def RCL_block(l,activation_function=PReLU(),features=64,kernel_size=kernelSz,name="RCL"):
#INITIAL CONVOLUTION BLOCK
conv1 = Conv1D(features, kernel_size, border_mode='same',name=name)
stack1 = conv1(l)
stack2 = activation_function(stack1)
stack3 = BatchNormalization()(stack2)
#UNROLLED RECURRENT BLOCK(s)
conv2 = Conv1D(features, kernel_size, border_mode='same', init = 'he_normal')
stack4 = conv2(stack3)
stack5 = merge([stack1, stack4], mode='sum')
stack6 = activation_function(stack5)
stack7 = BatchNormalization()(stack6)
conv3 = Convolution1D_tied(features, kernel_size, border_mode='same', tied_to = conv2)
stack8 = conv3(stack7)
stack9 = merge([stack1, stack8], mode='sum')
stack10 = activation_function(stack9)
stack11 = BatchNormalization()(stack10)
conv4 = Convolution1D_tied(features, kernel_size, border_mode='same', tied_to = conv2)
stack12 = conv4(stack11)
stack13 = merge([stack1, stack12], mode='sum')
stack14 = activation_function(stack13)
stack15 = BatchNormalization()(stack14)
return stack15
def defineNetwork(self):
print("Setting up network...")
conv = [None] * self.n_conv_layers
deconv = [None] * self.n_conv_layers
inputs = Input(shape=(self.nx, self.ny, self.n_diversity))
conv[0] = Convolution2D(self.n_filters, 3, 3, activation='relu', subsample=(self.stride,self.stride), border_mode='same', init='he_normal')(inputs)
for i in range(self.n_conv_layers-1):
conv[i+1] = Convolution2D(self.n_filters, 3, 3, activation='relu', subsample=(self.stride,self.stride), border_mode='same', init='he_normal')(conv[i])
deconv[0] = Deconvolution2D(self.n_filters, 3, 3, activation='relu', output_shape=(self.batch_size, self.nx, self.ny,self.n_filters), subsample=(self.stride,self.stride), border_mode='same', init='he_normal')(conv[-1])
for i in range(self.n_conv_layers-1):
if (i % self.skip_frequency == 0):
x = Deconvolution2D(self.n_filters, 3, 3, output_shape=(self.batch_size,self.nx, self.ny,self.n_filters), activation='relu', subsample=(self.stride,self.stride), border_mode='same', init='he_normal')(deconv[i])
x = merge([conv[self.n_conv_layers-i-2], x], mode='sum')
deconv[i+1] = Activation('relu')(x)
else:
deconv[i+1] = Deconvolution2D(self.n_filters, 3, 3, output_shape=(self.batch_size,self.nx, self.ny,self.n_filters), activation='relu', subsample=(self.stride,self.stride), border_mode='same', init='he_normal')(deconv[i])
x = Deconvolution2D(1, 1, 1, output_shape=(self.batch_size,self.nx, self.ny, 1), activation='linear', subsample=(self.stride,self.stride), border_mode='same', init='he_normal')(deconv[-1])
focused = Lambda(lambda x: x[:,:,:,0:1], output_shape=(self.nx, self.ny, 1))(inputs)
final = merge([x, focused], 'sum')
self.model = Model(input=inputs, output=final)
self.model.load_weights("{0}_weights.hdf5".format(self.root))
def inception_stem(input):
c = Convolution2D(32, 3, 3, activation='relu', subsample=(2, 2))(input)
# c = Convolution2D(32, 3, 3, activation='relu', subsample=(1, 1))(input)
c = Convolution2D(32, 3, 3, activation='relu', )(c)
c = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(c)
c1 = MaxPooling2D((3, 3), strides=(2, 2))(c)
c2 = Convolution2D(64, 3, 3, activation='relu', subsample=(2, 2))(c)
m = merge([c1, c2], mode='concat', concat_axis=1)
c1 = Convolution2D(64, 1, 1, activation='relu', border_mode='same')(m)
c1 = Convolution2D(96, 3, 3, activation='relu', )(c1)
c2 = Convolution2D(64, 1, 1, activation='relu', border_mode='same')(m)
c2 = Convolution2D(64, 7, 1, activation='relu', border_mode='same')(c2)
c2 = Convolution2D(64, 1, 7, activation='relu', border_mode='same')(c2)
c2 = Convolution2D(96, 3, 3, activation='relu', border_mode='valid')(c2)
m2 = merge([c1, c2], mode='concat', concat_axis=1)
p1 = MaxPooling2D((3,3), strides=(2, 2), )(m2)
p2 = Convolution2D(96, 3, 3, activation='relu', subsample=(2,2))(m2)
m3 = merge([p1, p2], mode='concat', concat_axis=1)
m3 = BatchNormalization()(m3)
m3 = Activation('relu')(m3)
return m3
def Inception_A(Input):
Input = Activation('relu')(Input)
path1 = Convolution2D(32 // nb_filters_reduction_factor,
1, 1, border_mode='same', activation='relu')(Input)
path1 = Convolution2D(48 // nb_filters_reduction_factor,
3, 3, border_mode='same', activation='relu')(path1)
path1 = Convolution2D(64 // nb_filters_reduction_factor,
3, 3, border_mode='same', activation='relu')(path1)
path2 = Convolution2D(32 // nb_filters_reduction_factor,
1, 1, border_mode='same', activation='relu')(Input)
path2 = Convolution2D(32 // nb_filters_reduction_factor,
3, 3, border_mode='same', activation='relu')(path2)
path3 = Convolution2D(32 // nb_filters_reduction_factor,
1, 1, border_mode='same', activation='relu')(Input)
out = merge([path3, path1, path2], mode='concat', concat_axis=channel_axis)
out = Convolution2D(384 // nb_filters_reduction_factor,
1, 1, border_mode='same')(out)
out = Lambda(lambda x: x * alpha)(out)
output = merge([out, Input], mode='sum')
output = BatchNormalization(axis=channel_axis)(output)
output = Activation('relu')(output)
return output
def __init__(self, state_dim, action_dim):
''' Create the Learner Networks | Three callable structures = net1, net2, actor
net1 --> holds actor fixed and trains critic to match target values
net2 --> holds critic fixed and trains actor to produce action that maximizes the Q value
best_net --> for just keeping track of best_net so far
actor --> can be used to predict the action given state
'''
# Make placeholders
self.t_state = Input(shape=(state_dim,))
self.t_exploration = Input(shape=(action_dim,))
a1 = Dense(A_UNITS[0], activation=A_ACT[0])(self.t_state)
a2 = Dense(A_UNITS[1], activation=A_ACT[1])(a1)
a3 = Dense(output_dim=action_dim, activation=A_ACT[2])(a2)
m1 = merge([self.t_exploration, a3], mode='sum')
m2 = merge([self.t_state, m1], mode='concat')
c1 = Dense(C_UNITS[0], activation=C_ACT[0])(m2)
c2 = Dense(C_UNITS[1], activation=C_ACT[1])(c1)
c3 = Dense(output_dim=1, activation=C_ACT[2])(c2)
self.actor = Model(input=[self.t_state], output=[a3])
self.net1 = Model(input=[self.t_state, self.t_exploration], output=[c3])
self.net2 = Model(input=[self.t_state, self.t_exploration], output=[c3])
# Compile networks
opt_set1 = RMSprop(lr=LEARN_RATE)
self.net1.compile(optimizer=opt_set1, loss='mse')
opt_set2 = SGD(lr=LEARN_RATE)
self.net2.compile(optimizer=opt_set2, loss='nmo')
# Collect layers of actor and critic
self.alayers = [1,2,3]
self.clayers = [7,8,9]
def __init__(self, state_dim, action_dim, param_dim):
''' Create the target networks | Two callable structures: actor and critic '''
# Make placeholders
self.t_state = Input(shape=(state_dim,))
self.t_exploration = Input(shape=(action_dim,))
self.t_param = Input(shape=(param_dim,))
self.t_target = Input(shape=(1,))
a1 = Dense(10, activation='relu')(self.t_state)
a2 = Dense(10, activation='relu')(a1)
a3 = Dense(output_dim=action_dim, activation='linear')(a2)
m1 = merge([self.t_exploration, a3], mode='sum')
m2 = merge([self.t_state, m1], mode='concat')
c1 = Dense(10, activation='relu')(m2)
c2 = Dense(5, activation='tanh')(c1)
p1 = Dense(10, activation='relu')(self.t_param)
p2 = Dense(5, activation='tanh')(p1)
m3 = merge([c2, p2], mode='concat')
c3 = Dense(10, activation='relu')(m3)
c4 = Dense(5, activation='relu')(c3)
c5 = Dense(output_dim=1, activation='linear')(c4)
self.actor = Model(input=[self.t_state], output=[a3])
self.critic = Model(input=[self.t_state, self.t_exploration, self.t_param], output=[c5])
def I2INet3D(input_shape=(64,64,64,1), Nfilters=32, l2_reg=0.0):
inp = Input(shape=input_shape)
#First convolution layer
x = Conv3D(Nfilters,(3,3,3), padding='same',activation='relu')(inp)
x = Conv3D(Nfilters,(3,3,3), padding='same',activation='relu')(x)
out_1 = x
x = AveragePooling3D()(x)
#second convolution layer
Nfilter2 = 4*Nfilters
x = Conv3D(Nfilter2,(3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter2,(3,3,3), padding='same',activation='relu')(x)
out_2 = x
x = AveragePooling3D()(x)
#third convolution layer
Nfilter3 = 8*Nfilters
x = Conv3D(Nfilter3,(3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter3,(3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter3,(3,3,3), padding='same',activation='relu')(x)
out_3 = x
x = AveragePooling3D()(x)
#fourth convolution layer
Nfilter4 = 16*Nfilters
x = Conv3D(Nfilter4,(3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter4,(3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter4,(3,3,3), padding='same',activation='relu')(x)
out_4 = UpSampling3D()(x)
####################
# Second branch
####################
s = merge([out_3,out_4], mode='concat', concat_axis=4)
x = Conv3D(Nfilter4, (1,1,1), padding='same',activation='relu')(s)
x = Conv3D(Nfilter3, (3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter3, (3,3,3), padding='same',activation='relu')(x)
s_out_1 = Conv3D(1, (1,1,1), padding='same',activation='sigmoid')(x)
#second upsample
x = UpSampling3D()(x)
s = merge([out_2,x], mode='concat', concat_axis=4)
x = Conv3D(Nfilter3, (1,1,1), padding='same',activation='relu')(s)
x = Conv3D(Nfilter2, (3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilter2, (3,3,3), padding='same',activation='relu')(x)
s_out_2 = Conv3D(1, (1,1,1), padding='same',activation='sigmoid')(x)
#final upsample
x = UpSampling3D()(x)
s = merge([out_1,x], mode='concat', concat_axis=4)
x = Conv3D(Nfilter2, (1,1,1), padding='same',activation='relu')(s)
x = Conv3D(Nfilters, (3,3,3), padding='same',activation='relu')(x)
x = Conv3D(Nfilters, (3,3,3), padding='same',activation='relu')(x)
s_out_3 = Conv3D(1, (1,1,1), padding='same',activation='sigmoid')(x)
i2i = Model(inp,[s_out_3,s_out_2,s_out_1])
return i2i
def __init__(self):
'''
Constructor
'''
super().__init__()
# input tensor
lr_image = Input(shape=(None,None,4))
# contraction part
conv1 = Convolution2D(64, 7, 7, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same', subsample=(2, 2))(lr_image)
conv2 = Convolution2D(128, 5, 5, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same', subsample=(2, 2))(conv1)
# extension part
# upconvolution level 2
upconv2 = UpSampling2D(size=(2,2), dim_ordering='tf')(conv1)
upconv2 = Convolution2D(128, 4, 4, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same')(upconv2)
upconv2 = Convolution2D(128, 3, 3, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same')(upconv2)
# upconvolution level 1
upconv1 = merge([conv2, upconv2], mode='concat', concat_axis=-1)
#TODO: Does -1 correspond to last axis?
print('Does -1 correspond to last axis?')
upconv1 = UpSampling2D(size=(2,2), dim_ordering='tf')(upconv1)
upconv1 = Convolution2D(64, 4, 4, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same')(upconv1)
upconv1 = Convolution2D(64, 3, 3, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same')(upconv1)
# upconvolution level 0 (sr)
sr_prediction = merge([conv1, upconv1], mode='concat', concat_axis=-1)
#TODO: Does -1 correspond to last axis?
print('Does -1 correspond to last axis?')
sr_prediction = UpSampling2D(size=(2,2), dim_ordering='tf')(sr_prediction)
sr_prediction = Convolution2D(4, 4, 4, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same')(sr_prediction)
sr_prediction = Convolution2D(4, 3, 3, dim_ordering='tf', init=he_normal, activation=relu, border_mode='same')(sr_prediction)
# configuration
self.model = Model(input=lr_image, output=sr_prediction)
self.model.compile(optimizer='rmsprop', loss='mse')
def compile(self, optimizer, metrics=[]):
metrics += [mean_q] # register default metrics
# Create target V model. We don't need targets for mu or L.
self.target_V_model = clone_model(self.V_model, self.custom_model_objects)
self.target_V_model.compile(optimizer='sgd', loss='mse')
# Build combined model.
observation_shape = self.V_model.input._keras_shape[1:]
a_in = Input(shape=(self.nb_actions,), name='action_input')
o_in = Input(shape=observation_shape, name='observation_input')
L_out = self.L_model([a_in, o_in])
V_out = self.V_model(o_in)
mu_out = self.mu_model(o_in)
A_out = NAFLayer(self.nb_actions)(merge([L_out, mu_out, a_in], mode='concat'))
combined_out = merge([A_out, V_out], mode='sum')
combined = Model(input=[a_in, o_in], output=combined_out)
# Compile combined model.
if self.target_model_update < 1.:
# We use the `AdditionalUpdatesOptimizer` to efficiently soft-update the target model.
updates = get_soft_target_model_updates(self.target_V_model, self.V_model, self.target_model_update)
optimizer = AdditionalUpdatesOptimizer(optimizer, updates)
def clipped_mse(y_true, y_pred):
delta = K.clip(y_true - y_pred, self.delta_range[0], self.delta_range[1])
return K.mean(K.square(delta), axis=-1)
combined.compile(loss=clipped_mse, optimizer=optimizer, metrics=metrics)
self.combined_model = combined
self.compiled = True
def residual_block(l, increase_dim=False, first=False, filters=16):
if increase_dim:
first_stride = (2,2)
else:
first_stride = (1,1)
if first:
pre_act = l
else:
# BN -> ReLU
bn = BatchNormalization(axis=1)(l)
pre_act = Activation('relu')(bn)
conv_1 = Convolution2D(filters, 3,3, init='he_normal', border_mode='same', subsample=first_stride, activation='linear')(pre_act)
bn_1 = BatchNormalization(axis=1)(conv_1)
relu_1 = Activation('relu')(bn_1)
conv_2 = Convolution2D(filters, 3,3, init='he_normal', border_mode='same', activation='linear')(relu_1)
# add shorcut
if increase_dim:
# projection shortcut
projection = Convolution2D(filters, 1,1, subsample=(2,2), border_mode='same', activation='linear')(pre_act)
block = merge([conv_2, projection], mode='sum')
else:
block = merge([conv_2, pre_act], mode='sum')
return block
def test_merge_mask_2d():
from keras.layers import Input, merge, Masking
from keras.models import Model
rand = lambda *shape: np.asarray(np.random.random(shape) > 0.5, dtype='int32')
# inputs
input_a = Input(shape=(3,))
input_b = Input(shape=(3,))
# masks
masked_a = Masking(mask_value=0)(input_a)
masked_b = Masking(mask_value=0)(input_b)
# three different types of merging
merged_sum = merge([masked_a, masked_b], mode='sum')
merged_concat = merge([masked_a, masked_b], mode='concat', concat_axis=1)
merged_concat_mixed = merge([masked_a, input_b], mode='concat', concat_axis=1)
# test sum
model_sum = Model([input_a, input_b], [merged_sum])
model_sum.compile(loss='mse', optimizer='sgd')
model_sum.fit([rand(2, 3), rand(2, 3)], [rand(2, 3)], nb_epoch=1)
# test concatenation
model_concat = Model([input_a, input_b], [merged_concat])
model_concat.compile(loss='mse', optimizer='sgd')
model_concat.fit([rand(2, 3), rand(2, 3)], [rand(2, 6)], nb_epoch=1)
# test concatenation with masked and non-masked inputs
model_concat = Model([input_a, input_b], [merged_concat_mixed])
model_concat.compile(loss='mse', optimizer='sgd')
model_concat.fit([rand(2, 3), rand(2, 3)], [rand(2, 6)], nb_epoch=1)
def Stem(Input):
print(Input)
x = Convolution2D(32 // nb_filters_reduction_factor,
3, 3, subsample=(1, 1))(Input)
x = Convolution2D(32 // nb_filters_reduction_factor, 3, 3)(x)
x = Convolution2D(32 // nb_filters_reduction_factor, 3, 3)(x)
x = Convolution2D(64 // nb_filters_reduction_factor,
3, 3, border_mode='same')(x)
path1 = MaxPooling2D((3, 3), strides=(1, 1))(x) # changed
path2 = Convolution2D(96 // nb_filters_reduction_factor,
3, 3, subsample=(1, 1))(x) # changed
y = merge([path1, path2], mode='concat')
a = Convolution2D(64 // nb_filters_reduction_factor,
1, 1, border_mode='same')(y)
a = Convolution2D(64 // nb_filters_reduction_factor,
3, 1, border_mode='same')(a)
a = Convolution2D(64 // nb_filters_reduction_factor,
1, 3, border_mode='same')(a)
a = Convolution2D(96 // nb_filters_reduction_factor,
3, 3, border_mode='valid')(a)
b = Convolution2D(64 // nb_filters_reduction_factor,
1, 1, border_mode='same')(y)
b = Convolution2D(96 // nb_filters_reduction_factor,
3, 3, border_mode='valid')(b)
z = merge([a, b], mode='concat')
z1 = MaxPooling2D((3, 3), strides=(2, 2))(z)
z2 = Convolution2D(192 // nb_filters_reduction_factor, 3,
3, subsample=(2, 2), border_mode='valid')(z)
c = merge([z1, z2], mode='concat', concat_axis=-1)
return c
def block_inception_c(input):
if K.image_dim_ordering() == "th":
channel_axis = 1
else:
channel_axis = -1
branch_0 = conv2d_bn(input, 256, 1, 1)
branch_1 = conv2d_bn(input, 384, 1, 1)
branch_10 = conv2d_bn(branch_1, 256, 1, 3)
branch_11 = conv2d_bn(branch_1, 256, 3, 1)
branch_1 = merge([branch_10, branch_11], mode='concat', concat_axis=channel_axis)
branch_2 = conv2d_bn(input, 384, 1, 1)
branch_2 = conv2d_bn(branch_2, 448, 3, 1)
branch_2 = conv2d_bn(branch_2, 512, 1, 3)
branch_20 = conv2d_bn(branch_2, 256, 1, 3)
branch_21 = conv2d_bn(branch_2, 256, 3, 1)
branch_2 = merge([branch_20, branch_21], mode='concat', concat_axis=channel_axis)
branch_3 = AveragePooling2D((3, 3), strides=(1, 1), border_mode='same')(input)
branch_3 = conv2d_bn(branch_3, 256, 1, 1)
x = merge([branch_0, branch_1, branch_2, branch_3], mode='concat', concat_axis=channel_axis)
return x
def get_similarities(self):
if self._models is None:
self._models = self.build()
assert len(self._models) == 2, 'build() should make question and answer language models'
if self._similarities is None:
question_model, answer_model = self._models
answers_use_question = len(answer_model.internal_input_shapes) == 2
question = question_model(self.question, self.answer_good)
if answers_use_question:
good = answer_model([self.question, self.answer_good])
bad = answer_model([self.question, self.answer_bad])
else:
good = answer_model([self.answer_good])
bad = answer_model([self.answer_bad])
similarity = self.get_similarity()
good_sim = merge([question, good], mode=similarity, output_shape=lambda x: x[:-1])
bad_sim = merge([question, bad], mode=similarity, output_shape=lambda x: x[:-1])
self._similarities = [good_sim, bad_sim]
return self._similarities
def inception_stem(input): # Input (299,299,3)
# Input Shape is 299 x 299 x 3 (th) or 3 x 299 x 299 (th)
c = Convolution2D(32, 3, 3, activation='relu', subsample=(2,2))(input)
c = Convolution2D(32, 3, 3, activation='relu', )(c)
c = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(c)
c1 = MaxPooling2D((3,3), strides=(2,2))(c)
c2 = Convolution2D(96, 3, 3, activation='relu', subsample=(2,2))(c)
m = merge([c1, c2], mode='concat', concat_axis=1)
c1 = Convolution2D(64, 1, 1, activation='relu', border_mode='same')(m)
c1 = Convolution2D(96, 3, 3, activation='relu', )(c1)
c2 = Convolution2D(64, 1, 1, activation='relu', border_mode='same')(m)
c2 = Convolution2D(64, 7, 1, activation='relu', border_mode='same')(c2)
c2 = Convolution2D(64, 1, 7, activation='relu', border_mode='same')(c2)
c2 = Convolution2D(96, 3, 3, activation='relu', border_mode='valid')(c2)
m2 = merge([c1, c2], mode='concat', concat_axis=1)
p1 = MaxPooling2D((3,3), strides=(2,2), )(m2)
p2 = Convolution2D(192, 3, 3, activation='relu', subsample=(2,2))(m2)
m3 = merge([p1, p2], mode='concat', concat_axis=1)
return m3
def get_unet():
inputs = Input((1, img_rows, img_cols))
conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(inputs)
conv1 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(pool1)
conv2 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(pool2)
conv3 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(pool3)
conv4 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Convolution2D(512, 3, 3, activation='relu', border_mode='same')(pool4)
conv5 = Convolution2D(512, 3, 3, activation='relu', border_mode='same')(conv5)
pool5 = MaxPooling2D(pool_size=(2, 2))(conv5)
poll5_flat = Flatten()(pool5)
dense1 = Dense(1024, activation='relu')(poll5_flat)
dense2 = Dense(512, activation='relu')(dense1)
dense3 = Dense(1, activation='sigmoid', name='has_mask_output')(dense1)
up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4], mode='concat', concat_axis=1)
conv6 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(up6)
conv6 = Convolution2D(256, 3, 3, activation='relu', border_mode='same')(conv6)
up7 = merge([UpSampling2D(size=(2, 2))(conv6), conv3], mode='concat', concat_axis=1)
conv7 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(up7)
conv7 = Convolution2D(128, 3, 3, activation='relu', border_mode='same')(conv7)
up8 = merge([UpSampling2D(size=(2, 2))(conv7), conv2], mode='concat', concat_axis=1)
conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up8)
conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(conv8)
up9 = merge([UpSampling2D(size=(2, 2))(conv8), conv1], mode='concat', concat_axis=1)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up9)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(conv9)
conv10 = Convolution2D(1, 1, 1, activation='sigmoid')(conv9)
conv10_flat = Flatten()(conv10)
dense3_repeat = Flatten()(RepeatVector(img_rows * img_cols)(dense3))
merge_output = merge([conv10_flat, dense3_repeat], mode='mul')
out = Reshape((1, img_rows, img_cols), name='main_output')(merge_output)
model = Model(input=inputs, output=[out, dense3])
model.compile(optimizer=Adam(lr=1e-5),
loss={'main_output': dice_coef_loss, 'has_mask_output': 'binary_crossentropy'},
metrics={'main_output': dice_coef, 'has_mask_output': 'accuracy'})
return model
def build(self):
assert self.config['question_len'] == self.config['answer_len']
question = self.question
answer = self.get_answer()
# add embedding layers
embedding = Embedding(self.config['n_words'], self.model_params.get('n_embed_dims', 100))
question_embedding = embedding(question)
answer_embedding = embedding(answer)
# turn off layer updating
embedding.params = []
embedding.updates = []
# dropout
dropout = Dropout(0.25)
question_dropout = dropout(question_embedding)
answer_dropout = dropout(answer_embedding)
# dense
dense = TimeDistributed(Dense(self.model_params.get('n_hidden', 200), activation='tanh'))
question_dense = dense(question_dropout)
answer_dense = dense(answer_dropout)
# regularization
question_dense = ActivityRegularization(l2=0.0001)(question_dense)
answer_dense = ActivityRegularization(l2=0.0001)(answer_dense)
# dropout
question_dropout = dropout(question_dense)
answer_dropout = dropout(answer_dense)
# cnn
cnns = [Convolution1D(filter_length=filter_length,
nb_filter=self.model_params.get('nb_filters', 1000),
activation=self.model_params.get('conv_activation', 'relu'),
border_mode='same') for filter_length in [2, 3, 5, 7]]
question_cnn = merge([cnn(question_dropout) for cnn in cnns], mode='concat')
answer_cnn = merge([cnn(answer_dropout) for cnn in cnns], mode='concat')
# regularization
question_cnn = ActivityRegularization(l2=0.0001)(question_cnn)
answer_cnn = ActivityRegularization(l2=0.0001)(answer_cnn)
# dropout
question_dropout = dropout(question_cnn)
answer_dropout = dropout(answer_cnn)
# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False), output_shape=lambda x: (x[0], x[2]))
question_pool = maxpool(question_dropout)
answer_pool = maxpool(answer_dropout)
# activation
activation = Activation('tanh')
question_output = activation(question_pool)
answer_output = activation(answer_pool)
return question_output, answer_output
def build_model(self):
lstm_branch = []
input_branch = []
for i in range(self.layers_number):
main_input = Input(shape=(self.lstm_timesteps, self.input_dim), name="main_input_" + str(i))
input_branch.append(main_input)
lstm_out = LSTM(self.lstm_hidden, return_sequences=True)(main_input)
# auxiliary_input = Input(batch_shape=(self.batch_size,1,self.lstm_timesteps), name='auxiliary_input'+str(i))
auxiliary_input = Input(shape=(1, self.lstm_timesteps), name="auxiliary_input" + str(i))
input_branch.append(auxiliary_input)
"""
x1 = Merge([lstm_out, auxiliary_input], mode=lambda x, y: (x*y).sum(axis=0),
name='merge_lstm_auxi'+str(i))
"""
x1 = merge([auxiliary_input, lstm_out], mode="dot", dot_axes=[2, 1], name="merge_lstm_auxi" + str(i))
assert x1
flatten = Reshape((self.lstm_hidden,))(x1)
c_input = Input(shape=(6,), name="c_input" + str(i))
input_branch.append(c_input)
x2 = merge([flatten, c_input], mode="concat")
x2 = Dense(
self.lstm_hidden, activation="relu", W_regularizer=l2(0.001), activity_regularizer=activity_l2(0.001)
)(x2)
assert x2
lstm_branch.append(x2)
lstm_all_out = merge(lstm_branch, mode="sum", name="lstm_all_out")
"""
dense_relu = Dense(self.lstm_hidden, activation='relu', W_regularizer=l2(0.001),
activity_regularizer=activity_l2(0.001))(lstm_all_out)
"""
final_loss = Dense(self.output_dim, name="main_output")(lstm_all_out)
self.model = Model(input_branch, output=final_loss)
self.model.compile(loss="mean_squared_error", optimizer="adagrad")
plot(self.model, to_file="multiple_model.png", show_shapes=True)
def qmodel(self, trainable):
inp_s = Input( shape=(xp.STATE_DIM,), name="inp_s") # batch (None, ...) added automatically
inp_a = Input( shape=(xp.ACTION_DIM,), name="inp_a")
a1 = Dense(640, activation='relu', W_regularizer=l2(0.01))
a2 = Dense(640, activation='relu', W_regularizer=l2(0.01))
#a3 = Dense(320, activation='relu', W_regularizer=l2(0.01), activity_regularizer=activity_l2(0.01))
a_out = Dense(1, W_regularizer=l2(0.01)) #W_constraint=nonneg())
v1 = Dense(640, activation='relu', W_regularizer=l2(0.01))
v2 = Dense(640, activation='relu', W_regularizer=l2(0.01))
v_out = Dense(1, W_regularizer=l2(0.01))
gaussian = Lambda(gaussian_of_x, output_shape=gaussian_of_x_shape)
parabolized_action = merge( [
a_out( a2(a1( merge([inp_s, inp_a], mode='concat') )) ),
gaussian(inp_a)
], mode='mul')
out_tensor = a_out( a2(a1( merge([inp_s, inp_a], mode='concat') )) )
#out_tensor = merge( [
# parabolized_action,
# v_out( v2(v1(inp_s)) )
# ], mode='sum' )
Qmod = Model( input=[inp_s,inp_a], output=out_tensor )
Qmod.compile(loss='mse', optimizer=Adam(lr=0.0005, beta_2=0.9999))
if not trainable:
for layer in [v1,v2,v_out, a1,a2,a_out]:
layer.trainable = False # model already compiled (that's where this flag used), this assignment avoids learning Q layers by learning policy
return Qmod
def build(self):
assert self.config['question_len'] == self.config['answer_len']
question = self.question
answer = self.get_answer()
# add embedding layers
weights = np.load(self.config['initial_embed_weights'])
embedding = Embedding(input_dim=self.config['n_words'],
output_dim=weights.shape[1],
weights=[weights])
question_embedding = embedding(question)
answer_embedding = embedding(answer)
# cnn
cnns = [Convolution1D(filter_length=filter_length,
nb_filter=500,
activation='tanh',
border_mode='same') for filter_length in [2, 3, 5, 7]]
question_cnn = merge([cnn(question_embedding) for cnn in cnns], mode='concat')
answer_cnn = merge([cnn(answer_embedding) for cnn in cnns], mode='concat')
# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False), output_shape=lambda x: (x[0], x[2]))
maxpool.supports_masking = True
enc = Dense(100, activation='tanh')
question_pool = enc(maxpool(question_cnn))
answer_pool = enc(maxpool(answer_cnn))
return question_pool, answer_pool
def build(self):
question = self.question
answer = self.get_answer()
# add embedding layers
weights = np.load(self.config['initial_embed_weights'])
embedding = Embedding(input_dim=self.config['n_words'],
output_dim=weights.shape[1],
# mask_zero=True,
weights=[weights])
question_embedding = embedding(question)
answer_embedding = embedding(answer)
# question rnn part
f_rnn = LSTM(141, return_sequences=True, consume_less='mem')
b_rnn = LSTM(141, return_sequences=True, consume_less='mem', go_backwards=True)
question_f_rnn = f_rnn(question_embedding)
question_b_rnn = b_rnn(question_embedding)
# maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False), output_shape=lambda x: (x[0], x[2]))
maxpool.supports_masking = True
question_pool = merge([maxpool(question_f_rnn), maxpool(question_b_rnn)], mode='concat', concat_axis=-1)
# answer rnn part
from attention_lstm import AttentionLSTMWrapper
f_rnn = AttentionLSTMWrapper(f_rnn, question_pool, single_attention_param=True)
b_rnn = AttentionLSTMWrapper(b_rnn, question_pool, single_attention_param=True)
answer_f_rnn = f_rnn(answer_embedding)
answer_b_rnn = b_rnn(answer_embedding)
answer_pool = merge([maxpool(answer_f_rnn), maxpool(answer_b_rnn)], mode='concat', concat_axis=-1)
return question_pool, answer_pool
def VGG19_hieratt(query_in_size, query_embed_size, nb_classes):
"""Stack hierarchical attention on pre-trained VGG19.
Requires https://github.com/fchollet/deep-learning-models"""
base_model = VGG19(weights='imagenet')
input_image = base_model.input
input_question = Input(shape=(query_in_size,)) # question vector
# Model up to 3rd block
f_1 = Model(input=img_in, output=base_model.get_layer('block3_pool').output)
f_1 = f_1(img_in)
f_1 = Reshape((256, 28*28))(f_1)
f_1 = Permute((2,1))(f_1)
q_1 = Dense(query_embed_size, activation='relu')(input_question) # Encode question
# Add question embedding to each feature column
q_1 = RepeatVector(28*28)(q_1)
q_f = merge([f_1, q_1], 'concat')
# Estimate and apply attention per feature
att_1 = TimeDistributedDense(1, activation="sigmoid")(q_f)
att_1 = Lambda(repeat_1, output_shape=(28*28, 256))(att_1)
att_1 = merge([f_1, att_1], 'mul')
# Reshape to the original feature map from previous layer
att_1 = Permute((2,1))(att_1)
f_1_att = Reshape((256, 28, 28))(att_1)
model = Model(input=[img_in, input_question], output=f_1_att)
print model.summary()
def create_model(self, height=32, width=32, channels=3, load_weights=False, batch_size=128,
small_train_images=False):
"""
Creates a model to remove / reduce noise from upscaled images.
"""
from keras.layers.convolutional import Deconvolution2D
assert height % 4 == 0, "Height of the image must be divisible by 4"
assert width % 4 == 0, "Width of the image must be divisible by 4"
if K.image_dim_ordering() == "th":
shape = (channels, width, height)
else:
shape = (width, height, channels)
init = Input(shape=shape)
level1_1 = Convolution2D(self.n1, 3, 3, activation='relu', border_mode='same')(init)
level2_1 = Convolution2D(self.n1, 3, 3, activation='relu', border_mode='same')(level1_1)
level2_2 = Deconvolution2D(self.n1, 3, 3, activation='relu', output_shape=(None, channels, height, width), border_mode='same')(level2_1)
level2 = merge([level2_1, level2_2], mode='sum')
level1_2 = Deconvolution2D(self.n1, 3, 3, activation='relu', output_shape=(None, channels, height, width), border_mode='same')(level2)
level1 = merge([level1_1, level1_2], mode='sum')
decoded = Convolution2D(channels, 5, 5, activation='linear', border_mode='same')(level1)
model = Model(init, decoded)
model.compile(optimizer='adam', loss='mse', metrics=[PSNRLoss])
if load_weights: model.load_weights("weights/Denoising AutoEncoder.h5")
self.model = model
return model
def inception_stem(input):
"""Create inception stem."""
if K.image_dim_ordering() == "th":
channel_axis = 1
else:
channel_axis = -1
# Input Shape is 299 x 299 x 3 (th) or 3 x 299 x 299 (th)
x = conv_block(input, 32, 3, 3, subsample=(2, 2), border_mode='valid')
x = conv_block(x, 32, 3, 3, border_mode='valid')
x = conv_block(x, 64, 3, 3)
x1 = MaxPooling2D((3, 3), strides=(2, 2), border_mode='valid')(x)
x2 = conv_block(x, 96, 3, 3, subsample=(2, 2), border_mode='valid')
x = merge([x1, x2], mode='concat', concat_axis=channel_axis)
x1 = conv_block(x, 64, 1, 1)
x1 = conv_block(x1, 96, 3, 3, border_mode='valid')
x2 = conv_block(x, 64, 1, 1)
x2 = conv_block(x2, 64, 1, 7)
x2 = conv_block(x2, 64, 7, 1)
x2 = conv_block(x2, 96, 3, 3, border_mode='valid')
x = merge([x1, x2], mode='concat', concat_axis=channel_axis)
x1 = conv_block(x, 192, 3, 3, subsample=(2, 2), border_mode='valid')
x2 = MaxPooling2D((3, 3), strides=(2, 2), border_mode='valid')(x)
x = merge([x1, x2], mode='concat', concat_axis=channel_axis)
return x
def inception_resnet_v2_A(input, scale_residual=True):
if K.image_dim_ordering() == "th":
channel_axis = 1
else:
channel_axis = -1
# Input is relu activation
init = input
ir1 = Convolution2D(32, 1, 1, activation='relu', border_mode='same')(input)
ir2 = Convolution2D(32, 1, 1, activation='relu', border_mode='same')(input)
ir2 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(ir2)
ir3 = Convolution2D(32, 1, 1, activation='relu', border_mode='same')(input)
ir3 = Convolution2D(48, 3, 3, activation='relu', border_mode='same')(ir3)
ir3 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(ir3)
ir_merge = merge([ir1, ir2, ir3], concat_axis=channel_axis, mode='concat')
ir_conv = Convolution2D(384, 1, 1, activation='linear', border_mode='same')(ir_merge)
if scale_residual: ir_conv = Lambda(lambda x: x * 0.1)(ir_conv)
out = merge([init, ir_conv], mode='sum')
out = BatchNormalization(axis=channel_axis)(out)
out = Activation("relu")(out)
return out
def inception_C(input):
if K.image_dim_ordering() == "th":
channel_axis = 1
else:
channel_axis = -1
c1 = conv_block(input, 256, 1, 1)
c2 = conv_block(input, 384, 1, 1)
c2_1 = conv_block(c2, 256, 1, 3)
c2_2 = conv_block(c2, 256, 3, 1)
c2 = merge([c2_1, c2_2], mode='concat', concat_axis=channel_axis)
c3 = conv_block(input, 384, 1, 1)
c3 = conv_block(c3, 448, 3, 1)
c3 = conv_block(c3, 512, 1, 3)
c3_1 = conv_block(c3, 256, 1, 3)
c3_2 = conv_block(c3, 256, 3, 1)
c3 = merge([c3_1, c3_2], mode='concat', concat_axis=channel_axis)
c4 = AveragePooling2D((3, 3), strides=(1, 1), border_mode='same')(input)
c4 = conv_block(c4, 256, 1, 1)
m = merge([c1, c2, c3, c4], mode='concat', concat_axis=channel_axis)
return m
def inception_resnet_stem(input):
if K.image_dim_ordering() == "th":
channel_axis = 1
else:
channel_axis = -1
# Input Shape is 299 x 299 x 3 (th) or 3 x 299 x 299 (th)
c = Convolution2D(32, 3, 3, activation='relu', subsample=(2, 2))(input)
c = Convolution2D(32, 3, 3, activation='relu', )(c)
c = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(c)
c1 = MaxPooling2D((3, 3), strides=(2, 2))(c)
c2 = Convolution2D(96, 3, 3, activation='relu', subsample=(2, 2))(c)
m = merge([c1, c2], mode='concat', concat_axis=channel_axis)
c1 = Convolution2D(64, 1, 1, activation='relu', border_mode='same')(m)
c1 = Convolution2D(96, 3, 3, activation='relu', )(c1)
c2 = Convolution2D(64, 1, 1, activation='relu', border_mode='same')(m)
c2 = Convolution2D(64, 7, 1, activation='relu', border_mode='same')(c2)
c2 = Convolution2D(64, 1, 7, activation='relu', border_mode='same')(c2)
c2 = Convolution2D(96, 3, 3, activation='relu', border_mode='valid')(c2)
m2 = merge([c1, c2], mode='concat', concat_axis=channel_axis)
p1 = MaxPooling2D((3, 3), strides=(2, 2), )(m2)
p2 = Convolution2D(192, 3, 3, activation='relu', subsample=(2, 2))(m2)
m3 = merge([p1, p2], mode='concat', concat_axis=channel_axis)
m3 = BatchNormalization(axis=channel_axis)(m3)
m3 = Activation('relu')(m3)
return m3
def build_model(nb_classes, word_vocab_size, chars_vocab_size,
word_count, word_length, batch_size):
print('Build model...')
CONSUME_LESS='gpu'
char_input = Input(batch_shape=(batch_size,word_count, word_length,),
dtype='int32', name='char_input')
character_embedding = TimeDistributed(Embedding(chars_vocab_size, 15,
input_length=word_length,
name='char_embedding'),
name='td_char_embedding')(char_input)
forward_gru = TimeDistributed(GRU(16,name='char_gru_forward',
consume_less=CONSUME_LESS),
name='td_char_gru_forward')(character_embedding)
backward_gru = TimeDistributed(GRU(16,name='char_gru_backward',
consume_less=CONSUME_LESS,
go_backwards=True),
name='td_char_gru_backward')(character_embedding)
char_embedding = merge([forward_gru,backward_gru],mode='concat')
word_input = Input(batch_shape=(batch_size,word_count,),
dtype='int32', name='word_input')
word_embedding = Embedding(word_vocab_size, 32,
input_length=word_count,
name='word_embedding')(word_input)
embedding = merge([char_embedding,word_embedding],mode='concat')
word_gru = GRU(32, name='word_lstm', consume_less=CONSUME_LESS)(embedding)
dense = Dense(nb_classes, activation='sigmoid', name='dense')(word_gru)
output = Activation('softmax', name='output')(dense)
model = Model(input=[char_input,word_input], output=output)
model.compile(loss='categorical_crossentropy', optimizer='adam',
metrics=['accuracy'])
return model
def fit(inp_layer, inp_embed, X, y, *params): #X_val,y_val
print('fitting...')
#inp_layer, inp_embed = feature_generate(X, cate_columns, cont_columns)
input = merge(inp_embed, mode = 'concat')
print('\tinput shape: ', input.shape)
# deep layer
for i in range(4):
if i == 0:
deep = Dense(272, activation='relu')(Flatten()(input))
else:
deep = Dense(272, activation='relu')(deep)
# cross layer
cross = CrossLayer(output_dim = input.shape[2].value, num_layer = 8, name = "cross_layer")(input)
#concat both layers
output = merge([deep, cross], mode = 'concat')
output = Dense(1, activation = 'sigmoid')(output)
model = Model(inp_layer, output)
print(model.summary())
# plot_model(model, to_file = '/Users/ewenwang/Documents/practice_data/conversion_rate/model.png', show_shapes = True)
model.compile(optimizer = 'Adam', loss = 'binary_crossentropy', metrics = ["accuracy"])
if len(params) == 2:
X_val = params[0]
y_val = params[1]
model.fit([X[c] for c in X.columns], y, batch_size = 1024, epochs = 5, validation_data = ([X_val[c] for c in X_val.columns], y_val))
else:
model.fit([X[c] for c in X.columns], y, batch_size = 1024, epochs = 1)
return model
def test_merge():
from keras.layers import Input, merge
from keras.models import Model
# test modes: 'sum', 'mul', 'concat', 'ave', 'cos', 'dot'.
input_shapes = [(3, 2), (3, 2)]
inputs = [np.random.random(shape) for shape in input_shapes]
# test functional API
for mode in ['sum', 'mul', 'concat', 'ave']:
print(mode)
input_a = Input(shape=input_shapes[0][1:])
input_b = Input(shape=input_shapes[1][1:])
merged = merge([input_a, input_b], mode=mode)
model = Model([input_a, input_b], merged)
model.compile('rmsprop', 'mse')
expected_output_shape = model.get_output_shape_for(input_shapes)
actual_output_shape = model.predict(inputs).shape
assert expected_output_shape == actual_output_shape
config = model.get_config()
model = Model.from_config(config)
model.compile('rmsprop', 'mse')
# test lambda with output_shape lambda
input_a = Input(shape=input_shapes[0][1:])
input_b = Input(shape=input_shapes[1][1:])
merged = merge([input_a, input_b],
mode=lambda tup: K.concatenate([tup[0], tup[1]]),
output_shape=lambda tup: (tup[0][:-1],) + (tup[0][-1] + tup[1][-1],))
expected_output_shape = model.get_output_shape_for(input_shapes)
actual_output_shape = model.predict(inputs).shape
assert expected_output_shape == actual_output_shape
config = model.get_config()
model = Model.from_config(config)
model.compile('rmsprop', 'mse')
# test function with output_shape function
def fn_mode(tup):
x, y = tup
return K.concatenate([x, y])
def fn_output_shape(tup):
s1, s2 = tup
return (s1[:-1],) + (s1[-1] + s2[-1],)
input_a = Input(shape=input_shapes[0][1:])
input_b = Input(shape=input_shapes[1][1:])
merged = merge([input_a, input_b],
mode=fn_mode,
output_shape=fn_output_shape)
expected_output_shape = model.get_output_shape_for(input_shapes)
actual_output_shape = model.predict(inputs).shape
assert expected_output_shape == actual_output_shape
config = model.get_config()
model = Model.from_config(config)
model.compile('rmsprop', 'mse')
def unet_model():
inputs = Input(shape=(1, max_slices, img_size, img_size))
conv1 = Convolution3D(width,
3,
3,
3,
activation='relu',
border_mode='same')(inputs)
conv1 = BatchNormalization(axis=1)(conv1)
conv1 = Convolution3D(width * 2,
3,
3,
3,
activation='relu',
border_mode='same')(conv1)
conv1 = BatchNormalization(axis=1)(conv1)
pool1 = MaxPooling3D(pool_size=(2, 2, 2),
strides=(2, 2, 2),
border_mode='same')(conv1)
conv2 = Convolution3D(width * 2,
3,
3,
3,
activation='relu',
border_mode='same')(pool1)
conv2 = BatchNormalization(axis=1)(conv2)
conv2 = Convolution3D(width * 4,
3,
3,
3,
activation='relu',
border_mode='same')(conv2)
conv2 = BatchNormalization(axis=1)(conv2)
pool2 = MaxPooling3D(pool_size=(2, 2, 2),
strides=(2, 2, 2),
border_mode='same')(conv2)
conv3 = Convolution3D(width * 4,
3,
3,
3,
activation='relu',
border_mode='same')(pool2)
conv3 = BatchNormalization(axis=1)(conv3)
conv3 = Convolution3D(width * 8,
3,
3,
3,
activation='relu',
border_mode='same')(conv3)
conv3 = BatchNormalization(axis=1)(conv3)
pool3 = MaxPooling3D(pool_size=(2, 2, 2),
strides=(2, 2, 2),
border_mode='same')(conv3)
conv4 = Convolution3D(width * 8,
3,
3,
3,
activation='relu',
border_mode='same')(pool3)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Convolution3D(width * 8,
3,
3,
3,
activation='relu',
border_mode='same')(conv4)
conv4 = BatchNormalization(axis=1)(conv4)
conv4 = Convolution3D(width * 16,
3,
3,
3,
activation='relu',
border_mode='same')(conv4)
conv4 = BatchNormalization(axis=1)(conv4)
up5 = merge([UpSampling3D(size=(2, 2, 2))(conv4), conv3],
mode='concat',
concat_axis=1)
conv5 = SpatialDropout3D(0.2)(up5)
conv5 = Convolution3D(width * 8,
3,
3,
3,
activation='relu',
border_mode='same')(conv5)
#conv5 = BatchNormalization()(conv5)
conv5 = Convolution3D(width * 8,
3,
3,
3,
activation='relu',
border_mode='same')(conv5)
#conv5 = BatchNormalization()(conv5)
up6 = merge([UpSampling3D(size=(2, 2, 2))(conv5), conv2],
mode='concat',
concat_axis=1)
conv6 = SpatialDropout3D(0.2)(up6)
conv6 = Convolution3D(width * 4,
3,
3,
3,
activation='relu',
border_mode='same')(conv6)
#conv6 = BatchNormalization()(conv6)
conv6 = Convolution3D(width * 4,
3,
3,
3,
activation='relu',
border_mode='same')(conv6)
#conv6 = BatchNormalization()(conv6)
up7 = merge([UpSampling3D(size=(2, 2, 2))(conv6), conv1],
mode='concat',
concat_axis=1)
conv7 = SpatialDropout3D(0.2)(up7)
conv7 = Convolution3D(width * 2,
3,
3,
3,
activation='relu',
border_mode='same')(conv7)
#conv7 = BatchNormalization()(conv7)
conv7 = Convolution3D(width * 2,
3,
3,
3,
activation='relu',
border_mode='same')(conv7)
#conv7 = BatchNormalization()(conv7)
conv8 = Convolution3D(1, 1, 1, 1, activation='sigmoid')(conv7)
model = Model(input=inputs, output=conv8)
model.compile(optimizer=Adam(lr=1e-5),
loss=dice_coef_loss,
metrics=[dice_coef])
return model
def build_model(input_shape):
xin = Input(input_shape)
#shift the below down by one
x1 = conv_block(xin, 8, activation='relu')
x1_ident = AveragePooling3D()(xin)
x1_merged = merge([x1, x1_ident], mode='concat', concat_axis=1)
x2_1 = conv_block(x1_merged, 24, activation='relu') #outputs 37 ch
x2_ident = AveragePooling3D()(x1_ident)
x2_merged = merge([x2_1, x2_ident], mode='concat', concat_axis=1)
#by branching we reduce the #params
x3_ident = AveragePooling3D()(x2_ident)
x3_malig = conv_block(x2_merged, 48,
activation='relu') #outputs 25 + 16 ch = 41
x3_malig_merged = merge([x3_malig, x3_ident], mode='concat', concat_axis=1)
x4_ident = AveragePooling3D()(x3_ident)
x4_malig = conv_block(x3_malig_merged, 64,
activation='relu') #outputs 25 + 16 ch = 41
x4_merged = merge([x4_malig, x4_ident], mode='concat', concat_axis=1)
x5_malig = conv_block(x4_merged, 64) #outputs 25 + 16 ch = 41
xpool_malig = BatchNormalization(momentum=0.995)(
GlobalMaxPooling3D()(x5_malig))
xout_malig = Dense(1, name='o_mal',
activation='softplus')(xpool_malig) #relu output
x5_diam = conv_block(x4_merged, 64) #outputs 25 + 16 ch = 41
xpool_diam = BatchNormalization(momentum=0.995)(
GlobalMaxPooling3D()(x5_diam))
xout_diam = Dense(1, name='o_diam',
activation='softplus')(xpool_diam) #relu output
x5_lob = conv_block(x4_merged, 64) #outputs 25 + 16 ch = 41
xpool_lob = BatchNormalization(momentum=0.995)(
GlobalMaxPooling3D()(x5_lob))
xout_lob = Dense(1, name='o_lob',
activation='softplus')(xpool_lob) #relu output
x5_spic = conv_block(x4_merged, 64) #outputs 25 + 16 ch = 41
xpool_spic = BatchNormalization(momentum=0.995)(
GlobalMaxPooling3D()(x5_spic))
xout_spic = Dense(1, name='o_spic',
activation='softplus')(xpool_spic) #relu output
model = Model(input=xin,
output=[xout_diam, xout_lob, xout_spic, xout_malig])
if input_shape[1] == 32:
lr_start = .01
elif input_shape[1] == 64:
lr_start = .003
elif input_shape[1] == 128:
lr_start = .002
# elif input_shape[1] == 96:
# lr_start = 5e-4
opt = Nadam(lr_start, clipvalue=1.0)
print 'compiling model'
model.compile(optimizer=opt,
loss='mse',
loss_weights={
'o_diam': 0.06,
'o_lob': 0.5,
'o_spic': 0.5,
'o_mal': 1.0
})
return model
def main():
# load data
print("loading data...")
ts = time.time()
datapath = os.path.join(Paramater.DATAPATH, "2016", month)
if is_mmn:
fname = os.path.join(
datapath, 'CACHE', 'TaxiBJ_C{}_P{}_T{}_{}_mmn_speed.h5'.format(
len_closeness, len_period, len_trend,
"External" if hasExternal else "noExternal"))
else:
fname = os.path.join(
datapath, 'CACHE', 'TaxiBJ_C{}_P{}_T{}_{}_speed.h5'.format(
len_closeness, len_period, len_trend,
"External" if hasExternal else "noExternal"))
pkl = fname + '.preprocessing_speed.pkl'
fn = "48_48_20_LinearInterpolationFixed"
if os.path.exists(fname) and CACHEDATA:
X_train, Y_train, X_test, Y_test, mmn, external_dim, \
timestamp_train, timestamp_test, noConditionRegions, x_num, y_num, z_num = read_cache(fname, is_mmn,
pkl)
print("load %s successfully" % fname)
else:
datapaths = [os.path.join(datapath, fn)]
noConditionRegionsPath = os.path.join(datapath,
"48_48_20_noSpeedRegion_0.05")
X_train, Y_train, X_test, Y_test, mmn, external_dim, timestamp_train, timestamp_test, noConditionRegions, \
x_num, y_num, z_num = Data.loadDataFromRaw(
paths=datapaths, noSpeedRegionPath=noConditionRegionsPath, nb_flow=nb_flow, len_closeness=len_closeness,
len_period=len_period, len_trend=len_trend
, len_test=len_test, maxMinNormalization=is_mmn, preprocess_name=pkl,
meta_data=hasExternal,
meteorol_data=hasExternal,
holiday_data=hasExternal, isComplete=False)
if CACHEDATA:
cache(fname, X_train, Y_train, X_test, Y_test, external_dim,
timestamp_train, timestamp_test, noConditionRegions, is_mmn,
x_num, y_num, nb_flow)
z_num = nb_flow
# print("\n days (test): ", [v[:8] for v in timestamp_test[0::72]])
print("\nelapsed time (loading data): %.3f seconds\n" % (time.time() - ts))
print('=' * 10)
print("compiling model_train...")
print(
"**at the first time, it takes a few minites to compile if you use [Theano] as the backend**"
)
ts = time.time()
X_train, Y_train = Data.getSequenceXY(X_train, Y_train, step)
Y_train_final = Y_train[:, -1]
X_test, Y_test = Data.getSequenceXY(X_test, Y_test, step)
Y_test_final = Y_test[:, -1]
X_train.append(Y_train)
X_test.append(Y_test)
timestamp_train = timestamp_train[step - 1:]
timestamp_test = timestamp_test[step - 1:]
if use_diff_test:
X_test_old = X_test
Y_test_old = Y_test
import pandas as pd
df_diff = pd.read_csv("./data/2016/all/" + fn + "_diff.csv",
index_col=0)
# 大于200 有335个作为test
test_time = df_diff[df_diff["diff"] > 200]["time"].values
timestamp_train_dict = dict(
zip(timestamp_train, range(len(timestamp_train))))
timestamp_test_dict = dict(
zip(timestamp_test, range(len(timestamp_test))))
new_X_test = []
new_Y_test = []
if isinstance(X_train, list):
for _ in range(len(X_train)):
new_X_test.append([])
for _test_time in test_time:
_test_time = str(_test_time)
if (_test_time in timestamp_train_dict):
index = timestamp_train_dict[_test_time]
if isinstance(X_train, list):
for i in range(len(X_train)):
new_X_test[i].append(X_train[i][index])
else:
new_X_test.append(X_train[index])
new_Y_test.append(Y_train[index])
if (_test_time in timestamp_test_dict):
index = timestamp_test_dict[_test_time]
if isinstance(X_test_old, list):
for i in range(len(X_test_old)):
new_X_test[i].append(X_test_old[i][index])
else:
new_X_test.append(X_test_old[index])
new_Y_test.append(Y_test_old[index])
# if (_test_time not in timestamp_train_dict and _test_time not in timestamp_test_dict):
# print(_test_time)
if isinstance(new_X_test, list):
for i in range(len(new_X_test)):
new_X_test[i] = np.stack(new_X_test[i], axis=0)
else:
new_X_test = np.stack(new_X_test, axis=0)
new_Y_test = np.stack(new_Y_test, axis=0)
# if isinstance(new_X_test, list):
# for i in range(len(new_X_test)):
# print(new_X_test[i].shape)
# else:
# print(new_X_test.shape)
# print(new_Y_test.shape)
X_test = new_X_test
Y_test = new_Y_test
Y_test_final = Y_test[:, -1]
# print "X_test len:", len(X_test)
# for x in X_test:
# print x.shape
# print Y_test.shape
# print z_num, x_num, y_num
print "start build model_train"
outputs = []
inputs = []
resUnit_share_layers = []
resUnit_share_layers2 = []
resUnit_share_layers3 = []
shared_conv1 = Convolution2D(filters=64,
kernel_size=(3, 3),
padding="same")
shared_conv2 = Convolution2D(nb_filter=nb_flow,
nb_row=3,
nb_col=3,
border_mode="same")
shared_conv3 = Convolution2D(filters=64,
kernel_size=(3, 3),
padding="same")
shared_conv4 = Convolution2D(nb_filter=nb_flow,
nb_row=3,
nb_col=3,
border_mode="same")
shared_conv5 = Convolution2D(filters=64,
kernel_size=(3, 3),
padding="same")
shared_conv6 = Convolution2D(nb_filter=nb_flow,
nb_row=3,
nb_col=3,
border_mode="same")
shared_convLSTM_period = ConvLSTM2D(nb_filter=32,
nb_row=3,
nb_col=3,
border_mode="same")
shared_conv_period = Convolution2D(nb_filter=nb_flow,
nb_row=3,
nb_col=3,
border_mode="same")
shared_convLSTM_trend = ConvLSTM2D(nb_filter=32,
nb_row=3,
nb_col=3,
border_mode="same")
shared_conv_trend = Convolution2D(nb_filter=nb_flow,
nb_row=3,
nb_col=3,
border_mode="same")
shared_ilayers = []
shared_embeding = Dense(output_dim=10)
shared_embeding2 = Dense(output_dim=nb_flow * x_num * y_num)
assert l < step
for _ in range(step):
main_outputs = []
if len_closeness > 0:
input = Input(shape=(nb_flow * len_closeness, x_num, y_num))
inputs.append(input)
# Conv1
conv1 = shared_conv1(input)
# [nb_residual_unit] Residual Units
resUnit_share_index = [0]
residual_output = ResUnits(_residual_unit,
nb_filter=64,
repetations=nb_residual_unit,
share=True,
shareIndex=resUnit_share_index,
shares=resUnit_share_layers)(conv1)
# Conv2
activation = Activation('relu')(residual_output)
conv2 = shared_conv2(activation)
main_outputs.append(conv2)
# input = Input(shape=(nb_flow * len_closeness, x_num, y_num))
# inputs.append(input)
# # conv1 = Convolution2D(nb_filter=64, nb_row=3, nb_col=3, border_mode="same")(input)
# # act1 = Activation("relu")(conv1)
# reshape = Reshape((len_closeness, nb_flow, x_num, y_num))(input)
# convLSTM = ConvLSTM2D(nb_filter=32, nb_row=3, nb_col=3, border_mode="same")(reshape)
# act2 = Activation("relu")(convLSTM)
# conv2 = Convolution2D(nb_filter=nb_flow, nb_row=3, nb_col=3, border_mode="same")(act2)
# main_outputs.append(conv2)
if len_period > 0:
input = Input(shape=(nb_flow * len_period, x_num, y_num))
inputs.append(input)
# Conv1
conv1 = shared_conv3(input)
# [nb_residual_unit] Residual Units
resUnit_share_index = [0]
residual_output = ResUnits(_residual_unit,
nb_filter=64,
repetations=nb_residual_unit,
share=True,
shareIndex=resUnit_share_index,
shares=resUnit_share_layers2)(conv1)
# Conv2
activation = Activation('relu')(residual_output)
conv2 = shared_conv4(activation)
main_outputs.append(conv2)
# input = Input(shape=(nb_flow * len_period, x_num, y_num))
# inputs.append(input)
# # conv1 = Convolution2D(nb_filter=64, nb_row=3, nb_col=3, border_mode="same")(input)
# # act1 = Activation("relu")(conv1)
# input = Reshape((len_period, nb_flow, x_num, y_num))(input)
# convLSTM = shared_convLSTM_period(input)
# act2 = Activation("relu")(convLSTM)
# conv2 = shared_conv_period(act2)
# main_outputs.append(conv2)
if len_trend > 0:
input = Input(shape=(nb_flow * len_trend, x_num, y_num))
inputs.append(input)
# Conv1
conv1 = shared_conv5(input)
# [nb_residual_unit] Residual Units
resUnit_share_index = [0]
residual_output = ResUnits(_residual_unit,
nb_filter=64,
repetations=nb_residual_unit,
share=True,
shareIndex=resUnit_share_index,
shares=resUnit_share_layers3)(conv1)
# Conv2
activation = Activation('relu')(residual_output)
conv2 = shared_conv6(activation)
main_outputs.append(conv2)
# input = Input(shape=(nb_flow * len_trend, x_num, y_num))
# inputs.append(input)
# # conv1 = Convolution2D(nb_filter=64, nb_row=3, nb_col=3, border_mode="same")(input)
# # act1 = Activation("relu")(conv1)
# reshape = Reshape((len_trend, nb_flow, x_num, y_num))(input)
# convLSTM = shared_convLSTM_trend(reshape)
# act2 = Activation("relu")(convLSTM)
# conv2 = shared_conv_trend(act2)
# main_outputs.append(conv2)
if len(main_outputs) == 1:
main_output = main_outputs[0]
else:
new_outputs = []
for index, output in enumerate(main_outputs):
if (len(shared_ilayers) <= index):
shared_ilayers.append(iLayer())
new_outputs.append(shared_ilayers[index](output))
main_output = merge(new_outputs, mode='sum')
if external_dim != None and external_dim > 0:
# external input
external_input = Input(shape=(external_dim, ))
inputs.append(external_input)
embedding = shared_embeding(external_input)
embedding = Activation('relu')(embedding)
h1 = shared_embeding2(embedding)
activation = Activation('relu')(h1)
external_output = Reshape((nb_flow, x_num, y_num))(activation)
main_output = merge([main_output, external_output], mode='sum')
main_output = Activation('tanh')(main_output)
outputs.append(main_output)
main_output = merge(outputs, mode="concat", concat_axis=1)
predict_sequence = Reshape((step, z_num, x_num, y_num))(main_output)
input_targets = Input(shape=(step, z_num, x_num, y_num),
name="input_targets")
inputs.append(input_targets)
main_output = eRNN(error_hidden_dim, (z_num, x_num, y_num), l,
False)([predict_sequence, input_targets])
model_train = Model(inputs=inputs, outputs=[predict_sequence, main_output])
adam = Adam(lr=lr)
model_train.compile(loss=['mse', 'mse'],
loss_weights=[0.2, 1],
optimizer=adam,
metrics=[metrics.rmse])
# model_train.compile(loss=lambda y_true,y_preiod: K.mean(K.square(y_preiod - y_true), axis=-1), optimizer=adam, metrics=[metrics.rmse])
# model_predict = Model(input=inputs, output=main_output)
# model_predict.compile(optimizer=adam,loss="mse",metrics=metrics.rmse)
model_train.summary()
print "finish build model_train"
hyperparams_name = 'testMyModel4(ernn_{}_h{}_l{}_step{})_speed.c{}.p{}.t{}.resunit{}.lr{}.{}.{}'.format(
ernn_weight, error_hidden_dim, l, step, len_closeness, len_period,
len_trend, nb_residual_unit, lr,
"External" if hasExternal else "noExternal",
"MMN" if is_mmn else "noMMN")
fname_param = os.path.join(path_model,
'{}.best.h5'.format(hyperparams_name))
early_stopping = EarlyStopping(monitor='val_e_rnn_1_rmse',
patience=4,
mode='min')
model_checkpoint = ModelCheckpoint(fname_param,
monitor='val_e_rnn_1_rmse',
verbose=0,
save_best_only=True,
mode='min',
save_weights_only=True)
print("\nelapsed time (compiling model_train): %.3f seconds\n" %
(time.time() - ts))
print('=' * 10)
print("training model_train...")
ts = time.time()
history = model_train.fit(X_train, [Y_train, Y_train_final],
epochs=nb_epoch,
batch_size=batch_size,
validation_split=0.1,
callbacks=[early_stopping, model_checkpoint],
verbose=1)
model_train.save_weights(os.path.join(path_model,
'{}.h5'.format(hyperparams_name)),
overwrite=True)
pickle.dump((history.history),
open(
os.path.join(path_result,
'{}.history.pkl'.format(hyperparams_name)),
'wb'))
print("\nelapsed time (training): %.3f seconds\n" % (time.time() - ts))
print('=' * 10)
print(
'evaluating using the model_train that has the best loss on the valid set'
)
ts = time.time()
model_train.load_weights(fname_param)
score = model_train.evaluate(X_train, [Y_train, Y_train_final],
batch_size=Y_train.shape[0] // 48,
verbose=0)
if is_mmn:
print('Train score: %.6f rmse (norm): %.6f rmse (real): %.6f' %
(score[0], score[1], score[1] * (mmn._max - mmn._min) / 2.))
else:
print('Train score: %.6f rmse (real): %.6f' % (score[0], score[1]))
score = model_train.evaluate(X_test, [Y_test, Y_test_final],
batch_size=Y_test.shape[0] // 12,
verbose=0)
if is_mmn:
print('Test score: %.6f rmse (norm): %.6f rmse (real): %.6f' %
(score[0], score[1], score[1] * (mmn._max - mmn._min) / 2.))
else:
print('Test score: %.6f rmse (real): %.6f' % (score[0], score[1]))
if not is_mmn:
predict = model_train.predict(X_test)[1]
else:
predict = mmn.inverse_transform(model_train.predict(X_test)[1])
Y_test_final = mmn.inverse_transform(Y_test_final)
# predict = predict[:, -1]
# Y_test = Y_test[:, -1]
# print("predict", predict)
# print("test", Y_test_final)
rmse = round(Metric.RMSE(predict, Y_test_final, noConditionRegions), 5)
save_result(
predict, Y_test_final, timestamp_test,
"./result/{}_predict_rmse{}".format(hyperparams_name, str(rmse)))
print("RMSE:", rmse)
# print("accuracy", Metric.accuracy(predict, Y_test, noConditionRegions))
print("\nelapsed time (eval): %.3f seconds\n" % (time.time() - ts))
exit(1)
for n_gram in filter_sizes:
conv_name[i] = str('conv_' + str(n_gram))
conv_name[i] = Convolution1D(nb_filter=nb_filter,
filter_length=n_gram,
border_mode='valid',
activation='relu',
subsample_length=1,
input_dim=embeddings_dim,
input_length=max_sent_len)(Drop1)
pool_name[i] = str('maxpool_' + str(n_gram))
pool_name[i] = MaxPooling1D(pool_length=max_sent_len - n_gram + 1)(
conv_name[i])
flat_name[i] = str('flat_' + str(n_gram))
flat_name[i] = Flatten()(pool_name[i])
i += 1
merged = merge([flat_name[0], flat_name[1], flat_name[2]], mode='concat')
droput_final = Dropout(dropout_prob[1])(merged)
Dense_final = Dense(1,
input_dim=nb_filter * len(filter_sizes))(droput_final)
Out = Dense(1, activation='sigmoid')(Dense_final)
model = Model(inputs=main_input, outputs=Out)
# Print model summary
# ==================================================
print(model.summary())
#Visualize the model in a graph
#1st method
#from IPython.display import SVG
#from keras.utils.vis_utils import model_to_dot
#from keras.utils import vis_utils
def unet(X,
Y,
train_idxs,
test_idxs,
nb_epoch=20,
optomizer='adagrad',
batch_size=32):
# Architecture from: https://github.com/jocicmarko/ultrasound-nerve-segmentation
# Original paper: http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net/
import numpy as np
from keras.models import Model
from keras.layers import Input, merge, Convolution2D, MaxPooling2D, UpSampling2D
#from keras.optimizers import Adam
np.random.seed(2016) # For reproducibility:
#from sklearn.metrics import (precision_score, recall_score, f1_score, accuracy_score)
# Process data:
X_train, X_test, Y_train, Y_test, N_samples, channels, img_rows, img_cols = process_XY(
X, Y, train_idxs, test_idxs, mask=True)
#inputs = Input((1, X.shape[0], img_cols))
inputs = Input((channels, img_rows, img_cols))
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(inputs)
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(pool1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(pool2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(pool3)
conv4 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Convolution2D(512, 3, 3, activation='relu',
border_mode='same')(pool4)
conv5 = Convolution2D(512, 3, 3, activation='relu',
border_mode='same')(conv5)
up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4],
mode='concat',
concat_axis=1)
conv6 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(up6)
conv6 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(conv6)
up7 = merge([UpSampling2D(size=(2, 2))(conv6), conv3],
mode='concat',
concat_axis=1)
conv7 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(up7)
conv7 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv7)
up8 = merge([UpSampling2D(size=(2, 2))(conv7), conv2],
mode='concat',
concat_axis=1)
conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up8)
conv8 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv8)
up9 = merge([UpSampling2D(size=(2, 2))(conv8), conv1],
mode='concat',
concat_axis=1)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up9)
conv9 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv9)
conv10 = Convolution2D(1, 1, 1, activation='sigmoid')(conv9)
model = Model(input=inputs, output=conv10)
model.compile(optimizer=optomizer,
loss=dice_coef_loss,
metrics=[dice_coef]) #Adam(lr=1e-5)
model.fit(X_train,
Y_train,
batch_size=batch_size,
nb_epoch=nb_epoch,
verbose=1,
validation_data=(X_test, Y_test),
shuffle=True)
import time
time.sleep(1)
if len(X_test) > 0:
score = []
Y_pred = model.predict(X_test)
dc = dice_coef(Y_test, Y_pred)
score = dc
print('Dice coefficient: ' + str(dc))
# y_pred = model.predict(X_test)
# y_pred = y_pred[:,0]
# accuracy = accuracy_score(Y_test, y_pred)
# recall = recall_score(Y_test, y_pred)
# precision = precision_score(Y_test, y_pred)
# f1 = f1_score(Y_test, y_pred)
# print('Accuracy: {}'.format(accuracy))
# print('Recall: {}'.format(recall))
# print('Precision: {}'.format(precision))
# print('F1: {}'.format(f1))
# score = np.array([accuracy, recall, precision, f1])
else:
score = []
return model, score
# normalize test date
X_test_uid=np.array(X_test_uid)
X_test_uid=X_test_uid.reshape(X_test_uid.shape[0],1)
X_test_iid=np.array(X_test_iid)
X_test_iid=X_test_iid.reshape(X_test_iid.shape[0],1)
# define model
input_1=Input(shape=(1,), dtype='int32')
input_2=Input(shape=(1,), dtype='int32')
x1=Embedding(output_dim=128, input_dim=223970, input_length=1)(input_1)
x2=Embedding(output_dim=128, input_dim=14726, input_length=1)(input_2)
x1=Flatten()(x1)
x2=Flatten()(x2)
x = merge([x1, x2], mode='concat')
x = Dropout(0.2)(x)
x = Dense(512, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(64, activation='relu')(x)
x = Dropout(0.2)(x)
out = Dense(1, activation='sigmoid')(x)
model = Model(input=[input_1, input_2], output=out)
model.compile(optimizer='rmsprop',
loss='mean_squared_error',
metrics=[])
# train model
model.fit([X_train_uid, X_train_iid], Y_train_score,
nb_epoch=10, batch_size=1024*6)
# predict
def smart_merge(vectors, **kwargs):
return vectors[0] if len(vectors) == 1 else merge(vectors, **kwargs)
def create_model(X_vocab_len, X_max_len, y_vocab_len, y_max_len,
n_phonetic_features, y1, n1, y2, n2, y3, n3, y4, n4, y5, n5,
y6, n6, hidden_size, num_layers):
def smart_merge(vectors, **kwargs):
return vectors[0] if len(vectors) == 1 else merge(vectors, **kwargs)
current_word = Input(shape=(X_max_len, ), dtype='float32',
name='input1') # for encoder (shared)
decoder_input = Input(shape=(X_max_len, ), dtype='float32',
name='input3') # for decoder -- attention
right_word1 = Input(shape=(X_max_len, ), dtype='float32', name='input4')
right_word2 = Input(shape=(X_max_len, ), dtype='float32', name='input5')
right_word3 = Input(shape=(X_max_len, ), dtype='float32', name='input6')
right_word4 = Input(shape=(X_max_len, ), dtype='float32', name='input7')
left_word1 = Input(shape=(X_max_len, ), dtype='float32', name='input8')
left_word2 = Input(shape=(X_max_len, ), dtype='float32', name='input9')
left_word3 = Input(shape=(X_max_len, ), dtype='float32', name='input10')
left_word4 = Input(shape=(X_max_len, ), dtype='float32', name='input11')
phonetic_input = Input(shape=(n_phonetic_features, ),
dtype='float32',
name='input12')
emb_layer1 = Embedding(X_vocab_len,
EMBEDDING_DIM,
input_length=X_max_len,
mask_zero=False,
name='Embedding')
list_of_inputs = [
current_word, right_word1, right_word2, right_word3, right_word4,
left_word1, left_word2, left_word3, left_word4
]
current_word_embedding, right_word_embedding1, right_word_embedding2,right_word_embedding3, right_word_embedding4, \
left_word_embedding1, left_word_embedding2, left_word_embedding3, left_word_embedding4 = [emb_layer1(i) for i in list_of_inputs]
print("Type:: ", type(current_word_embedding))
list_of_embeddings1 = [current_word_embedding, right_word_embedding1, right_word_embedding2,right_word_embedding3, right_word_embedding4, \
left_word_embedding1, left_word_embedding2, left_word_embedding3, left_word_embedding4]
list_of_embeddings = [
Dropout(0.50, name='drop1_' + str(j))(i)
for i, j in zip(list_of_embeddings1, range(len(list_of_embeddings1)))
]
list_of_embeddings = [
GaussianNoise(0.05, name='noise1_' + str(j))(i)
for i, j in zip(list_of_embeddings, range(len(list_of_embeddings)))
]
conv4_curr, conv4_right1, conv4_right2, conv4_right3, conv4_right4, conv4_left1, conv4_left2, conv4_left3, conv4_left4 =\
[Conv1D(filters=no_filters,
kernel_size=4, padding='valid',activation='relu',
strides=1, name='conv4_'+str(j))(i) for i,j in zip(list_of_embeddings, range(len(list_of_embeddings)))]
conv4s = [
conv4_curr, conv4_right1, conv4_right2, conv4_right3, conv4_right4,
conv4_left1, conv4_left2, conv4_left3, conv4_left4
]
maxPool4 = [
MaxPooling1D(name='max4_' + str(j))(i)
for i, j in zip(conv4s, range(len(conv4s)))
]
avgPool4 = [
AveragePooling1D(name='avg4_' + str(j))(i)
for i, j in zip(conv4s, range(len(conv4s)))
]
pool4_curr, pool4_right1, pool4_right2, pool4_right3, pool4_right4, pool4_left1, pool4_left2, pool4_left3, pool4_left4 = \
[merge([i,j], name='merge_conv4_'+str(k)) for i,j,k in zip(maxPool4, avgPool4, range(len(maxPool4)))]
conv5_curr, conv5_right1, conv5_right2, conv5_right3, conv5_right4, conv5_left1, conv5_left2, conv5_left3, conv5_left4 = \
[Conv1D(filters=no_filters,
kernel_size=5,
padding='valid',
activation='relu',
strides=1, name='conv5_'+str(j))(i) for i,j in zip(list_of_embeddings, range(len(list_of_embeddings)))]
conv5s = [
conv5_curr, conv5_right1, conv5_right2, conv5_right3, conv5_right4,
conv5_left1, conv5_left2, conv5_left3, conv5_left4
]
maxPool5 = [
MaxPooling1D(name='max5_' + str(j))(i)
for i, j in zip(conv5s, range(len(conv5s)))
]
avgPool5 = [
AveragePooling1D(name='avg5_' + str(j))(i)
for i, j in zip(conv5s, range(len(conv5s)))
]
pool5_curr, pool5_right1, pool5_right2, pool5_right3, pool5_right4, pool5_left1, pool5_left2, pool5_left3, pool5_left4 = \
[merge([i,j], name='merge_conv5_'+str(k)) for i,j,k in zip(maxPool5, avgPool5, range(len(maxPool5)))]
maxPools = [pool4_curr, pool4_right1, pool4_right2, pool4_right3, pool4_right4, \
pool4_left1, pool4_left2, pool4_left3, pool4_left4, \
pool5_curr, pool5_right1, pool5_right2, pool5_right3, pool5_right4, \
pool5_left1, pool5_left2, pool5_left3, pool5_left4]
concat = merge(maxPools, mode='concat', name='main_merge')
x = Dropout(0.15, name='drop_single1')(concat)
x = Bidirectional(RNN(rnn_output_size), name='bidirec1')(x)
total_features = [x, phonetic_input]
concat2 = merge(total_features, mode='concat', name='phonetic_merging')
x = Dense(HIDDEN_DIM,
activation='relu',
kernel_initializer='he_normal',
kernel_constraint=maxnorm(3),
bias_constraint=maxnorm(3),
name='dense1')(concat2)
x = Dropout(0.15, name='drop_single2')(x)
x = Dense(HIDDEN_DIM,
kernel_initializer='he_normal',
activation='tanh',
kernel_constraint=maxnorm(3),
bias_constraint=maxnorm(3),
name='dense2')(x)
x = Dropout(0.15, name='drop_single3')(x)
out1 = Dense(n1,
kernel_initializer='he_normal',
activation='softmax',
name='output1')(x)
out2 = Dense(n2,
kernel_initializer='he_normal',
activation='softmax',
name='output2')(x)
out3 = Dense(n3,
kernel_initializer='he_normal',
activation='softmax',
name='output3')(x)
out4 = Dense(n4,
kernel_initializer='he_normal',
activation='softmax',
name='output4')(x)
out5 = Dense(n5,
kernel_initializer='he_normal',
activation='softmax',
name='output5')(x)
out6 = Dense(n6,
kernel_initializer='he_normal',
activation='softmax',
name='output6')(x)
# Luong et al. 2015 attention model
emb_layer = Embedding(X_vocab_len,
EMBEDDING_DIM,
input_length=X_max_len,
mask_zero=True,
name='Embedding_for_seq2seq')
current_word_embedding, right_word_embedding1, right_word_embedding2,right_word_embedding3, right_word_embedding4, \
left_word_embedding1, left_word_embedding2, left_word_embedding3, left_word_embedding4 = [emb_layer(i) for i in list_of_inputs]
# current_word_embedding = smart_merge([ current_word_embedding, right_word_embedding1, left_word_embedding1])
encoder, state = GRU(rnn_output_size,
return_sequences=True,
unroll=True,
return_state=True,
name='encoder')(current_word_embedding)
encoder_last = encoder[:, -1, :]
decoder = emb_layer(decoder_input)
decoder = GRU(rnn_output_size,
return_sequences=True,
unroll=True,
name='decoder')(decoder, initial_state=[encoder_last])
attention = dot([decoder, encoder], axes=[2, 2], name='dot')
attention = Activation('softmax', name='attention')(attention)
context = dot([attention, encoder], axes=[2, 1], name='dot2')
decoder_combined_context = concatenate([context, decoder],
name='concatenate')
outputs = TimeDistributed(Dense(64, activation='tanh'),
name='td1')(decoder_combined_context)
outputs = TimeDistributed(Dense(X_vocab_len, activation='softmax'),
name='td2')(outputs)
all_inputs = [current_word, decoder_input, right_word1, right_word2, right_word3, right_word4, left_word1, left_word2, left_word3,\
left_word4, phonetic_input]
all_outputs = [outputs, out1, out2, out3, out4, out5, out6]
model = Model(input=all_inputs, output=all_outputs)
opt = Adam()
return model
def building_residual_block(input_shape,
n_feature_maps,
kernel_sizes=None,
n_skip=2,
is_subsample=False,
subsample=None):
'''
[1] Building block of layers for residual learning.
Code based on https://github.com/ndronen/modeling/blob/master/modeling/residual.py
, but modification of (perhaps) incorrect relu(f)+x thing and it's for conv layer
[2] MaxPooling is used instead of strided convolution to make it easier
to set size(output of short-cut) == size(output of conv-layers).
If you want to remove MaxPooling,
i) change (border_mode in Convolution2D in shortcut), 'same'-->'valid'
ii) uncomment ZeroPadding2D in conv layers.
(Then the following Conv2D is not the first layer of this container anymore,
so you can remove the input_shape in the line 101, the line with comment #'OPTION' )
[3] It can be used for both cases whether it subsamples or not.
[4] In the short-cut connection, I used 1x1 convolution to increase #channel.
It occurs when is_expand_channels == True
input_shape = (None, num_channel, height, width)
n_feature_maps: number of feature maps. In ResidualNet it increases whenever image is downsampled.
kernel_sizes : list or tuple, (3,3) or [3,3] for example
n_skip : number of layers to skip
is_subsample : If it is True, the layers subsamples by *subsample* to reduce the size.
subsample : tuple, (2,2) or (1,2) for example. Used only if is_subsample==True
'''
# ***** VERBOSE_PART *****
print(' - New residual block with')
print(' input shape:', input_shape)
print(' kernel size:', kernel_sizes)
# is_expand_channels == True when num_channels increases.
# E.g. the very first residual block (e.g. 1->64, 3->128, 128->256, ...)
is_expand_channels = not (input_shape[0] == n_feature_maps)
if is_expand_channels:
print(' - Input channels: %d ---> num feature maps on out: %d' %
(input_shape[0], n_feature_maps))
if is_subsample:
print(' - with subsample:', subsample)
kernel_row, kernel_col = kernel_sizes
# set input
x = Input(shape=(input_shape))
# ***** SHORTCUT PATH *****
if is_subsample: # subsample (+ channel expansion if needed)
shortcut_y = Convolution2D(n_feature_maps,
kernel_row,
kernel_col,
subsample=subsample,
W_regularizer=l2(0.0001),
border_mode='valid')(x)
else: # channel expansion only (e.g. the very first layer of the whole networks)
if is_expand_channels:
shortcut_y = Convolution2D(n_feature_maps,
1,
1,
W_regularizer=l2(0.0001),
border_mode='same')(x)
else:
# if no subsample and no channel expension, there's nothing to add on the shortcut.
shortcut_y = x
# ***** CONVOLUTION_PATH *****
conv_y = x
for i in range(n_skip):
conv_y = BatchNormalization(axis=1)(conv_y)
conv_y = Activation('relu')(conv_y)
if i == 0 and is_subsample: # [Subsample at layer 0 if needed]
conv_y = Convolution2D(n_feature_maps,
kernel_row,
kernel_col,
subsample=subsample,
W_regularizer=l2(0.0001),
border_mode='valid')(conv_y)
else:
conv_y = Convolution2D(n_feature_maps,
kernel_row,
kernel_col,
W_regularizer=l2(0.0001),
border_mode='same')(conv_y)
# output
y = merge([shortcut_y, conv_y], mode='sum')
block = Model(input=x, output=y)
print(' -- model was built.')
return block
def get_unet():
inputs = Input((8, ISZ, ISZ))
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(inputs)
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(pool1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(pool2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(pool3)
conv4 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Convolution2D(512, 3, 3, activation='relu',
border_mode='same')(pool4)
conv5 = Convolution2D(512, 3, 3, activation='relu',
border_mode='same')(conv5)
up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4],
mode='concat',
concat_axis=1)
conv6 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(up6)
conv6 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(conv6)
up7 = merge([UpSampling2D(size=(2, 2))(conv6), conv3],
mode='concat',
concat_axis=1)
conv7 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(up7)
conv7 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv7)
up8 = merge([UpSampling2D(size=(2, 2))(conv7), conv2],
mode='concat',
concat_axis=1)
conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up8)
conv8 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv8)
up9 = merge([UpSampling2D(size=(2, 2))(conv8), conv1],
mode='concat',
concat_axis=1)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up9)
conv9 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv9)
conv10 = Convolution2D(N_Cls, 1, 1, activation='sigmoid')(conv9)
model = Model(input=inputs, output=conv10)
model.compile(optimizer=Adam(),
loss='binary_crossentropy',
metrics=[jaccard_coef, jaccard_coef_int, 'accuracy'])
return model
dropout_U=0.2,
return_sequences=True,
W_regularizer=l2(0.01))(x5)
else:
x5 = LSTM(300,
dropout_W=0.2,
dropout_U=0.2,
return_sequences=True)(x5)
attention5 = TimeDistributed(Dense(1, activation='tanh'))(x5)
attention5 = Flatten()(attention5)
attention5 = Activation('softmax')(attention5)
attention5 = RepeatVector(600)(attention5)
attention5 = Permute([2, 1])(attention5)
merge5 = merge([x5, attention5], mode='mul')
merge5 = Lambda(lambda xin: K.sum(xin, axis=1))(merge5)
merge5 = Dense(300, activation='softmax')(merge5)
model5 = Model(input=model5_ip, output=merge5)
print model5.summary()
else:
model5 = Sequential()
model5.add(
Embedding(len(word_index) + 1, 300, input_length=40, dropout=0.2))
if opts.cnn == 1:
model5.add(
Conv1D(64, 5, padding='valid', activation='relu', strides=1))
model5.add(MaxPooling1D(pool_size=4))
if opts.bilstm == 1:
if opts.regularize == 1:
def base_conv(inp):
branch = Convolution2D(32,
3,
3,
border_mode='same',
activation='relu',
init='glorot_uniform')(inp)
branch = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(branch)
return branch
for i in range(x_train.shape[0]):
x = base_conv(inputs[i])
convbranches.append(x)
merged_model = merge([branch for branch in convbranches],
mode='sum',
concat_axis=-1)
merged_model = Convolution2D(128,
3,
3,
border_mode='same',
activation='relu',
init='glorot_uniform')(merged_model)
merged_model = Convolution2D(128,
3,
3,
border_mode='same',
activation='relu',
init='glorot_uniform')(merged_model)
merged_model = MaxPooling2D(pool_size=(2, 2), strides=(2, 2))(merged_model)
#merged_model = Convolution2D(256, 3, 3, border_mode = 'same', activation = 'relu', init = 'normal')(merged_model)
def get_unet():
inputs = Input((1, img_rows, img_cols))
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(inputs)
conv1 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(conv1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(pool1)
conv2 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(pool2)
conv3 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(pool3)
conv4 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(conv4)
conv5 = Convolution2D(512, 3, 3, activation='relu',
border_mode='same')(pool4)
conv5 = Convolution2D(512, 3, 3, activation='relu',
border_mode='same')(conv5)
up6 = merge([UpSampling2D(size=(2, 2))(conv5), conv4],
mode='concat',
concat_axis=1)
conv6 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(up6)
conv6 = Convolution2D(256, 3, 3, activation='relu',
border_mode='same')(conv6)
up7 = merge([UpSampling2D(size=(2, 2))(conv6), conv3],
mode='concat',
concat_axis=1)
conv7 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(up7)
conv7 = Convolution2D(128, 3, 3, activation='relu',
border_mode='same')(conv7)
up8 = merge([UpSampling2D(size=(2, 2))(conv7), conv2],
mode='concat',
concat_axis=1)
conv8 = Convolution2D(64, 3, 3, activation='relu', border_mode='same')(up8)
conv8 = Convolution2D(64, 3, 3, activation='relu',
border_mode='same')(conv8)
up9 = merge([UpSampling2D(size=(2, 2))(conv8), conv1],
mode='concat',
concat_axis=1)
conv9 = Convolution2D(32, 3, 3, activation='relu', border_mode='same')(up9)
conv9 = Convolution2D(32, 3, 3, activation='relu',
border_mode='same')(conv9)
conv10 = Convolution2D(1, 1, 1, activation='sigmoid')(conv9)
model = Model(input=inputs, output=conv10)
model.compile(optimizer=Adam(lr=1.0e-5),
loss=dice_coef_loss,
metrics=[dice_coef])
return model
os.environ['KERAS_BACKEND'] = 'mxnet' # os.environ['KERAS_BACKEND']='tensorflow' from keras.layers import Input, Lambda, merge from keras.models import Model import numpy as np i1 = np.array([[2, 3], [2, 3], [2, 3], [2, 3], [2, 3]]) i2 = np.array([3, 3, 3, 3, 3]) input1_shape = (2, ) input2_shape = (1, ) model_input1 = Input(shape=input1_shape) model_input2 = Input(shape=input2_shape) z = model_input1 print 'keras shape = ', z._keras_shape print 'shape = ', z.shape z = Lambda(lambda x: x[:, 1:2])(model_input1) print 'keras shape = ', z._keras_shape print 'shape = ', z.shape output = merge([z, model_input2], mode='mul') model = Model([model_input1, model_input2], output) model.compile(optimizer='rmsprop', loss='categorical_crossentropy') print model.predict([i1, i2])
def calibrate(target, source, sourceIndex, predLabel, path):
mmdNetLayerSizes = [25, 25]
l2_penalty = 1e-2
init = lambda shape, name: initializers.normal(
shape, scale=.1e-4, name=name)
space_dim = target.X.shape[1]
calibInput = Input(shape=(space_dim, ))
block1_bn1 = BatchNormalization()(calibInput)
block1_a1 = Activation('relu')(block1_bn1)
block1_w1 = Dense(mmdNetLayerSizes[0],
activation='linear',
kernel_regularizer=l2(l2_penalty),
init=init)(block1_a1)
block1_bn2 = BatchNormalization()(block1_w1)
block1_a2 = Activation('relu')(block1_bn2)
block1_w2 = Dense(space_dim,
activation='linear',
kernel_regularizer=l2(l2_penalty),
init=init)(block1_a2)
block1_output = merge([block1_w2, calibInput], mode='sum')
block2_bn1 = BatchNormalization()(block1_output)
block2_a1 = Activation('relu')(block2_bn1)
block2_w1 = Dense(mmdNetLayerSizes[1],
activation='linear',
kernel_regularizer=l2(l2_penalty),
init=init)(block2_a1)
block2_bn2 = BatchNormalization()(block2_w1)
block2_a2 = Activation('relu')(block2_bn2)
block2_w2 = Dense(space_dim,
activation='linear',
kernel_regularizer=l2(l2_penalty),
init=init)(block2_a2)
block2_output = merge([block2_w2, block1_output], mode='sum')
block3_bn1 = BatchNormalization()(block2_output)
block3_a1 = Activation('relu')(block3_bn1)
block3_w1 = Dense(mmdNetLayerSizes[1],
activation='linear',
kernel_regularizer=l2(l2_penalty),
init=init)(block3_a1)
block3_bn2 = BatchNormalization()(block3_w1)
block3_a2 = Activation('relu')(block3_bn2)
block3_w2 = Dense(space_dim,
activation='linear',
kernel_regularizer=l2(l2_penalty),
init=init)(block3_a2)
block3_output = merge([block3_w2, block2_output], mode='sum')
calibMMDNet = Model(inputs=calibInput, outputs=block3_output)
n = target.X.shape[0]
p = np.random.permutation(n)
toTake = p[range(int(.2 * n))]
targetXMMD = target.X[toTake]
targetYMMD = target.y[toTake]
targetXMMD = targetXMMD[targetYMMD != 0]
targetYMMD = targetYMMD[targetYMMD != 0]
targetYMMD = np.reshape(targetYMMD, (-1, 1))
n = source.X.shape[0]
p = np.random.permutation(n)
toTake = p[range(int(.2 * n))]
sourceXMMD = source.X[toTake]
sourceYMMD = predLabel[toTake]
sourceXMMD = sourceXMMD[sourceYMMD != 0]
sourceYMMD = sourceYMMD[sourceYMMD != 0]
sourceYMMD = np.reshape(sourceYMMD, (-1, 1))
lrate = LearningRateScheduler(step_decay)
optimizer = opt.RMSprop(lr=0.0)
calibMMDNet.compile(
optimizer=optimizer,
loss=lambda y_true, y_pred: cf.MMD(
block3_output, targetXMMD, MMDTargetValidation_split=0.1).
KerasCost(y_true, y_pred))
sourceLabels = np.zeros(sourceXMMD.shape[0])
calibMMDNet.fit(sourceXMMD,
sourceLabels,
epochs=500,
batch_size=1000,
validation_split=0.1,
verbose=0,
callbacks=[
lrate,
mn.monitorMMD(sourceXMMD, sourceYMMD, targetXMMD,
targetYMMD, calibMMDNet.predict),
cb.EarlyStopping(monitor='val_loss',
patience=20,
mode='auto')
])
plt.close('all')
calibMMDNet.save_weights(
os.path.join(io.DeepLearningRoot(),
path + '/ResNet' + str(sourceIndex) + '.h5'))
calibrateSource = Sample(calibMMDNet.predict(source.X), source.y)
calibMMDNet = None
return calibrateSource
def generator_google_mnistM(noise_dim, img_source_dim,img_dest_dim,deterministic,pureGAN,wd,suffix=None):
"""DCGAN generator based on Upsampling and Conv2D
Args:
noise_dim: Dimension of the noise input
img_dim: dimension of the image output
bn_mode: keras batchnorm mode
model_name: model name (default: {"generator_upsampling"})
dset: dataset (default: {"mnist"})
Returns:
keras model
"""
s = img_source_dim[1]
# shp = np.expand_dims(img_dim[1:],1) # to make shp= (None, 1, 28, 28) but is not working
start_dim = int(s / 4)
if K.image_dim_ordering() == "th":
input_channels = img_source_dim[0]
output_channels = img_dest_dim[0]
reshape_shape = (input_channels, s, s)
shp=reshape_shape
else:
input_channels = img_source_dim[-1]
output_channels = img_dest_dim[-1]
reshape_shape = (s, s, input_channels)
shp=reshape_shape
gen_noise_input = Input(shape=noise_dim, name="generator_input")
gen_image_input = Input(shape=shp, name="generator_image_input")
# Noise input and reshaping
x = Dense(5*s*s, input_dim=noise_dim,W_regularizer=l2(wd))(gen_noise_input)
x = Reshape((5,s,s))(x)
x = Activation("relu")(x)
if deterministic: #here I link or not link the noise vector to the whole network
g = gen_image_input
elif pureGAN:
g = x
else:
g = merge([gen_image_input, x], mode='concat',concat_axis=1) # because of concat_axis=1, will it work on tensorflow NHWC too?
x1 = Conv2D(64, (3, 3), border_mode='same', kernel_initializer="he_normal",W_regularizer=l2(wd))(g) #convolved by 3x3 filter to get 64x55x35
x1 = Activation('relu')(x1)
for i in range(4):
x = Conv2D(64, (3, 3), border_mode='same',weight_norm=False, kernel_initializer="he_normal",W_regularizer=l2(wd))(x1)
x=BatchNormGAN(axis=1)(x)
x = Activation('relu')(x)
x = Conv2D(64, (3, 3), border_mode='same', weight_norm=False, kernel_initializer="he_normal",W_regularizer=l2(wd))(x)
x=BatchNormGAN(axis=1)(x)
x1 = merge([x, x1], mode='sum')
x1 = Activation('relu')(x1)
# Last Conv to get the output image
x1 = Conv2D(output_channels, (1, 1),name="gen_conv2d_final", border_mode='same', kernel_initializer="he_normal",W_regularizer=l2(wd))(x1)
x1 = Activation('tanh')(x1)
if suffix is None:
generator_model = Model(input=[gen_noise_input,gen_image_input], output=[x1], name="generator_google1")
else:
generator_model = Model(input=[gen_noise_input,gen_image_input], output=[x1], name="generator_google2")
visualize_model(generator_model)
return generator_model
word_context = np.array(word_context, dtype='int32')
print(couples[:10], labels[:10])
input_target = Input((1, ))
input_context = Input((1, ))
embedding = Embedding(vocab_size,
vector_dim,
input_length=1,
name='embedding')
target = embedding(input_target)
target = Reshape((vector_dim, 1))(target)
context = embedding(input_context)
context = Reshape((vector_dim, 1))(context)
similarity = merge([target, context], mode='cos', dot_axes=0)
# now perform the dot product operation to get a similarity measure
dot_product = merge([target, context], mode='dot', dot_axes=1)
dot_product = Reshape((1, ))(dot_product)
# add the sigmoid output layer
output = Dense(1, activation='sigmoid')(dot_product)
# create the primary training model
model = Model(input=[input_target, input_context], output=output)
model.compile(loss='binary_crossentropy', optimizer='rmsprop')
# create a secondary validation model to run our similarity checks during training
validation_model = Model(input=[input_target, input_context],
output=similarity)
sim_cb = SimilarityCallback(idx2word=index2word)
def googlenet_model(img_rows,
img_cols,
channel=1,
num_classes=None,
model_path="../imagenet_models/"):
"""
GoogLeNet a.k.a. Inception v1 for Keras
Model Schema is based on
https://gist.github.com/joelouismarino/a2ede9ab3928f999575423b9887abd14
ImageNet Pretrained Weights
https://drive.google.com/open?id=0B319laiAPjU3RE1maU9MMlh2dnc
Blog Post:
http://joelouismarino.github.io/blog_posts/blog_googlenet_keras.html
@param img_rows: Rows in input.
@param img_cols: Columns in input.
@param channel: 1 for grayscale, 3 for color.
@param num_classes: Number of class labels for our classification task.
@param model_path: Path containing the ImageNet model.
@return: Model object.
"""
input = Input(shape=(channel, img_rows, img_cols))
conv1_7x7_s2 = Convolution2D(64,
7,
7,
subsample=(2, 2),
border_mode='same',
activation='relu',
name='conv1/7x7_s2',
W_regularizer=l2(0.0002))(input)
conv1_zero_pad = ZeroPadding2D(padding=(1, 1))(conv1_7x7_s2)
pool1_helper = PoolHelper()(conv1_zero_pad)
pool1_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
border_mode='valid',
name='pool1/3x3_s2')(pool1_helper)
pool1_norm1 = LRN(name='pool1/norm1')(pool1_3x3_s2)
conv2_3x3_reduce = Convolution2D(64,
1,
1,
border_mode='same',
activation='relu',
name='conv2/3x3_reduce',
W_regularizer=l2(0.0002))(pool1_norm1)
conv2_3x3 = Convolution2D(192,
3,
3,
border_mode='same',
activation='relu',
name='conv2/3x3',
W_regularizer=l2(0.0002))(conv2_3x3_reduce)
conv2_norm2 = LRN(name='conv2/norm2')(conv2_3x3)
conv2_zero_pad = ZeroPadding2D(padding=(1, 1))(conv2_norm2)
pool2_helper = PoolHelper()(conv2_zero_pad)
pool2_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
border_mode='valid',
name='pool2/3x3_s2')(pool2_helper)
inception_3a_1x1 = Convolution2D(64,
1,
1,
border_mode='same',
activation='relu',
name='inception_3a/1x1',
W_regularizer=l2(0.0002))(pool2_3x3_s2)
inception_3a_3x3_reduce = Convolution2D(
96,
1,
1,
border_mode='same',
activation='relu',
name='inception_3a/3x3_reduce',
W_regularizer=l2(0.0002))(pool2_3x3_s2)
inception_3a_3x3 = Convolution2D(
128,
3,
3,
border_mode='same',
activation='relu',
name='inception_3a/3x3',
W_regularizer=l2(0.0002))(inception_3a_3x3_reduce)
inception_3a_5x5_reduce = Convolution2D(
16,
1,
1,
border_mode='same',
activation='relu',
name='inception_3a/5x5_reduce',
W_regularizer=l2(0.0002))(pool2_3x3_s2)
inception_3a_5x5 = Convolution2D(
32,
5,
5,
border_mode='same',
activation='relu',
name='inception_3a/5x5',
W_regularizer=l2(0.0002))(inception_3a_5x5_reduce)
inception_3a_pool = MaxPooling2D(pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_3a/pool')(pool2_3x3_s2)
inception_3a_pool_proj = Convolution2D(
32,
1,
1,
border_mode='same',
activation='relu',
name='inception_3a/pool_proj',
W_regularizer=l2(0.0002))(inception_3a_pool)
inception_3a_output = merge([
inception_3a_1x1, inception_3a_3x3, inception_3a_5x5,
inception_3a_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_3a/output')
inception_3b_1x1 = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_3b/1x1',
W_regularizer=l2(0.0002))(inception_3a_output)
inception_3b_3x3_reduce = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_3b/3x3_reduce',
W_regularizer=l2(0.0002))(inception_3a_output)
inception_3b_3x3 = Convolution2D(
192,
3,
3,
border_mode='same',
activation='relu',
name='inception_3b/3x3',
W_regularizer=l2(0.0002))(inception_3b_3x3_reduce)
inception_3b_5x5_reduce = Convolution2D(
32,
1,
1,
border_mode='same',
activation='relu',
name='inception_3b/5x5_reduce',
W_regularizer=l2(0.0002))(inception_3a_output)
inception_3b_5x5 = Convolution2D(
96,
5,
5,
border_mode='same',
activation='relu',
name='inception_3b/5x5',
W_regularizer=l2(0.0002))(inception_3b_5x5_reduce)
inception_3b_pool = MaxPooling2D(
pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_3b/pool')(inception_3a_output)
inception_3b_pool_proj = Convolution2D(
64,
1,
1,
border_mode='same',
activation='relu',
name='inception_3b/pool_proj',
W_regularizer=l2(0.0002))(inception_3b_pool)
inception_3b_output = merge([
inception_3b_1x1, inception_3b_3x3, inception_3b_5x5,
inception_3b_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_3b/output')
inception_3b_output_zero_pad = ZeroPadding2D(
padding=(1, 1))(inception_3b_output)
pool3_helper = PoolHelper()(inception_3b_output_zero_pad)
pool3_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
border_mode='valid',
name='pool3/3x3_s2')(pool3_helper)
inception_4a_1x1 = Convolution2D(192,
1,
1,
border_mode='same',
activation='relu',
name='inception_4a/1x1',
W_regularizer=l2(0.0002))(pool3_3x3_s2)
inception_4a_3x3_reduce = Convolution2D(
96,
1,
1,
border_mode='same',
activation='relu',
name='inception_4a/3x3_reduce',
W_regularizer=l2(0.0002))(pool3_3x3_s2)
inception_4a_3x3 = Convolution2D(
208,
3,
3,
border_mode='same',
activation='relu',
name='inception_4a/3x3',
W_regularizer=l2(0.0002))(inception_4a_3x3_reduce)
inception_4a_5x5_reduce = Convolution2D(
16,
1,
1,
border_mode='same',
activation='relu',
name='inception_4a/5x5_reduce',
W_regularizer=l2(0.0002))(pool3_3x3_s2)
inception_4a_5x5 = Convolution2D(
48,
5,
5,
border_mode='same',
activation='relu',
name='inception_4a/5x5',
W_regularizer=l2(0.0002))(inception_4a_5x5_reduce)
inception_4a_pool = MaxPooling2D(pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_4a/pool')(pool3_3x3_s2)
inception_4a_pool_proj = Convolution2D(
64,
1,
1,
border_mode='same',
activation='relu',
name='inception_4a/pool_proj',
W_regularizer=l2(0.0002))(inception_4a_pool)
inception_4a_output = merge([
inception_4a_1x1, inception_4a_3x3, inception_4a_5x5,
inception_4a_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_4a/output')
loss1_ave_pool = AveragePooling2D(
pool_size=(5, 5), strides=(3, 3),
name='loss1/ave_pool')(inception_4a_output)
loss1_conv = Convolution2D(128,
1,
1,
border_mode='same',
activation='relu',
name='loss1/conv',
W_regularizer=l2(0.0002))(loss1_ave_pool)
loss1_flat = Flatten()(loss1_conv)
loss1_fc = Dense(1024,
activation='relu',
name='loss1/fc',
W_regularizer=l2(0.0002))(loss1_flat)
loss1_drop_fc = Dropout(0.7)(loss1_fc)
loss1_classifier = Dense(1000,
name='loss1/classifier',
W_regularizer=l2(0.0002))(loss1_drop_fc)
loss1_classifier_act = Activation('softmax')(loss1_classifier)
inception_4b_1x1 = Convolution2D(
160,
1,
1,
border_mode='same',
activation='relu',
name='inception_4b/1x1',
W_regularizer=l2(0.0002))(inception_4a_output)
inception_4b_3x3_reduce = Convolution2D(
112,
1,
1,
border_mode='same',
activation='relu',
name='inception_4b/3x3_reduce',
W_regularizer=l2(0.0002))(inception_4a_output)
inception_4b_3x3 = Convolution2D(
224,
3,
3,
border_mode='same',
activation='relu',
name='inception_4b/3x3',
W_regularizer=l2(0.0002))(inception_4b_3x3_reduce)
inception_4b_5x5_reduce = Convolution2D(
24,
1,
1,
border_mode='same',
activation='relu',
name='inception_4b/5x5_reduce',
W_regularizer=l2(0.0002))(inception_4a_output)
inception_4b_5x5 = Convolution2D(
64,
5,
5,
border_mode='same',
activation='relu',
name='inception_4b/5x5',
W_regularizer=l2(0.0002))(inception_4b_5x5_reduce)
inception_4b_pool = MaxPooling2D(
pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_4b/pool')(inception_4a_output)
inception_4b_pool_proj = Convolution2D(
64,
1,
1,
border_mode='same',
activation='relu',
name='inception_4b/pool_proj',
W_regularizer=l2(0.0002))(inception_4b_pool)
inception_4b_output = merge([
inception_4b_1x1, inception_4b_3x3, inception_4b_5x5,
inception_4b_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_4b_output')
inception_4c_1x1 = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_4c/1x1',
W_regularizer=l2(0.0002))(inception_4b_output)
inception_4c_3x3_reduce = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_4c/3x3_reduce',
W_regularizer=l2(0.0002))(inception_4b_output)
inception_4c_3x3 = Convolution2D(
256,
3,
3,
border_mode='same',
activation='relu',
name='inception_4c/3x3',
W_regularizer=l2(0.0002))(inception_4c_3x3_reduce)
inception_4c_5x5_reduce = Convolution2D(
24,
1,
1,
border_mode='same',
activation='relu',
name='inception_4c/5x5_reduce',
W_regularizer=l2(0.0002))(inception_4b_output)
inception_4c_5x5 = Convolution2D(
64,
5,
5,
border_mode='same',
activation='relu',
name='inception_4c/5x5',
W_regularizer=l2(0.0002))(inception_4c_5x5_reduce)
inception_4c_pool = MaxPooling2D(
pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_4c/pool')(inception_4b_output)
inception_4c_pool_proj = Convolution2D(
64,
1,
1,
border_mode='same',
activation='relu',
name='inception_4c/pool_proj',
W_regularizer=l2(0.0002))(inception_4c_pool)
inception_4c_output = merge([
inception_4c_1x1, inception_4c_3x3, inception_4c_5x5,
inception_4c_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_4c/output')
inception_4d_1x1 = Convolution2D(
112,
1,
1,
border_mode='same',
activation='relu',
name='inception_4d/1x1',
W_regularizer=l2(0.0002))(inception_4c_output)
inception_4d_3x3_reduce = Convolution2D(
144,
1,
1,
border_mode='same',
activation='relu',
name='inception_4d/3x3_reduce',
W_regularizer=l2(0.0002))(inception_4c_output)
inception_4d_3x3 = Convolution2D(
288,
3,
3,
border_mode='same',
activation='relu',
name='inception_4d/3x3',
W_regularizer=l2(0.0002))(inception_4d_3x3_reduce)
inception_4d_5x5_reduce = Convolution2D(
32,
1,
1,
border_mode='same',
activation='relu',
name='inception_4d/5x5_reduce',
W_regularizer=l2(0.0002))(inception_4c_output)
inception_4d_5x5 = Convolution2D(
64,
5,
5,
border_mode='same',
activation='relu',
name='inception_4d/5x5',
W_regularizer=l2(0.0002))(inception_4d_5x5_reduce)
inception_4d_pool = MaxPooling2D(
pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_4d/pool')(inception_4c_output)
inception_4d_pool_proj = Convolution2D(
64,
1,
1,
border_mode='same',
activation='relu',
name='inception_4d/pool_proj',
W_regularizer=l2(0.0002))(inception_4d_pool)
inception_4d_output = merge([
inception_4d_1x1, inception_4d_3x3, inception_4d_5x5,
inception_4d_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_4d/output')
loss2_ave_pool = AveragePooling2D(
pool_size=(5, 5), strides=(3, 3),
name='loss2/ave_pool')(inception_4d_output)
loss2_conv = Convolution2D(128,
1,
1,
border_mode='same',
activation='relu',
name='loss2/conv',
W_regularizer=l2(0.0002))(loss2_ave_pool)
loss2_flat = Flatten()(loss2_conv)
loss2_fc = Dense(1024,
activation='relu',
name='loss2/fc',
W_regularizer=l2(0.0002))(loss2_flat)
loss2_drop_fc = Dropout(0.7)(loss2_fc)
loss2_classifier = Dense(1000,
name='loss2/classifier',
W_regularizer=l2(0.0002))(loss2_drop_fc)
loss2_classifier_act = Activation('softmax')(loss2_classifier)
inception_4e_1x1 = Convolution2D(
256,
1,
1,
border_mode='same',
activation='relu',
name='inception_4e/1x1',
W_regularizer=l2(0.0002))(inception_4d_output)
inception_4e_3x3_reduce = Convolution2D(
160,
1,
1,
border_mode='same',
activation='relu',
name='inception_4e/3x3_reduce',
W_regularizer=l2(0.0002))(inception_4d_output)
inception_4e_3x3 = Convolution2D(
320,
3,
3,
border_mode='same',
activation='relu',
name='inception_4e/3x3',
W_regularizer=l2(0.0002))(inception_4e_3x3_reduce)
inception_4e_5x5_reduce = Convolution2D(
32,
1,
1,
border_mode='same',
activation='relu',
name='inception_4e/5x5_reduce',
W_regularizer=l2(0.0002))(inception_4d_output)
inception_4e_5x5 = Convolution2D(
128,
5,
5,
border_mode='same',
activation='relu',
name='inception_4e/5x5',
W_regularizer=l2(0.0002))(inception_4e_5x5_reduce)
inception_4e_pool = MaxPooling2D(
pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_4e/pool')(inception_4d_output)
inception_4e_pool_proj = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_4e/pool_proj',
W_regularizer=l2(0.0002))(inception_4e_pool)
inception_4e_output = merge([
inception_4e_1x1, inception_4e_3x3, inception_4e_5x5,
inception_4e_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_4e/output')
inception_4e_output_zero_pad = ZeroPadding2D(
padding=(1, 1))(inception_4e_output)
pool4_helper = PoolHelper()(inception_4e_output_zero_pad)
pool4_3x3_s2 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
border_mode='valid',
name='pool4/3x3_s2')(pool4_helper)
inception_5a_1x1 = Convolution2D(256,
1,
1,
border_mode='same',
activation='relu',
name='inception_5a/1x1',
W_regularizer=l2(0.0002))(pool4_3x3_s2)
inception_5a_3x3_reduce = Convolution2D(
160,
1,
1,
border_mode='same',
activation='relu',
name='inception_5a/3x3_reduce',
W_regularizer=l2(0.0002))(pool4_3x3_s2)
inception_5a_3x3 = Convolution2D(
320,
3,
3,
border_mode='same',
activation='relu',
name='inception_5a/3x3',
W_regularizer=l2(0.0002))(inception_5a_3x3_reduce)
inception_5a_5x5_reduce = Convolution2D(
32,
1,
1,
border_mode='same',
activation='relu',
name='inception_5a/5x5_reduce',
W_regularizer=l2(0.0002))(pool4_3x3_s2)
inception_5a_5x5 = Convolution2D(
128,
5,
5,
border_mode='same',
activation='relu',
name='inception_5a/5x5',
W_regularizer=l2(0.0002))(inception_5a_5x5_reduce)
inception_5a_pool = MaxPooling2D(pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_5a/pool')(pool4_3x3_s2)
inception_5a_pool_proj = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_5a/pool_proj',
W_regularizer=l2(0.0002))(inception_5a_pool)
inception_5a_output = merge([
inception_5a_1x1, inception_5a_3x3, inception_5a_5x5,
inception_5a_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_5a/output')
inception_5b_1x1 = Convolution2D(
384,
1,
1,
border_mode='same',
activation='relu',
name='inception_5b/1x1',
W_regularizer=l2(0.0002))(inception_5a_output)
inception_5b_3x3_reduce = Convolution2D(
192,
1,
1,
border_mode='same',
activation='relu',
name='inception_5b/3x3_reduce',
W_regularizer=l2(0.0002))(inception_5a_output)
inception_5b_3x3 = Convolution2D(
384,
3,
3,
border_mode='same',
activation='relu',
name='inception_5b/3x3',
W_regularizer=l2(0.0002))(inception_5b_3x3_reduce)
inception_5b_5x5_reduce = Convolution2D(
48,
1,
1,
border_mode='same',
activation='relu',
name='inception_5b/5x5_reduce',
W_regularizer=l2(0.0002))(inception_5a_output)
inception_5b_5x5 = Convolution2D(
128,
5,
5,
border_mode='same',
activation='relu',
name='inception_5b/5x5',
W_regularizer=l2(0.0002))(inception_5b_5x5_reduce)
inception_5b_pool = MaxPooling2D(
pool_size=(3, 3),
strides=(1, 1),
border_mode='same',
name='inception_5b/pool')(inception_5a_output)
inception_5b_pool_proj = Convolution2D(
128,
1,
1,
border_mode='same',
activation='relu',
name='inception_5b/pool_proj',
W_regularizer=l2(0.0002))(inception_5b_pool)
inception_5b_output = merge([
inception_5b_1x1, inception_5b_3x3, inception_5b_5x5,
inception_5b_pool_proj
],
mode='concat',
concat_axis=1,
name='inception_5b/output')
pool5_7x7_s1 = AveragePooling2D(pool_size=(7, 7),
strides=(1, 1),
name='pool5/7x7_s2')(inception_5b_output)
loss3_flat = Flatten()(pool5_7x7_s1)
pool5_drop_7x7_s1 = Dropout(0.4)(loss3_flat)
loss3_classifier = Dense(1000,
name='loss3/classifier',
W_regularizer=l2(0.0002))(pool5_drop_7x7_s1)
loss3_classifier_act = Activation('softmax', name='prob')(loss3_classifier)
# Create model
model = Model(input=input,
output=[
loss1_classifier_act, loss2_classifier_act,
loss3_classifier_act
])
# Load ImageNet pre-trained data
model.load_weights(model_path + 'googlenet_weights.h5')
# Truncate and replace softmax layer for transfer learning
# Cannot use model.layers.pop() since model is not of Sequential() type
# The method below works since pre-trained weights are stored in layers but not in the model
loss1_classifier_statefarm = Dense(num_classes,
name='loss1/classifier',
W_regularizer=l2(0.0002))(loss1_drop_fc)
loss1_classifier_act_statefarm = Activation('softmax')(
loss1_classifier_statefarm)
# Create another model with our customized softmax
model = Model(input=input, output=[loss1_classifier_act_statefarm])
# Learning rate is changed to 0.001
sgd = SGD(lr=1e-3, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd,
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def main(layers1=[200],
layers2=[200],
leaky_alpha=0.10,
ENV_NAME='EnvPong',
show=False,
wall_reward=-0.1,
touch_reward=0.3,
n_steps=80000,
n_alternances=10,
L_R=0.0001,
only_test=False,
opp_aware=[1, 1],
myopie=[0.00, 0.00],
ball_speed=1.0,
weights1_name='',
weights2_name=''):
ENV_NAME = ENV_NAME
conf_name = "{}_layers1={}__layers2={}__leaky={}__lr={}__opp={}__myopia={}__speed={}".format(
ENV_NAME, layers1, layers2, leaky_alpha, L_R, opp_aware, myopie,
ball_speed)
#gym.undo_logger_setup()
# Get the environment and extract the number of actions.
if ENV_NAME == 'Env2D':
env = Game2D(2.)
elif ENV_NAME == 'Env2DSoloSpin':
env = Game2DSolo(2., spinRacket=True)
elif ENV_NAME == 'Env3DSolo':
env = Game3DSolo(2., 9.8, 0.5, 7., 3.)
elif ENV_NAME == 'EnvPong':
env = Pong(PongPlayer(None, opp_aware=(opp_aware[0] == 1)),
PongPlayer(None, opp_aware=(opp_aware[1] == 1)))
np.random.seed(123)
#env.seed(123)
assert len(env.action_space.shape) == 1
nb_actions = env.action_space.shape[0]
# Next, we build a very simple model.
actor = Sequential()
actor.add(Flatten(input_shape=(1, ) + env.observation_space_1.shape))
#actor.add(keras.layers.normalization.BatchNormalization())
for size in layers1:
actor.add(
Dense(size,
kernel_initializer=RandomUniform(minval=-0.005,
maxval=0.005,
seed=None)))
#actor.add(keras.layers.core.Dropout(0.2))
actor.add(LeakyReLU(leaky_alpha))
#actor.add(keras.layers.normalization.BatchNormalization())
actor.add(
Dense(nb_actions,
kernel_initializer=RandomUniform(minval=-0.005,
maxval=0.005,
seed=None),
bias_regularizer=regularizers.l2(0.01)))
#actor.add(keras.layers.core.Dropout(0.2))
actor.add(Activation('linear'))
print(actor.summary())
action_input = Input(shape=(nb_actions, ), name='action_input')
observation_input = Input(shape=(1, ) + env.observation_space_1.shape,
name='observation_input')
flattened_observation = Flatten()(observation_input)
x = merge([action_input, flattened_observation], mode='concat')
#x = keras.layers.normalization.BatchNormalization()(x)
for size in layers1:
x = Dense(size)(x)
#x = keras.layers.core.Dropout(0.2)(x)
x = LeakyReLU(alpha=leaky_alpha)(x)
#x = keras.layers.normalization.BatchNormalization()(x)
x = Dense(1)(x)
x = Activation('linear')(x)
critic = Model(input=[action_input, observation_input], output=x)
print(critic.summary())
actor2 = Sequential()
actor2.add(Flatten(input_shape=(1, ) + env.observation_space_2.shape))
#actor2.add(keras.layers.normalization.BatchNormalization())
for size in layers2:
actor2.add(
Dense(size,
kernel_initializer=RandomUniform(minval=-0.005,
maxval=0.005,
seed=None)))
#actor2.add(keras.layers.core.Dropout(0.2))
actor2.add(LeakyReLU(alpha=leaky_alpha))
actor2.add(
Dense(nb_actions,
kernel_initializer=RandomUniform(minval=-0.005,
maxval=0.005,
seed=None),
bias_regularizer=regularizers.l2(0.01)))
#actor2.add(keras.layers.core.Dropout(0.2))
actor2.add(Activation('linear'))
print(actor2.summary())
action_input2 = Input(shape=(nb_actions, ), name='action_input')
observation_input2 = Input(shape=(1, ) + env.observation_space_2.shape,
name='observation_input')
flattened_observation2 = Flatten()(observation_input2)
x2 = merge([action_input2, flattened_observation2], mode='concat')
#x2 = keras.layers.normalization.BatchNormalization()(x2)
for size in layers2:
x2 = Dense(size)(x2)
#x2 = keras.layers.core.Dropout(0.2)(x2)
x2 = LeakyReLU(alpha=leaky_alpha)(x2)
x2 = Dense(1)(x2)
x2 = Activation('linear')(x2)
critic2 = Model(input=[action_input2, observation_input2], output=x2)
print(critic2.summary())
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory1 = SequentialMemory(limit=50000, window_length=1)
if opp_aware[0] != opp_aware[1]:
memory2 = SequentialMemory(limit=50000, window_length=1)
else:
memory2 = memory1
random_process1 = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=.1,
mu=0.,
sigma=.15,
sigma_min=0.,
n_steps_annealing=n_steps /
4) # Explores less at the end ?
random_process2 = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=.1,
mu=0.,
sigma=.15,
sigma_min=0.,
n_steps_annealing=4 * n_steps)
agent1 = DDPGAgent(nb_actions=nb_actions,
actor=actor,
critic=critic,
critic_action_input=action_input,
memory=memory1,
nb_steps_warmup_critic=5000,
nb_steps_warmup_actor=5000,
random_process=random_process1,
gamma=.99,
target_model_update=1e-3,
batch_size=100)
agent2 = DDPGAgent(nb_actions=nb_actions,
actor=actor2,
critic=critic2,
critic_action_input=action_input2,
memory=memory2,
nb_steps_warmup_critic=5000,
nb_steps_warmup_actor=5000,
random_process=random_process2,
gamma=.99,
target_model_update=1e-3,
batch_size=100)
#agent.compile(Adam(lr=L_R, clipnorm=1., clipvalue=0.5), metrics=['mae'])
agent1.compile(Adam(lr=L_R, clipnorm=1.), metrics=['mae'])
agent2.compile(Adam(lr=L_R, clipnorm=1.), metrics=['mae'])
player1 = PongPlayer(agent1,
myopie=myopie[0],
opp_aware=(opp_aware[0] == 1))
player2 = PongPlayer(agent2,
myopie=myopie[1],
opp_aware=(opp_aware[1] == 1))
# Grid -4
# Add -1 when lost
# CEM method
directory_log = "logs/ddpg/{}".format(conf_name)
directory_weights = "weights/ddpg/{}".format(conf_name)
if not os.path.exists(directory_log):
os.makedirs(directory_log)
if not os.path.exists(directory_weights):
os.makedirs(directory_weights)
if only_test:
'''if weights1_name =='':
weights1_name = "{}/player1_final".format(directory_weights)
if weights2_name == '':
weights2_name = "{}/player2_final".format(directory_weights)
#if os.path.isfile(weights1_name) and os.path.isfile(weights2_name):
agent1.load_weights(weights1_name)
agent2.load_weights(weights2_name)'''
agent1.load_weights("{}/player1_{}".format(directory_weights, "final"))
agent2.load_weights("{}/player1_{}".format(directory_weights, "final"))
env = makeEnv(player1, player2, ENV_NAME, ball_speed=ball_speed)
for i in range(10):
playPong(env)
confrontPlayers(env)
plotStrategy(env)
else:
for i in range(n_alternances):
print "Alternance n {} \n".format(i)
def learning_rate_schedule(epoch):
return L_R
if ENV_NAME == 'Env2D':
env = Game2D(agent2,
wall_reward=wall_reward,
touch_reward=touch_reward)
elif ENV_NAME == 'EnvPong':
env = Pong(player1,
player2,
wall_reward=wall_reward,
touch_reward=touch_reward,
ball_speed=ball_speed)
agent1.fit(env,
nb_steps=n_steps,
visualize=False,
verbose=1,
until_score=True,
score_to_reach=0.5,
last_episodes=500,
nb_max_episode_steps=None,
callbacks=[
FileLogger("{}/player1_{}.h5f".format(
directory_log, i)),
keras.callbacks.LearningRateScheduler(
learning_rate_schedule)
])
agent1.test(env,
nb_episodes=100,
visualize=False,
nb_max_episode_steps=500,
verbose=1)
agent1.save_weights("{}/player1_{}".format(directory_weights, i),
overwrite=True)
agent1.memory = SequentialMemory(limit=500000, window_length=1)
wall_reward = wall_reward * 0.8
touch_reward = touch_reward * 0.8
agent2.load_weights("{}/player1_{}".format(directory_weights, i))
print "Fin de {}".format(conf_name)
env = Pong(player1,
player2,
wall_reward=wall_reward,
touch_reward=touch_reward,
ball_speed=ball_speed)
#agent1.fit(env, nb_steps=150000, visualize=False, verbose=2, nb_max_episode_steps=None,callbacks=[FileLogger("logs/ddpg/{}_weights_steps_leaky_reg_bias_drop_lr{}.h5f".format(ENV_NAME,L_R), interval=100)])
agent1.save_weights("{}/player1_final".format(directory_weights),
overwrite=True)
agent2.save_weights("{}/player2_final".format(directory_weights),
overwrite=True)
agent1.test(env,
nb_episodes=15,
visualize=False,
nb_max_episode_steps=500,
verbose=2)
if show == True:
if ENV_NAME == 'Env2D':
for i in range(10):
play2D(player1=agent1, player2=agent1)
elif ENV_NAME == 'EnvPong':
for i in range(10):
playPong(left=agent1, right=agent2)
def conv_block(input_tensor,
kernel_size,
filters,
stage,
block,
strides=(2, 2)):
'''conv_block is the block that has a conv layer at shortcut
# Arguments
input_tensor: input tensor
kernel_size: defualt 3, the kernel size of middle conv layer at main path
filters: list of integers, the nb_filters of 3 conv layer at main path
stage: integer, current stage label, used for generating layer names
block: 'a','b'..., current block label, used for generating layer names
Note that from stage 3, the first conv layer at main path is with subsample=(2,2)
And the shortcut should have subsample=(2,2) as well
'''
eps = 1.1e-5
nb_filter1, nb_filter2, nb_filter3 = filters
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
scale_name_base = 'scale' + str(stage) + block + '_branch'
x = Convolution2D(nb_filter1,
1,
1,
subsample=strides,
name=conv_name_base + '2a',
bias=False)(input_tensor)
x = BatchNormalization(epsilon=eps, axis=bn_axis,
name=bn_name_base + '2a')(x)
x = Scale(axis=bn_axis, name=scale_name_base + '2a')(x)
x = Activation('relu', name=conv_name_base + '2a_relu')(x)
x = ZeroPadding2D((1, 1), name=conv_name_base + '2b_zeropadding')(x)
x = Convolution2D(nb_filter2,
kernel_size,
kernel_size,
name=conv_name_base + '2b',
bias=False)(x)
x = BatchNormalization(epsilon=eps, axis=bn_axis,
name=bn_name_base + '2b')(x)
x = Scale(axis=bn_axis, name=scale_name_base + '2b')(x)
x = Activation('relu', name=conv_name_base + '2b_relu')(x)
x = Convolution2D(nb_filter3, 1, 1, name=conv_name_base + '2c',
bias=False)(x)
x = BatchNormalization(epsilon=eps, axis=bn_axis,
name=bn_name_base + '2c')(x)
x = Scale(axis=bn_axis, name=scale_name_base + '2c')(x)
shortcut = Convolution2D(nb_filter3,
1,
1,
subsample=strides,
name=conv_name_base + '1',
bias=False)(input_tensor)
shortcut = BatchNormalization(epsilon=eps,
axis=bn_axis,
name=bn_name_base + '1')(shortcut)
shortcut = Scale(axis=bn_axis, name=scale_name_base + '1')(shortcut)
x = merge([x, shortcut], mode='sum', name='res' + str(stage) + block)
x = Activation('relu', name='res' + str(stage) + block + '_relu')(x)
return x
def vaeBuild(self):
in_dim = self.in_dim
expr_in = Input(shape=(self.in_dim, ))
h0 = NoiseLayer(0.8)(expr_in)
#h0 = Dropout(0.2)(expr_in)
## Encoder layers
h1 = Dense(units=512, name='encoder_1')(h0)
#h1_relu = Activation('relu')(h1)
#h1_drop = Dropout(0.3)(h1)
h2 = Dense(units=128, name='encoder_2')(h1)
h2_relu = Activation('relu')(h2)
#h2_relu = Dropout(0.3)(h2_relu)
h3 = Dense(units=32, name='encoder_3')(h2_relu)
h3_relu = Activation('relu')(h3)
#h3_relu = Dropout(0.3)(h3_relu)
z_mean = Dense(units=2, name='z_mean')(h3_relu)
z_log_var = Dense(units=2, name='z_log_var')(h3_relu)
#z_log_var = Lambda( lambda x:K.log(x) )(z_log_var)
drop_ratio = Dense(units=1, name='drop_ratio',
activation='sigmoid')(h3_relu)
drop_ratio = RepeatVector(self.in_dim)(drop_ratio)
drop_ratio = Reshape(target_shape=(self.in_dim, ))(drop_ratio)
## sampling new samples
z = Lambda(sampling, output_shape=(2, ))([z_mean, z_log_var])
## Decoder layers
#decoder_h1 = Dense( units=32,name='decoder_1' )(z)
#decoder_h1_relu = Activation('relu')(decoder_h1)
decoder_h2 = Dense(units=128, name='decoder_2')(z)
decoder_h2_relu = Activation('relu')(decoder_h2)
decoder_h2_relu = Dropout(0.3)(decoder_h2_relu)
decoder_h3 = Dense(units=512, name='decoder_3')(decoder_h2_relu)
decoder_h3_relu = Activation('relu')(decoder_h3)
decoder_h3_relu = Dropout(0.3)(decoder_h3_relu)
expr_x_tanh = Dense(units=self.in_dim,
activation='tanh')(decoder_h3_relu)
#expr_x = Lambda( lambda x:x*0.6 )(expr_x)
expr_x = Activation('relu')(expr_x_tanh)
expr_x_drop = Lambda(lambda x: -x**2)(expr_x)
expr_x_drop_log = merge(
[drop_ratio, expr_x_drop],
mode='mul') ### log p_drop = log(exp(-\lambda x^2))
expr_x_drop_p = Lambda(lambda x: K.exp(x))(expr_x_drop_log)
expr_x_nondrop_p = Lambda(lambda x: 1 - x)(expr_x_drop_p)
expr_x_nondrop_log = Lambda(lambda x: K.log(x + 1e-20))(
expr_x_nondrop_p)
expr_x_drop_log = Reshape(target_shape=(self.in_dim,
1))(expr_x_drop_log)
expr_x_nondrop_log = Reshape(target_shape=(self.in_dim,
1))(expr_x_nondrop_log)
logits = merge([expr_x_drop_log, expr_x_nondrop_log],
mode='concat',
concat_axis=-1)
samples = Lambda(gumbel_softmax, output_shape=(
self.in_dim,
2,
))(logits)
samples = Lambda(lambda x: x[:, :, 1])(samples)
samples = Reshape(target_shape=(self.in_dim, ))(samples)
#print(samples.shape)
out = merge([expr_x, samples], mode='mul')
class VariationalLayer(Layer):
def __init__(self, **kwargs):
self.is_placeholder = True
super(VariationalLayer, self).__init__(**kwargs)
def vae_loss(self, x, x_decoded_mean):
xent_loss = in_dim * metrics.binary_crossentropy(
x, x_decoded_mean)
kl_loss = -0.5 * K.sum(
1 + z_log_var - K.square(z_mean) - K.exp(z_log_var),
axis=-1)
return K.mean(xent_loss + kl_loss)
def call(self, inputs):
x = inputs[0]
x_decoded_mean = inputs[1]
loss = self.vae_loss(x, x_decoded_mean)
self.add_loss(loss, inputs=inputs)
# We won't actually use the output.
return x
y = VariationalLayer()([expr_in, out])
vae = Model(inputs=expr_in, outputs=y)
opt = RMSprop(lr=0.0001)
vae.compile(optimizer=opt, loss=None)
ae = Model(inputs=expr_in,
outputs=[h1, h2, h3, z_mean, decoder_h2, decoder_h3])
aux = Model(inputs=expr_in,
outputs=[expr_x_tanh, expr_x, samples, out, drop_ratio])
self.vae = vae
self.ae = ae
self.aux = aux
def get_unet(self):
inputs = Input((self.img_rows, self.img_cols, 1))
conv1 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(inputs)
conv1 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv1)
pool1 = MaxPooling2D(pool_size=(2, 2))(
conv1) #pool1=MaxPolong2D()(b)是指张量b作为输入，其他与此类同
conv2 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool1)
conv2 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv2)
pool2 = MaxPooling2D(pool_size=(2, 2))(conv2)
conv3 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool2)
conv3 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv3)
pool3 = MaxPooling2D(pool_size=(2, 2))(conv3)
conv4 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool3)
conv4 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv4)
drop4 = Dropout(0.5)(conv4)
pool4 = MaxPooling2D(pool_size=(2, 2))(drop4)
conv5 = Conv2D(1024,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(pool4)
conv5 = Conv2D(1024,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv5)
drop5 = Dropout(0.5)(conv5)
up6 = Conv2D(512,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(drop5))
merge6 = merge([drop4, up6], mode='concat', concat_axis=3)
conv6 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge6)
conv6 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv6)
up7 = Conv2D(256,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv6))
merge7 = merge([conv3, up7], mode='concat', concat_axis=3)
conv7 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge7)
conv7 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv7)
up8 = Conv2D(128,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv7))
merge8 = merge([conv2, up8], mode='concat', concat_axis=3)
conv8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge8)
conv8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv8)
up9 = Conv2D(64,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal')(
UpSampling2D(size=(2, 2))(conv8))
merge9 = merge([conv1, up9], mode='concat', concat_axis=3)
conv9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(merge9)
conv9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv9)
conv9 = Conv2D(2,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(conv9)
conv10 = Conv2D(1, 1, activation='sigmoid')(conv9)
model = Model(input=inputs, output=conv10)
model.compile(optimizer=Adam(lr=1e-4),
loss='binary_crossentropy',
metrics=['accuracy'])
print('model compile')
return model
def AlexNet(weights_path=None, heatmap=False):
if heatmap:
inputs = Input(shape=(3, None, None))
else:
inputs = Input(shape=(3, 227, 227))
conv_1 = Convolution2D(96, (11, 11),
strides=(4, 4),
activation='relu',
name='conv_1')(inputs)
conv_2 = MaxPooling2D((3, 3), strides=(2, 2))(conv_1)
conv_2 = crosschannelnormalization(name='convpool_1')(conv_2)
conv_2 = ZeroPadding2D((2, 2))(conv_2)
conv_2 = merge([
Convolution2D(
128, (5, 5), activation='relu', name='conv_2_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_2))
for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_2')
conv_3 = MaxPooling2D((3, 3), strides=(2, 2))(conv_2)
conv_3 = crosschannelnormalization()(conv_3)
conv_3 = ZeroPadding2D((1, 1))(conv_3)
conv_3 = Convolution2D(384, (3, 3), activation='relu',
name='conv_3')(conv_3)
conv_4 = ZeroPadding2D((1, 1))(conv_3)
conv_4 = merge([
Convolution2D(
192, (3, 3), activation='relu', name='conv_4_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_4))
for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_4')
conv_5 = ZeroPadding2D((1, 1))(conv_4)
conv_5 = merge([
Convolution2D(
128, (3, 3), activation='relu', name='conv_5_' + str(i + 1))(
splittensor(ratio_split=2, id_split=i)(conv_5))
for i in range(2)
],
mode='concat',
concat_axis=1,
name='conv_5')
dense_1 = MaxPooling2D((3, 3), strides=(2, 2), name='convpool_5')(conv_5)
if heatmap:
dense_1 = Convolution2D(4096, (6, 6),
activation='relu',
name='dense_1')(dense_1)
dense_2 = Convolution2D(4096, (1, 1),
activation='relu',
name='dense_2')(dense_1)
dense_3 = Convolution2D(1000, (1, 1), name='dense_3')(dense_2)
prediction = Softmax4D(axis=1, name='softmax')(dense_3)
else:
dense_1 = Flatten(name='flatten')(dense_1)
dense_1 = Dense(4096, activation='relu', name='dense_1')(dense_1)
dense_2 = Dropout(0.5)(dense_1)
dense_2 = Dense(4096, activation='relu', name='dense_2')(dense_2)
dense_3 = Dropout(0.5)(dense_2)
dense_3 = Dense(1000, name='dense_3')(dense_3)
prediction = Activation('softmax', name='softmax')(dense_3)
model = Model(input=inputs, output=prediction)
if weights_path:
model.load_weights(weights_path)
return model
lstm_input_r13=tf.concat(1,[lstm_input_r12,tf.expand_dims(atten_output_r15, 1)]) lstm_input_r14=tf.concat(1,[lstm_input_r13,tf.expand_dims(atten_output_r16, 1)]) lstm_input_r15=tf.concat(1,[lstm_input_r13,tf.expand_dims(atten_output_r17, 1)]) lstm_input_r16=tf.concat(1,[lstm_input_r13,tf.expand_dims(atten_output_r18, 1)]) var_c0_c= LSTM(output_dim=256,activation='relu',return_sequences=False,W_regularizer=l2(0.01), inner_activation='sigmoid')(lstm_input_c4) var_r0_r= LSTM(output_dim=256,activation='relu',return_sequences=False,W_regularizer=l2(0.01), inner_activation='sigmoid')(lstm_input_r16) var_r1=Dense(100,bias=False,activation='tanh')(var_r0_r) var_c1=Dense(100,bias=False,activation='tanh')(var_c0_c) #dropout layer var_c= Dropout(0.5)(var_c1) var_r= Dropout(0.5)(var_r1) var = merge([var_c,var_r], mode= 'concat') predictions= Dense(2,bias=False,activation='softmax')(var) labels = tf.placeholder(tf.float32, shape=(None,2)) loss = tf.reduce_mean(binary_crossentropy(labels, predictions)) train_w_step = tf.train.AdagradOptimizer(learning_rate=0.1).minimize(loss) acc_value = accuracy(labels, predictions)
def get_2k_twopath_twoconv(img_w, img_h):
input_img = Input(shape=(img_w, img_h, 1))
normal = BatchNormalization(input_shape=(img_w, img_h, 1))(input_img)
conv1 = Convolution2D(184,
4,
4,
init='glorot_normal',
input_shape=(img_w, img_h, 1),
border_mode='valid',
activation='relu')(normal)
max1 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2),
border_mode='same')(conv1)
drop1 = Dropout(0.25)(max1)
conv2 = Convolution2D(60,
4,
4,
init='glorot_normal',
activation='relu',
border_mode='valid')(drop1)
max2 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2),
border_mode='same')(conv2)
drop2 = Dropout(0.25)(max2)
conv1_bigk = Convolution2D(184,
8,
8,
init='glorot_normal',
input_shape=(img_w, img_h, 1),
border_mode='valid',
activation='relu')(normal)
max1_bigk = MaxPooling2D(pool_size=(2, 2),
strides=(2, 2),
border_mode='same')(conv1_bigk)
drop1_bigk = Dropout(0.25)(max1_bigk)
merge_all = merge([drop2, drop1_bigk], mode='concat', concat_axis=-1)
conv3 = Convolution2D(40,
4,
4,
init='glorot_normal',
activation='relu',
border_mode='same')(merge_all)
max3 = MaxPooling2D(pool_size=(2, 2), strides=(2, 2),
border_mode='same')(conv3)
drop3 = Dropout(0.25)(max3)
flat = Flatten()(drop3)
# model.add(BatchNormalization())
dense1 = Dense(500, activation='relu')(flat)
drop_final = Dropout(0.5)(dense1)
choice = Dense(3, init='glorot_normal', activation='softmax')(drop_final)
# model.add(Activation('softmax'))
model = Model(input=input_img, output=choice)
for i in range(len(model.layers)):
print(model.layers[i].output_shape, model.layers[i].output_shape)
plot(model, to_file='model_twopath.png')
return model
actor.add(Flatten(input_shape=(1, ) + env.observation_space.shape))
actor.add(Dense(32))
actor.add(Activation('relu'))
actor.add(Dense(32))
actor.add(Activation('relu'))
actor.add(Dense(32))
actor.add(Activation('relu'))
actor.add(Dense(nb_actions))
actor.add(Activation('sigmoid'))
print(actor.summary())
action_input = Input(shape=(nb_actions, ), name='action_input')
observation_input = Input(shape=(1, ) + env.observation_space.shape,
name='observation_input')
flattened_observation = Flatten()(observation_input)
x = merge([action_input, flattened_observation], mode='concat')
x = Dense(64)(x)
x = Activation('relu')(x)
x = Dense(64)(x)
x = Activation('relu')(x)
x = Dense(64)(x)
x = Activation('relu')(x)
x = Dense(1)(x)
x = Activation('linear')(x)
critic = Model(input=[action_input, observation_input], output=x)
print(critic.summary())
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory = SequentialMemory(limit=100000, window_length=1)
random_process = OrnsteinUhlenbeckProcess(theta=float(args.theta),
def _res_unit(self,
inputs,
nb_filter,
size=3,
stride=1,
atrous=1,
stage=1,
block=1):
name = '%02d-%02d/' % (stage, block)
id_name = '%sid_' % (name)
res_name = '%sres_' % (name)
# Residual branch
# 1x1 down-sample conv
x = kl.BatchNormalization(name=res_name + 'bn1')(inputs)
x = kl.Activation('relu', name=res_name + 'act1')(x)
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
x = kl.Conv1D(nb_filter[0],
1,
name=res_name + 'conv1',
subsample_length=stride,
init=self.init,
W_regularizer=w_reg)(x)
# LxL conv
x = kl.BatchNormalization(name=res_name + 'bn2')(x)
x = kl.Activation('relu', name=res_name + 'act2')(x)
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
x = kl.AtrousConv1D(nb_filter[1],
size,
atrous_rate=atrous,
name=res_name + 'conv2',
border_mode='same',
init=self.init,
W_regularizer=w_reg)(x)
# 1x1 up-sample conv
x = kl.BatchNormalization(name=res_name + 'bn3')(x)
x = kl.Activation('relu', name=res_name + 'act3')(x)
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
x = kl.Conv1D(nb_filter[2],
1,
name=res_name + 'conv3',
init=self.init,
W_regularizer=w_reg)(x)
# Identity branch
if nb_filter[-1] != inputs._keras_shape[-1] or stride > 1:
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
identity = kl.Conv1D(nb_filter[2],
1,
name=id_name + 'conv1',
subsample_length=stride,
init=self.init,
W_regularizer=w_reg)(inputs)
else:
identity = inputs
x = kl.merge([identity, x], name=name + 'merge', mode='sum')
return x
def _shortcut(input, residual):
return merge([input, residual], mode='sum')
rnn_kwargs = dict(output_dim=SENT_HIDDEN_SIZE, dropout_W=DP, dropout_U=DP)
premise = Input(shape=(MAX_LEN, ), dtype='int32')
hypothesis = Input(shape=(MAX_LEN, ), dtype='int32')
prem = embed(premise)
hypo = embed(hypothesis)
rnn_prem = RNN(return_sequences=False, **rnn_kwargs)
rnn_hypo = RNN(return_sequences=False, **rnn_kwargs)
prem = rnn_prem(prem)
prem = Dropout(DP)(prem)
hypo = rnn_hypo(hypo)
hypo = Dropout(DP)(hypo)
joint = merge([prem, hypo], mode='concat')
joint = Dense(output_dim=50, activation='tanh', W_regularizer=l2(0.01))(joint)
pred = Dense(len(LABELS), activation='softmax', W_regularizer=l2(0.01))(joint)
model = Model(input=[premise, hypothesis], output=pred)
model.compile(optimizer=OPTIMIZER,
loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
print('Training')
_, tmpfn = tempfile.mkstemp()
# Save the best model during validation and bail out of training early if we're not improving
callbacks = [
EarlyStopping(patience=PATIENCE),
def Xception(include_top=True, weights='imagenet',
input_tensor=None):
'''Instantiate the Xception architecture,
optionally loading weights pre-trained
on ImageNet. This model is available for TensorFlow only,
and can only be used with inputs following the TensorFlow
dimension ordering `(width, height, channels)`.
You should set `image_dim_ordering="tf"` in your Keras config
located at ~/.keras/keras.json.
Note that the default input image size for this model is 299x299.
# Arguments
include_top: whether to include the fully-connected
layer at the top of the network.
weights: one of `None` (random initialization)
or "imagenet" (pre-training on ImageNet).
input_tensor: optional Keras tensor (i.e. output of `layers.Input()`)
to use as image input for the model.
# Returns
A Keras model instance.
'''
if weights not in {'imagenet', None}:
raise ValueError('The `weights` argument should be either '
'`None` (random initialization) or `imagenet` '
'(pre-training on ImageNet).')
if K.backend() != 'tensorflow':
raise Exception('The Xception model is only available with '
'the TensorFlow backend.')
if K.image_dim_ordering() != 'tf':
warnings.warn('The Xception model is only available for the '
'input dimension ordering "tf" '
'(width, height, channels). '
'However your settings specify the default '
'dimension ordering "th" (channels, width, height). '
'You should set `image_dim_ordering="tf"` in your Keras '
'config located at ~/.keras/keras.json. '
'The model being returned right now will expect inputs '
'to follow the "tf" dimension ordering.')
K.set_image_dim_ordering('tf')
old_dim_ordering = 'th'
else:
old_dim_ordering = None
# Determine proper input shape
if include_top:
input_shape = (299, 299, 3)
else:
input_shape = (None, None, 3)
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
x = Conv2D(32, 3, 3, subsample=(2, 2), bias=False, name='block1_conv1')(img_input)
x = BatchNormalization(name='block1_conv1_bn')(x)
x = Activation('relu', name='block1_conv1_act')(x)
x = Conv2D(64, 3, 3, bias=False, name='block1_conv2')(x)
x = BatchNormalization(name='block1_conv2_bn')(x)
x = Activation('relu', name='block1_conv2_act')(x)
residual = Conv2D(128, 1, 1, subsample=(2, 2),
border_mode='same', bias=False)(x)
residual = BatchNormalization()(residual)
x = SeparableConv2D(128, 3, 3, border_mode='same', bias=False, name='block2_sepconv1')(x)
x = BatchNormalization(name='block2_sepconv1_bn')(x)
x = Activation('relu', name='block2_sepconv2_act')(x)
x = SeparableConv2D(128, 3, 3, border_mode='same', bias=False, name='block2_sepconv2')(x)
x = BatchNormalization(name='block2_sepconv2_bn')(x)
x = MaxPooling2D((3, 3), strides=(2, 2), border_mode='same', name='block2_pool')(x)
x = merge([x, residual], mode='sum')
residual = Conv2D(256, 1, 1, subsample=(2, 2),
border_mode='same', bias=False)(x)
residual = BatchNormalization()(residual)
x = Activation('relu', name='block3_sepconv1_act')(x)
x = SeparableConv2D(256, 3, 3, border_mode='same', bias=False, name='block3_sepconv1')(x)
x = BatchNormalization(name='block3_sepconv1_bn')(x)
x = Activation('relu', name='block3_sepconv2_act')(x)
x = SeparableConv2D(256, 3, 3, border_mode='same', bias=False, name='block3_sepconv2')(x)
x = BatchNormalization(name='block3_sepconv2_bn')(x)
x = MaxPooling2D((3, 3), strides=(2, 2), border_mode='same', name='block3_pool')(x)
x = merge([x, residual], mode='sum')
residual = Conv2D(728, 1, 1, subsample=(2, 2),
border_mode='same', bias=False)(x)
residual = BatchNormalization()(residual)
x = Activation('relu', name='block4_sepconv1_act')(x)
x = SeparableConv2D(728, 3, 3, border_mode='same', bias=False, name='block4_sepconv1')(x)
x = BatchNormalization(name='block4_sepconv1_bn')(x)
x = Activation('relu', name='block4_sepconv2_act')(x)
x = SeparableConv2D(728, 3, 3, border_mode='same', bias=False, name='block4_sepconv2')(x)
x = BatchNormalization(name='block4_sepconv2_bn')(x)
x = MaxPooling2D((3, 3), strides=(2, 2), border_mode='same', name='block4_pool')(x)
x = merge([x, residual], mode='sum')
for i in range(8):
residual = x
prefix = 'block' + str(i + 5)
x = Activation('relu', name=prefix + '_sepconv1_act')(x)
x = SeparableConv2D(728, 3, 3, border_mode='same', bias=False, name=prefix + '_sepconv1')(x)
x = BatchNormalization(name=prefix + '_sepconv1_bn')(x)
x = Activation('relu', name=prefix + '_sepconv2_act')(x)
x = SeparableConv2D(728, 3, 3, border_mode='same', bias=False, name=prefix + '_sepconv2')(x)
x = BatchNormalization(name=prefix + '_sepconv2_bn')(x)
x = Activation('relu', name=prefix + '_sepconv3_act')(x)
x = SeparableConv2D(728, 3, 3, border_mode='same', bias=False, name=prefix + '_sepconv3')(x)
x = BatchNormalization(name=prefix + '_sepconv3_bn')(x)
x = merge([x, residual], mode='sum')
residual = Conv2D(1024, 1, 1, subsample=(2, 2),
border_mode='same', bias=False)(x)
residual = BatchNormalization()(residual)
x = Activation('relu', name='block13_sepconv1_act')(x)
x = SeparableConv2D(728, 3, 3, border_mode='same', bias=False, name='block13_sepconv1')(x)
x = BatchNormalization(name='block13_sepconv1_bn')(x)
x = Activation('relu', name='block13_sepconv2_act')(x)
x = SeparableConv2D(1024, 3, 3, border_mode='same', bias=False, name='block13_sepconv2')(x)
x = BatchNormalization(name='block13_sepconv2_bn')(x)
x = MaxPooling2D((3, 3), strides=(2, 2), border_mode='same', name='block13_pool')(x)
x = merge([x, residual], mode='sum')
x = SeparableConv2D(1536, 3, 3, border_mode='same', bias=False, name='block14_sepconv1')(x)
x = BatchNormalization(name='block14_sepconv1_bn')(x)
x = Activation('relu', name='block14_sepconv1_act')(x)
x = SeparableConv2D(2048, 3, 3, border_mode='same', bias=False, name='block14_sepconv2')(x)
x = BatchNormalization(name='block14_sepconv2_bn')(x)
x = Activation('relu', name='block14_sepconv2_act')(x)
if include_top:
x = GlobalAveragePooling2D(name='avg_pool')(x)
x = Dense(1000, activation='softmax', name='predictions')(x)
# Create model
model = Model(img_input, x)
# load weights
if weights == 'imagenet':
if include_top:
weights_path = get_file('xception_weights_tf_dim_ordering_tf_kernels.h5',
TF_WEIGHTS_PATH,
cache_subdir='models')
else:
weights_path = get_file('xception_weights_tf_dim_ordering_tf_kernels_notop.h5',
TF_WEIGHTS_PATH_NO_TOP,
cache_subdir='models')
model.load_weights(weights_path)
if old_dim_ordering:
K.set_image_dim_ordering(old_dim_ordering)
return model