputs,
outputs=[encoder_inf_out, encoder_inf_h, encoder_inf_c])
# Decoder Inference model
encoder_inf_states = Input(batch_shape=(1, src_timesteps, 2 * units),
name='encoder_inf_states')
decoder_inf_inputs = Input(batch_shape=(1, 1, target_vocab),
name='decoder_word_inputs')
decoder_init_fwd_state = Input(batch_shape=(1, units * 2),
name='decoder_fwd_init')
decoder_init_back_state = Input(batch_shape=(1, units * 2),
name='decoder_back_init')
decoder_states_inputs = [decoder_init_fwd_state, decoder_init_back_state]
decoder_inf_out, decoder_inf_fwd_state, decoder_inf_back_state = decoder_lstm(
decoder_inf_inputs, initial_state=decoder_states_inputs)
attn_inf_out, attn_inf_states = attn_layer(
[encoder_inf_states, decoder_inf_out])
decoder_inf_concat = Concatenate(
axis=-1, name='concat')([decoder_inf_out, attn_inf_out])
decoder_inf_pred = decoder_dense(decoder_inf_concat)
decoder_model = Model(inputs=[
encoder_inf_states, decoder_init_fwd_state, decoder_init_back_state,
decoder_inf_inputs
],
outputs=[
decoder_inf_pred, attn_inf_states,
decoder_inf_fwd_state, decoder_inf_back_state
])
return model, encoder_model, decoder_model
def EEGNet(nb_classes, Chans = 64, Samples = 128,
dropoutRate = 0.5, kernLength = 64, F1 = 8,
D = 2, F2 = 16, norm_rate = 0.25, dropoutType = 'Dropout'):
""" Keras Implementation of EEGNet
http://iopscience.iop.org/article/10.1088/1741-2552/aace8c/meta
Note that this implements the newest version of EEGNet and NOT the earlier
version (version v1 and v2 on arxiv). We strongly recommend using this
architecture as it performs much better and has nicer properties than
our earlier version. For example:
1. Depthwise Convolutions to learn spatial filters within a
temporal convolution. The use of the depth_multiplier option maps
exactly to the number of spatial filters learned within a temporal
filter. This matches the setup of algorithms like FBCSP which learn
spatial filters within each filter in a filter-bank. This also limits
the number of free parameters to fit when compared to a fully-connected
convolution.
2. Separable Convolutions to learn how to optimally combine spatial
filters across temporal bands. Separable Convolutions are Depthwise
Convolutions followed by (1x1) Pointwise Convolutions.
While the original paper used Dropout, we found that SpatialDropout2D
sometimes produced slightly better results for classification of ERP
signals. However, SpatialDropout2D significantly reduced performance
on the Oscillatory dataset (SMR, BCI-IV Dataset 2A). We recommend using
the default Dropout in most cases.
Assumes the input signal is sampled at 128Hz. If you want to use this model
for any other sampling rate you will need to modify the lengths of temporal
kernels and average pooling size in blocks 1 and 2 as needed (double the
kernel lengths for double the sampling rate, etc). Note that we haven't
tested the model performance with this rule so this may not work well.
The model with default parameters gives the EEGNet-8,2 model as discussed
in the paper. This model should do pretty well in general, although it is
advised to do some model searching to get optimal performance on your
particular dataset.
We set F2 = F1 * D (number of input filters = number of output filters) for
the SeparableConv2D layer. We haven't extensively tested other values of this
parameter (say, F2 < F1 * D for compressed learning, and F2 > F1 * D for
overcomplete). We believe the main parameters to focus on are F1 and D.
Inputs:
nb_classes : int, number of classes to classify
Chans, Samples : number of channels and time points in the EEG data
dropoutRate : dropout fraction
kernLength : length of temporal convolution in first layer. We found
that setting this to be half the sampling rate worked
well in practice. For the SMR dataset in particular
since the data was high-passed at 4Hz we used a kernel
length of 32.
F1, F2 : number of temporal filters (F1) and number of pointwise
filters (F2) to learn. Default: F1 = 8, F2 = F1 * D.
D : number of spatial filters to learn within each temporal
convolution. Default: D = 2
dropoutType : Either SpatialDropout2D or Dropout, passed as a string.
"""
if dropoutType == 'SpatialDropout2D':
dropoutType = SpatialDropout2D
elif dropoutType == 'Dropout':
dropoutType = Dropout
else:
raise ValueError('dropoutType must be one of SpatialDropout2D '
'or Dropout, passed as a string.')
input1 = Input(shape = (1, Chans, Samples))
##################################################################
block1 = Conv2D(F1, (1, kernLength), padding = 'same',
input_shape = (1, Chans, Samples),
use_bias = False)(input1)
block1 = BatchNormalization(axis = 1)(block1)
block1 = DepthwiseConv2D((Chans, 1), use_bias = False,
depth_multiplier = D,
depthwise_constraint = max_norm(1.))(block1)
block1 = BatchNormalization(axis = 1)(block1)
block1 = Activation('elu')(block1)
block1 = AveragePooling2D((1, 4))(block1)
block1 = dropoutType(dropoutRate)(block1)
block2 = SeparableConv2D(F2, (1, 16),
use_bias = False, padding = 'same')(block1)
block2 = BatchNormalization(axis = 1)(block2)
block2 = Activation('elu')(block2)
block2 = AveragePooling2D((1, 8))(block2)
block2 = dropoutType(dropoutRate)(block2)
flatten = Flatten(name = 'flatten')(block2)
dense = Dense(nb_classes, name = 'dense',
kernel_constraint = max_norm(norm_rate))(flatten)
softmax = Activation('softmax', name = 'softmax')(dense)
return Model(inputs=input1, outputs=softmax)
def DeepConvNet(nb_classes, Chans = 64, Samples = 256,
dropoutRate = 0.5):
""" Keras implementation of the Deep Convolutional Network as described in
Schirrmeister et. al. (2017), Human Brain Mapping.
This implementation assumes the input is a 2-second EEG signal sampled at
128Hz, as opposed to signals sampled at 250Hz as described in the original
paper. We also perform temporal convolutions of length (1, 5) as opposed
to (1, 10) due to this sampling rate difference.
Note that we use the max_norm constraint on all convolutional layers, as
well as the classification layer. We also change the defaults for the
BatchNormalization layer. We used this based on a personal communication
with the original authors.
ours original paper
pool_size 1, 2 1, 3
strides 1, 2 1, 3
conv filters 1, 5 1, 10
Note that this implementation has not been verified by the original
authors.
"""
# start the model
input_main = Input((1, Chans, Samples))
block1 = Conv2D(25, (1, 5),
input_shape=(1, Chans, Samples),
kernel_constraint = max_norm(2., axis=(0,1,2)))(input_main)
block1 = Conv2D(25, (Chans, 1),
kernel_constraint = max_norm(2., axis=(0,1,2)))(block1)
block1 = BatchNormalization(axis=1, epsilon=1e-05, momentum=0.1)(block1)
block1 = Activation('elu')(block1)
block1 = MaxPooling2D(pool_size=(1, 2), strides=(1, 2))(block1)
block1 = Dropout(dropoutRate)(block1)
block2 = Conv2D(50, (1, 5),
kernel_constraint = max_norm(2., axis=(0,1,2)))(block1)
block2 = BatchNormalization(axis=1, epsilon=1e-05, momentum=0.1)(block2)
block2 = Activation('elu')(block2)
block2 = MaxPooling2D(pool_size=(1, 2), strides=(1, 2))(block2)
block2 = Dropout(dropoutRate)(block2)
block3 = Conv2D(100, (1, 5),
kernel_constraint = max_norm(2., axis=(0,1,2)))(block2)
block3 = BatchNormalization(axis=1, epsilon=1e-05, momentum=0.1)(block3)
block3 = Activation('elu')(block3)
block3 = MaxPooling2D(pool_size=(1, 2), strides=(1, 2))(block3)
block3 = Dropout(dropoutRate)(block3)
block4 = Conv2D(200, (1, 5),
kernel_constraint = max_norm(2., axis=(0,1,2)))(block3)
block4 = BatchNormalization(axis=1, epsilon=1e-05, momentum=0.1)(block4)
block4 = Activation('elu')(block4)
block4 = MaxPooling2D(pool_size=(1, 2), strides=(1, 2))(block4)
block4 = Dropout(dropoutRate)(block4)
flatten = Flatten()(block4)
dense = Dense(nb_classes, kernel_constraint = max_norm(0.5))(flatten)
softmax = Activation('softmax')(dense)
return Model(inputs=input_main, outputs=softmax)
