def __init__(self, inshape, hidshape, noutputs):
ninput = np.prod(inshape)
nhid = np.prod(hidshape)
nparams = (ninput + 1) * nhid + (nhid * noutputs)
# TODO:
self.params = np.empty(nparams)
self._grad = np.empty(nparams)
inhidwts = ninput * nhid
hidoutwts = nhid * noutputs
self.layers = [
LinearLayer(
inshape,
hidshape,
params=self.params[0:inhidwts],
grad=self._grad[0:inhidwts]
),
LogisticLayer(
hidshape,
params=self.params[inhidwts:(inhidwts + nhid)],
grad=self._grad[inhidwts:(inhidwts + nhid)]
),
LinearLayer(
hidshape,
noutputs,
params=self.params[(inhidwts + nhid):],
grad=self._grad[(inhidwts + nhid):]
),
SoftmaxLayer()
]
def build_model(text_len, negative_size, optimizer, word_size, entity_size,
dim_size, word_static, entity_static, word_embedding, entity_embedding):
text_input_layer = Input(shape=(text_len,), dtype='int32')
word_embed_layer = Embedding(
word_size, dim_size, input_length=text_len, name='word_embedding',
weights=[word_embedding], trainable=not word_static
)(text_input_layer)
text_layer = TextRepresentationLayer(name='text_layer')(
[word_embed_layer, text_input_layer]
)
entity_input_layer = Input(shape=(negative_size + 1,), dtype='int32')
entity_embed_layer = Embedding(
entity_size, dim_size, input_length=negative_size + 1,
name='entity_embedding', weights=[entity_embedding],
trainable=not entity_static
)(entity_input_layer)
similarity_layer = DotLayer(name='dot_layer')(
[RepeatVector(negative_size + 1)(text_layer), entity_embed_layer]
)
predictions = SoftmaxLayer()(similarity_layer)
model = Model(input=[text_input_layer, entity_input_layer],
output=predictions)
model.compile(optimizer=optimizer, loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def test_softmax_layer():
x_train = np.array([[5.1, 3.5, 1.4, 0.2], [4.9, 3.0, 1.4, 0.2],
[7.0, 3.2, 4.7, 1.4], [6.4, 3.2, 4.5, 1.5],
[6.3, 3.3, 6.0, 2.5], [5.8, 2.7, 5.1, 1.9]])
y_train = np.array([0, 0, 1, 1, 2, 2])
softmax = SoftmaxLayer()
W = 0.001 * np.random.randn(4, 3)
b = np.zeros((1, 3))
reg_parameter = 0.001
g_numerical_W = eval_numerical_gradient(softmax.forward_pass, x_train,
y_train, W, b, reg_parameter)
g_analytical_W = eval_analytical_gradient(softmax, x_train, y_train, W, b,
reg_parameter)
assert check_gradient(
g_numerical_W, g_analytical_W
) <= 1e-7, "Error in calculating gradient of the SoftmaxLayer"
def test_tanh_layer():
x_train = np.array([[5.1, 3.5, 1.4, 0.2], [4.9, 3.0, 1.4, 0.2],
[7.0, 3.2, 4.7, 1.4], [6.4, 3.2, 4.5, 1.5],
[6.3, 3.3, 6.0, 2.5], [5.8, 2.7, 5.1, 1.9]])
y_train = np.array([0, 0, 1, 1, 2, 2])
W1 = np.random.randn(4, 10) * 0.001
b1 = np.zeros((1, 10))
W2 = np.random.randn(10, 6) * 0.001
b2 = np.zeros((1, 6))
softmax = SoftmaxLayer()
tanh = TanhLayer()
reg_parameter = 0.001
g_numerical_W = eval_hidden_numerical_gradient(tanh, softmax, x_train,
y_train, W1, b1, W2, b2,
reg_parameter)
g_analytical_W = eval_hidden_analytical_gradient(tanh, softmax, x_train,
y_train, W1, b1, W2, b2,
reg_parameter)
assert check_gradient(
g_numerical_W, g_analytical_W
) <= 1e-7, "Error in calculating gradient of the TanhLayer"
def __init__(self,
input_shape=(None, 3, None, None),
n_classes=6,
n_filters_first_conv=48,
n_pool=4,
growth_rate=12,
n_layers_per_block=5,
dropout_p=0.5):
"""
This code implements the Fully Convolutional DenseNet described in https://arxiv.org/abs/1611.09326
The network consist of a downsampling path, where dense blocks and transition down are applied, followed
by an upsampling path where transition up and dense blocks are applied.
Skip connections are used between the downsampling path and the upsampling path
Each layer is a composite function of BN - ReLU - Conv and the last layer is a softmax layer.
:param input_shape: shape of the input batch. Only the first dimension (n_channels) is needed
:param n_classes: number of classes
:param n_filters_first_conv: number of filters for the first convolution applied
:param n_pool: number of pooling layers = number of transition down = number of transition up
:param growth_rate: number of new feature maps created by each layer in a dense block
:param n_layers_per_block: number of layers per block. Can be an int or a list of size 2 * n_pool + 1
:param dropout_p: dropout rate applied after each convolution (0. for not using)
"""
if type(n_layers_per_block) == list:
assert (len(n_layers_per_block) == 2 * n_pool + 1)
elif type(n_layers_per_block) == int:
n_layers_per_block = [n_layers_per_block] * (2 * n_pool + 1)
else:
raise ValueError
# Theano variables
self.input_var = T.tensor4('input_var', dtype='float32') # input image
self.target_var = T.tensor4('target_var', dtype='int32') # target
#####################
# First Convolution #
#####################
inputs = InputLayer(input_shape, self.input_var)
# We perform a first convolution. All the features maps will be stored in the tensor called stack (the Tiramisu)
stack = Conv2DLayer(inputs,
n_filters_first_conv,
filter_size=3,
pad='same',
W=HeUniform(gain='relu'),
nonlinearity=linear,
flip_filters=False)
# The number of feature maps in the stack is stored in the variable n_filters
n_filters = n_filters_first_conv
#####################
# Downsampling path #
#####################
skip_connection_list = []
for i in range(n_pool):
# Dense Block
for j in range(n_layers_per_block[i]):
# Compute new feature maps
l = BN_ReLU_Conv(stack, growth_rate, dropout_p=dropout_p)
# And stack it : the Tiramisu is growing
stack = ConcatLayer([stack, l])
n_filters += growth_rate
# At the end of the dense block, the current stack is stored in the skip_connections list
skip_connection_list.append(stack)
# Transition Down
stack = TransitionDown(stack, n_filters, dropout_p)
skip_connection_list = skip_connection_list[::-1]
#####################
# Bottleneck #
#####################
# We store now the output of the next dense block in a list. We will only upsample these new feature maps
block_to_upsample = []
# Dense Block
for j in range(n_layers_per_block[n_pool]):
l = BN_ReLU_Conv(stack, growth_rate, dropout_p=dropout_p)
block_to_upsample.append(l)
stack = ConcatLayer([stack, l])
#######################
# Upsampling path #
#######################
for i in range(n_pool):
# Transition Up ( Upsampling + concatenation with the skip connection)
n_filters_keep = growth_rate * n_layers_per_block[n_pool + i]
stack = TransitionUp(skip_connection_list[i], block_to_upsample,
n_filters_keep)
# Dense Block
block_to_upsample = []
for j in range(n_layers_per_block[n_pool + i + 1]):
l = BN_ReLU_Conv(stack, growth_rate, dropout_p=dropout_p)
block_to_upsample.append(l)
stack = ConcatLayer([stack, l])
#####################
# Softmax #
#####################
self.output_layer = SoftmaxLayer(stack, n_classes)
import skimage.measure
import pickle
from readlabel import read_image
from network import Network
from layers import ConvPoolLayer, FullyConnectedLayer, SoftmaxLayer, ReLU, Sigmoid
whole_data = read_image(path1 = 'test_images/', path2 = './test_annotation', data_size = 1050)
whole_x = whole_data[0]
mean = whole_x.mean(axis=0)
std = whole_x.std(axis=0)
whole_x = (whole_x - mean) / std
whole_y = whole_data[1]
test_x = whole_x
test_y = whole_y
test_data = [test_x, test_y]
mini_batch_size = 1
# final
net = Network([ConvPoolLayer(filter_shape=(5, 5, 3, 9), image_shape=(mini_batch_size, 64, 64, 3), poolsize=2, activation_fn=ReLU),
ConvPoolLayer(filter_shape=(5, 5, 9, 18), image_shape=(mini_batch_size, 30, 30, 9), poolsize=2, activation_fn=ReLU),
ConvPoolLayer(filter_shape=(4, 4, 18, 36), image_shape=(mini_batch_size, 13, 13, 18), poolsize=2, activation_fn=ReLU),
FullyConnectedLayer(n_in=900, n_out=225, activation_fn=ReLU),
FullyConnectedLayer(n_in=225, n_out=50, activation_fn=ReLU),
SoftmaxLayer(n_in=50, n_out=20, activation_fn=None)], mini_batch_size)
print('start')
net.load_test(mini_batch_size, test_data, path='./finalparams_noact.pickle')
def __init__(self, config, testMode):
self.config = config
batch_size = config['batch_size']
lib_conv = config['lib_conv']
useLayers = config['useLayers']
#imgWidth = config['imgWidth']
#imgHeight = config['imgHeight']
initWeights = config['initWeights'] #if we wish to initialize alexnet with some weights. #need to make changes in layers.py to accept initilizing weights
if initWeights:
weightsDir = config['weightsDir']
weightFileTag = config['weightFileTag']
prob_drop = config['prob_drop']
# ##################### BUILD NETWORK ##########################
x = T.ftensor4('x')
mean = T.ftensor4('mean')
#y = T.lvector('y')
print '... building the model'
self.layers = []
params = []
weight_types = []
if useLayers >= 1:
convpool_layer1 = ConvPoolLayer(input=x-mean,
image_shape=(3, None, None, batch_size),
filter_shape=(3, 11, 11, 96),
convstride=4, padsize=0, group=1,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=True,
lib_conv=lib_conv,
initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W_0'+weightFileTag, 'b_0'+weightFileTag]
)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
if useLayers >= 2:
convpool_layer2 = ConvPoolLayer(input=convpool_layer1.output,
image_shape=(96, None, None, batch_size), #change from 27 to appropriate value sbased on conv1's output
filter_shape=(96, 5, 5, 256),
convstride=1, padsize=2, group=2,
poolsize=3, poolstride=2,
bias_init=0.1, lrn=True,
lib_conv=lib_conv,
initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W0_1'+weightFileTag, 'W1_1'+weightFileTag, 'b0_1'+weightFileTag, 'b1_1'+weightFileTag]
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
if useLayers >= 3:
convpool_layer3 = ConvPoolLayer(input=convpool_layer2.output,
image_shape=(256, None, None, batch_size),
filter_shape=(256, 3, 3, 384),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=0,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W_2'+weightFileTag, 'b_2'+weightFileTag]
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
if useLayers >= 4:
convpool_layer4 = ConvPoolLayer(input=convpool_layer3.output,
image_shape=(384, None, None, batch_size),
filter_shape=(384, 3, 3, 384),
convstride=1, padsize=1, group=2,
poolsize=1, poolstride=0,
bias_init=0.1, lrn=False,
lib_conv=lib_conv,
initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W0_3'+weightFileTag, 'W1_3'+weightFileTag, 'b0_3'+weightFileTag, 'b1_3'+weightFileTag]
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
if useLayers >= 5:
convpool_layer5 = ConvPoolLayer(input=convpool_layer4.output,
image_shape=(384, None, None, batch_size),
filter_shape=(384, 3, 3, 256),
convstride=1, padsize=1, group=2,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W0_4'+weightFileTag, 'W1_4'+weightFileTag, 'b0_4'+weightFileTag, 'b1_4'+weightFileTag]
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
if useLayers >= 6:
fc_layer6_input = T.flatten(convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=9216, n_out=4096, initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W_5'+weightFileTag, 'b_5'+weightFileTag])
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
if testMode:
dropout_layer6 = fc_layer6
else:
dropout_layer6 = DropoutLayer(fc_layer6.output, n_in=4096, n_out=4096, prob_drop=prob_drop)
if useLayers >= 7:
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096, initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W_6'+weightFileTag, 'b_6'+weightFileTag])
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
if testMode:
dropout_layer6 = fc_layer7
else:
dropout_layer7 = DropoutLayer(fc_layer7.output, n_in=4096, n_out=4096, prob_drop=prob_drop)
if useLayers >= 8:
softmax_layer8 = SoftmaxLayer(input=dropout_layer7.output, n_in=4096, n_out=1000, initWeights=initWeights, weightsDir=weightsDir, weightFiles=['W_7'+weightFileTag, 'b_7'+weightFileTag])
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.output = self.layers[useLayers-1]
self.params = params
self.x = x
self.mean = mean
self.weight_types = weight_types
self.batch_size = batch_size
self.useLayers = useLayers
self.outLayer = self.layers[useLayers-1]
meanVal = np.load(config['mean_file'])
meanVal = meanVal[:, :, :, np.newaxis].astype('float32') #x is 4d, with 'batch' number of images. meanVal has only '1' in the 'batch' dimension. subtraction wont work.
meanVal = np.tile(meanVal,(1,1,1,batch_size))
self.meanVal = meanVal
#meanVal = np.zeros([3,imgHeight,imgWidth,2], dtype='float32')
if useLayers >= 8: #if last layer is softmax, then its output is y_pred
finalOut = self.outLayer.y_pred
else:
finalOut = self.outLayer.output
self.forwardFunction = theano.function([self.x, In(self.mean, value=meanVal)], [finalOut])
i = T.lscalar() # mini-batch index
self.test_mb_predictions = theano.function([i],
self.layers[-1].y_out,
givens={self.x: observation},
on_unused_input='warn')
return self.test_mb_predictions(0)
#Initialize network
layers = [
FullyConnectedLayer(n_in=4, n_out=10),
FullyConnectedLayer(n_in=10, n_out=10),
SoftmaxLayer(n_in=10, n_out=2)
]
params = [param for layer in layers for param in layer.params]
iterations = mini_batch_size
x = T.vector("x")
y = T.ivector("y")
init_layer = layers[0]
init_layer.set_inpt(x, 1)
for j in xrange(1, len(layers)):
prev_layer, layer = layers[j - 1], layers[j]
layer.set_inpt(prev_layer.output, 1)
cost = T.argmax(T.log(layers[-1].output))
def __init__(self, config):
ModelBase.__init__(self)
self.config = config
self.verbose = self.config['verbose']
self.name = 'alexnet'
batch_size = config['batch_size']
flag_datalayer = config['use_data_layer']
lib_conv = config['lib_conv']
n_softmax_out = config['n_softmax_out']
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x = T.ftensor4('x')
y = T.lvector('y')
rand = T.fvector('rand')
lr = T.scalar('lr')
if self.verbose: print 'AlexNet 2/16'
self.layers = []
params = []
weight_types = []
if flag_datalayer:
data_layer = DataLayer(input=x,
image_shape=(3, 256, 256, batch_size),
cropsize=227,
rand=rand,
mirror=True,
flag_rand=config['rand_crop'])
layer1_input = data_layer.output
else:
layer1_input = x
convpool_layer1 = ConvPoolLayer(input=layer1_input,
image_shape=(3, 227, 227, batch_size),
filter_shape=(3, 11, 11, 96),
convstride=4,
padsize=0,
group=1,
poolsize=3,
poolstride=2,
bias_init=0.0,
lrn=True,
lib_conv=lib_conv,
verbose=self.verbose)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(input=convpool_layer1.output,
image_shape=(96, 27, 27, batch_size),
filter_shape=(96, 5, 5, 256),
convstride=1,
padsize=2,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.1,
lrn=True,
lib_conv=lib_conv,
verbose=self.verbose)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(input=convpool_layer2.output,
image_shape=(256, 13, 13, batch_size),
filter_shape=(256, 3, 3, 384),
convstride=1,
padsize=1,
group=1,
poolsize=1,
poolstride=0,
bias_init=0.0,
lrn=False,
lib_conv=lib_conv,
verbose=self.verbose)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(input=convpool_layer3.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 384),
convstride=1,
padsize=1,
group=2,
poolsize=1,
poolstride=0,
bias_init=0.1,
lrn=False,
lib_conv=lib_conv,
verbose=self.verbose)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(input=convpool_layer4.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 256),
convstride=1,
padsize=1,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.0,
lrn=False,
lib_conv=lib_conv,
verbose=self.verbose)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
fc_layer6_input = T.flatten(
convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
fc_layer6 = FCLayer(input=fc_layer6_input,
n_in=9216,
n_out=4096,
verbose=self.verbose)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output,
n_in=4096,
n_out=4096,
verbose=self.verbose)
fc_layer7 = FCLayer(input=dropout_layer6.output,
n_in=4096,
n_out=4096,
verbose=self.verbose)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output,
n_in=4096,
n_out=4096,
verbose=self.verbose)
softmax_layer8 = SoftmaxLayer(input=dropout_layer7.output,
n_in=4096,
n_out=n_softmax_out,
verbose=self.verbose)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.p_y_given_x = softmax_layer8.p_y_given_x
self.y_pred = softmax_layer8.y_pred
self.output = self.p_y_given_x
self.cost = softmax_layer8.negative_log_likelihood(y)
self.error = softmax_layer8.errors(y)
if n_softmax_out < 5:
self.error_top_5 = softmax_layer8.errors_top_x(y, n_softmax_out)
else:
self.error_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
# inputs
self.x = x
self.y = y
self.rand = rand
self.lr = lr
self.shared_x = theano.shared(
np.zeros(
(3, config['input_width'], config['input_height'],
config['file_batch_size']), # for loading large batch
dtype=theano.config.floatX),
borrow=True)
self.shared_y = theano.shared(np.zeros((config['file_batch_size'], ),
dtype=int),
borrow=True)
self.shared_lr = theano.shared(np.float32(config['learning_rate']))
# training related
self.base_lr = np.float32(config['learning_rate'])
self.step_idx = 0
self.mu = config['momentum'] # def: 0.9 # momentum
self.eta = config['weight_decay'] #0.0002 # weight decay
self.weight_types = weight_types
self.batch_size = batch_size
self.grads = T.grad(self.cost, self.params)
subb_ind = T.iscalar('subb') # sub batch index
#print self.shared_x[:,:,:,subb_ind*self.batch_size:(subb_ind+1)*self.batch_size].shape.eval()
self.subb_ind = subb_ind
self.shared_x_slice = self.shared_x[:, :, :, subb_ind *
self.batch_size:(subb_ind + 1) *
self.batch_size]
self.shared_y_slice = self.shared_y[subb_ind *
self.batch_size:(subb_ind + 1) *
self.batch_size]
def __init__(self, config):
self.config = config
batch_size = config['batch_size']
flag_datalayer = config['use_data_layer']
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x = T.ftensor4('x')
y = T.lvector('y')
rand = T.fvector('rand')
print '... building the model'
self.layers = []
params = []
weight_types = []
if flag_datalayer:
data_layer = DataLayer(input=x, image_shape=(batch_size, 3, 256, 256),
cropsize=227, rand=rand, mirror=True,
flag_rand=config['rand_crop'])
layer1_input = data_layer.output
else:
layer1_input = x
convpool_layer1 = ConvPoolLayer(input=layer1_input,
image_shape=(batch_size, 3, 227, 227),
filter_shape=(96, 3, 11, 11),
convstride=4, padsize=0, group=1,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=True,
)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(input=convpool_layer1.output,
image_shape=(batch_size, 96, 27, 27),
filter_shape=(256, 96, 5, 5),
#convstride=1, padsize=2, group=1,
convstride=1, padsize=2, group=2,
poolsize=3, poolstride=2,
bias_init=0.1, lrn=True,
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(input=convpool_layer2.output,
image_shape=(batch_size, 256, 13, 13),
filter_shape=(384, 256, 3, 3),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=0,
bias_init=0.0, lrn=False,
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(input=convpool_layer3.output,
image_shape=(batch_size, 384, 13, 13),
filter_shape=(384, 384, 3, 3),
#convstride=1, padsize=1, group=1,
convstride=1, padsize=1, group=2,
poolsize=1, poolstride=0,
bias_init=0.1, lrn=False,
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(input=convpool_layer4.output,
image_shape=(batch_size, 384, 13, 13),
filter_shape=(256, 384, 3, 3),
#convstride=1, padsize=1, group=1,
convstride=1, padsize=1, group=2,
poolsize=3, poolstride=2,
bias_init=0.1, lrn=False,
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
fc_layer6_input = T.flatten(convpool_layer5.output, 2)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=9216, n_out=4096)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output, n_in=4096, n_out=4096)
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output, n_in=4096, n_out=4096)
softmax_layer8 = SoftmaxLayer(
input=dropout_layer7.output, n_in=4096, n_out=1000)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
self.x = x
self.y = y
self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
def __init__(self, config):
self.config = config
batch_size = config['batch_size']
num_seq = config['num_seq']
self.n_timesteps = config['num_timesteps']
num_joints = config['num_joints']
classes_num = config['classes_num']
# ##################### BUILD NETWORK ##########################
mask = T.fvector('mask')
y = T.lvector('y')
target = T.ftensor3('target')
rand = T.fvector('rand')
trng = RandomStreams(1234)
use_noise = T.fscalar('use_noise')
print '... building the model'
self.layers = []
params = []
weight_types = []
conv_fea = T.ftensor4('conv_fea') #(49, 16, 8, 1024)
lstm_att_layer15 = JointAttentionLstmLayer(config,
num_joints,
conv_fea=conv_fea,
mask=mask,
batch_size=batch_size,
num_seq=num_seq,
trng=trng,
use_noise=use_noise,
n_in=1024 * 5,
n_out=1024,
dim_part=32)
self.layers.append(lstm_att_layer15)
params += lstm_att_layer15.params
weight_types += lstm_att_layer15.weight_type
self.conv_fea = conv_fea
softmax_input = lstm_att_layer15.output
softmax_layer15 = SoftmaxLayer(input=softmax_input,
n_in=1024,
n_out=21)
self.layers.append(softmax_layer15)
params += softmax_layer15.params
weight_types += softmax_layer15.weight_type
# #################### NETWORK BUILT #######################
self.cost_nll = softmax_layer15.negative_log_likelihood(y, mask)
self.cost_jhmdb_attention = T.mean(T.sum(T.sum(
0.5 * (lstm_att_layer15.attention - target)**2, axis=1),
axis=1),
axis=0,
dtype=theano.config.floatX)
self.cost = self.cost_nll + self.cost_jhmdb_attention
self.errors_video = softmax_layer15.errors_video(
y, mask, batch_size, num_seq)
self.params = params
self.prob = softmax_layer15.p_y_given_x
self.mask = mask
self.y = y
self.target = target
self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
self.num_seq = num_seq
self.use_noise = use_noise
x_train, y_train = load_mnist(MNIST_TRAINING_X , MNIST_TRAINING_y)
x_train = x_train.reshape(MNIST_NUM_TRAINING, MNIST_NUM_FEATURES)
y_train = y_train.reshape(MNIST_NUM_TRAINING)
# initialize parameters randomly
HIDDEN_LAYER_SIZE = 500
W1 = 0.1 * np.random.randn(MNIST_NUM_FEATURES, HIDDEN_LAYER_SIZE)
b1 = np.zeros((1, HIDDEN_LAYER_SIZE))
W2 = 0.1 * np.random.randn(HIDDEN_LAYER_SIZE, MNIST_NUM_OUTPUT)
b2 = np.zeros((1, MNIST_NUM_OUTPUT))
learning_rate = 0.1 # step size of the gradient descent algorithm
reg_parameter = 0.001 # regularization strength
softmax = SoftmaxLayer()
hidden = TanhLayer()
num_iter = 12000
BATCH_SIZE = 100
for i in range(num_iter):
idx = np.random.choice(MNIST_NUM_TRAINING, BATCH_SIZE, replace=True)
x_batch = x_train[idx, :]
y_batch = y_train[idx]
pre_activation, hidden_output = hidden.forward_pass(x_batch, W1, b1)
hidden_layer_weights = [W1]
output_prob, loss = softmax.forward_pass(hidden_output, y_batch, W2, b2, reg_parameter, hidden_layer_weights)
def __init__(self, config):
self.config = config
self.name = "alexnet"
batch_size = config["batch_size"]
flag_datalayer = config["use_data_layer"]
lib_conv = config["lib_conv"]
n_softmax_out = config["n_softmax_out"]
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x = T.ftensor4("x")
y = T.lvector("y")
rand = T.fvector("rand")
print "... building the model"
self.layers = []
params = []
weight_types = []
if flag_datalayer:
data_layer = DataLayer(
input=x,
image_shape=(3, 256, 256, batch_size),
cropsize=227,
rand=rand,
mirror=True,
flag_rand=config["rand_crop"],
)
layer1_input = data_layer.output
else:
layer1_input = x
convpool_layer1 = ConvPoolLayer(
input=layer1_input,
image_shape=(3, 227, 227, batch_size),
filter_shape=(3, 11, 11, 96),
convstride=4,
padsize=0,
group=1,
poolsize=3,
poolstride=2,
bias_init=0.0,
lrn=True,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(
input=convpool_layer1.output,
image_shape=(96, 27, 27, batch_size),
filter_shape=(96, 5, 5, 256),
convstride=1,
padsize=2,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.1,
lrn=True,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(
input=convpool_layer2.output,
image_shape=(256, 13, 13, batch_size),
filter_shape=(256, 3, 3, 384),
convstride=1,
padsize=1,
group=1,
poolsize=1,
poolstride=0,
bias_init=0.0,
lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(
input=convpool_layer3.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 384),
convstride=1,
padsize=1,
group=2,
poolsize=1,
poolstride=0,
bias_init=0.1,
lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(
input=convpool_layer4.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 256),
convstride=1,
padsize=1,
group=2,
poolsize=3,
poolstride=2,
bias_init=0.0,
lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
fc_layer6_input = T.flatten(convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=9216, n_out=4096)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output, n_in=4096, n_out=4096)
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output, n_in=4096, n_out=4096)
softmax_layer8 = SoftmaxLayer(input=dropout_layer7.output, n_in=4096, n_out=n_softmax_out)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.p_y_given_x = softmax_layer8.p_y_given_x ###############ADDED##########
self.y_pred = softmax_layer8.y_pred ########ADDED##############
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
if n_softmax_out < 5:
self.errors_top_5 = softmax_layer8.errors_top_x(y, n_softmax_out) ##########MODIFIED##########
else:
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5) ##########MODIFIED##########
self.params = params
self.x = x
self.y = y
self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
def __init__(self, config):
self.config = config
batch_size = config['batch_size']
flag_datalayer = config['use_data_layer']
lib_conv = config['lib_conv']
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x = T.ftensor4('x')
y = T.lvector('y')
rand = T.fvector('rand')
print '... building the model'
self.layers = []
params = []
weight_types = []
if flag_datalayer:
data_layer = DataLayer(input=x, image_shape=(3, 256, 256,
batch_size),
cropsize=224, rand=rand, mirror=True,
flag_rand=config['rand_crop'])
layer1_input = data_layer.output
else:
layer1_input = x
convpool_layer1_1 = ConvPoolLayer(input=layer1_input,
image_shape=(3, 224, 224, batch_size),
filter_shape=(3, 3, 3, 64),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer1_1)
params += convpool_layer1_1.params
weight_types += convpool_layer1_1.weight_type
convpool_layer1_2 = ConvPoolLayer(input=convpool_layer1_1.output,
image_shape=(64, 224, 224, batch_size),
filter_shape=(64, 3, 3, 64),
convstride=1, padsize=1, group=1,
poolsize=2, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer1_2)
params += convpool_layer1_2.params
weight_types += convpool_layer1_2.weight_type
convpool_layer2_1 = ConvPoolLayer(input=convpool_layer1_2.output,
image_shape=(64, 112, 112, batch_size),
filter_shape=(64, 3, 3, 128),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer2_1)
params += convpool_layer2_1.params
weight_types += convpool_layer2_1.weight_type
convpool_layer2_2 = ConvPoolLayer(input=convpool_layer2_1.output,
image_shape=(128, 112, 112, batch_size),
filter_shape=(128, 3, 3, 128),
convstride=1, padsize=1, group=1,
poolsize=2, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer2_2)
params += convpool_layer2_2.params
weight_types += convpool_layer2_2.weight_type
convpool_layer3_1 = ConvPoolLayer(input=convpool_layer2_2.output,
image_shape=(128, 56, 56, batch_size),
filter_shape=(128, 3, 3, 256),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3_1)
params += convpool_layer3_1.params
weight_types += convpool_layer3_1.weight_type
convpool_layer3_2 = ConvPoolLayer(input=convpool_layer3_1.output,
image_shape=(256, 56, 56, batch_size),
filter_shape=(256, 3, 3, 256),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3_2)
params += convpool_layer3_2.params
weight_types += convpool_layer3_2.weight_type
convpool_layer3_3 = ConvPoolLayer(input=convpool_layer3_2.output,
image_shape=(256, 56, 56, batch_size),
filter_shape=(256, 3, 3, 256),
convstride=1, padsize=1, group=1,
poolsize=2, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3_3)
params += convpool_layer3_3.params
weight_types += convpool_layer3_3.weight_type
convpool_layer4_1 = ConvPoolLayer(input=convpool_layer3_3.output,
image_shape=(256, 28, 28, batch_size),
filter_shape=(256, 3, 3, 512),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4_1)
params += convpool_layer4_1.params
weight_types += convpool_layer4_1.weight_type
convpool_layer4_2 = ConvPoolLayer(input=convpool_layer4_1.output,
image_shape=(512, 28, 28, batch_size),
filter_shape=(512, 3, 3, 512),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4_2)
params += convpool_layer4_2.params
weight_types += convpool_layer4_2.weight_type
convpool_layer4_3 = ConvPoolLayer(input=convpool_layer4_2.output,
image_shape=(512, 28, 28, batch_size),
filter_shape=(512, 3, 3, 512),
convstride=1, padsize=1, group=1,
poolsize=2, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4_3)
params += convpool_layer4_3.params
weight_types += convpool_layer4_3.weight_type
convpool_layer5_1 = ConvPoolLayer(input=convpool_layer4_3.output,
image_shape=(512, 14, 14, batch_size),
filter_shape=(512, 3, 3, 512),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5_1)
params += convpool_layer5_1.params
weight_types += convpool_layer5_1.weight_type
convpool_layer5_2 = ConvPoolLayer(input=convpool_layer5_1.output,
image_shape=(512, 14, 14, batch_size),
filter_shape=(512, 3, 3, 512),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=1,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5_2)
params += convpool_layer5_2.params
weight_types += convpool_layer5_2.weight_type
convpool_layer5_3 = ConvPoolLayer(input=convpool_layer5_2.output,
image_shape=(512, 14, 14, batch_size),
filter_shape=(512, 3, 3, 512),
convstride=1, padsize=1, group=1,
poolsize=2, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5_3)
params += convpool_layer5_3.params
weight_types += convpool_layer5_3.weight_type
fc_layer6_input = T.flatten(convpool_layer5_3.output.dimshuffle(3, 0, 1, 2), 2)
# dimshuffle is changing the batch dimension to the first dimension
# flatten crash the output to (batch, evethingelse) 2-D array
# fc_layer6 = FCLayer(input=fc_layer6_input, n_in=224*224*64, n_out=4096)
# fc_layer6 = FCLayer(input=fc_layer6_input, n_in=401408, n_out=4096)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=25088, n_out=4096)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output, n_in=4096, n_out=4096)
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output, n_in=4096, n_out=4096)
softmax_layer8 = SoftmaxLayer(
input=dropout_layer7.output, n_in=4096, n_out=1000)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
self.x = x
self.y = y
self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
def __init__(self, config):
self.config = config
batch_size = config.batch_size
lib_conv = config.lib_conv
group = (2 if config.grouping else 1)
LRN = (True if config.LRN else False)
print 'LRN, group', LRN, group
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
x = T.ftensor4('x')
y = T.lvector('y')
print '... building the model with ConvLib %s, LRN %s, grouping %i ' \
% (lib_conv, LRN, group)
self.layers = []
params = []
weight_types = []
layer1_input = x
convpool_layer1 = ConvPoolLayer(
input=layer1_input,
image_shape=((3, 224, 224,
batch_size) if lib_conv == 'cudaconvnet' else
(batch_size, 3, 227, 227)),
filter_shape=((3, 11, 11, 96) if lib_conv == 'cudaconvnet' else
(96, 3, 11, 11)),
convstride=4,
padsize=(0 if lib_conv == 'cudaconvnet' else 3),
group=1,
poolsize=3,
poolstride=2,
bias_init=0.0,
lrn=LRN,
lib_conv=lib_conv)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(
input=convpool_layer1.output,
image_shape=((96, 27, 27,
batch_size) if lib_conv == 'cudaconvnet' else
(batch_size, 96, 27, 27)),
filter_shape=((96, 5, 5, 256) if lib_conv == 'cudaconvnet' else
(256, 96, 5, 5)),
convstride=1,
padsize=2,
group=group,
poolsize=3,
poolstride=2,
bias_init=0.1,
lrn=LRN,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(
input=convpool_layer2.output,
image_shape=((256, 13, 13,
batch_size) if lib_conv == 'cudaconvnet' else
(batch_size, 256, 13, 13)),
filter_shape=((256, 3, 3, 384) if lib_conv == 'cudaconvnet' else
(384, 256, 3, 3)),
convstride=1,
padsize=1,
group=1,
poolsize=1,
poolstride=0,
bias_init=0.0,
lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(
input=convpool_layer3.output,
image_shape=((384, 13, 13,
batch_size) if lib_conv == 'cudaconvnet' else
(batch_size, 384, 13, 13)),
filter_shape=((384, 3, 3, 384) if lib_conv == 'cudaconvnet' else
(384, 384, 3, 3)),
convstride=1,
padsize=1,
group=group,
poolsize=1,
poolstride=0,
bias_init=0.1,
lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(
input=convpool_layer4.output,
image_shape=((384, 13, 13,
batch_size) if lib_conv == 'cudaconvnet' else
(batch_size, 384, 13, 13)),
filter_shape=((384, 3, 3, 256) if lib_conv == 'cudaconvnet' else
(256, 384, 3, 3)),
convstride=1,
padsize=1,
group=group,
poolsize=3,
poolstride=2,
bias_init=0.0,
lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
if lib_conv == 'cudaconvnet':
fc_layer6_input = T.flatten(
convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
else:
fc_layer6_input = convpool_layer5.output.flatten(2)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=9216, n_out=4096)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output)
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output)
softmax_layer8 = SoftmaxLayer(input=dropout_layer7.output,
n_in=4096,
n_out=1000)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
self.x = x
self.y = y
# self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
def __init__(self,
input,
n_in=28**2,
n_hidden_1=1024,
n_hidden_2=1024,
n_hidden_3=1024,
n_hidden_4=1024,
n_out=10,
W_hidden_1=None,
W_hidden_2=None,
W_hidden_3=None,
W_hidden_4=None,
W_out=None,
dropout=0.0,
seed=None):
relu_activation = lambda x: T.nnet.relu(x, 0.1)
# relu_activation = T.nnet.relu
seed = np.random.randint(int(1e5)) if seed is None else seed
self.dropout_layer_1 = DropoutLayer(input=input,
seed=seed,
dropout=dropout)
self.hidden_1 = HiddenLayer(
seed=seed + 1,
# input=input,
input=self.dropout_layer_1.output,
# input=self.dropout_layer.output,
n_in=n_in,
n_out=n_hidden_1,
activation=relu_activation,
W=W_hidden_1,
)
self.dropout_layer_2 = DropoutLayer(input=self.hidden_1.output,
seed=seed + 2,
dropout=dropout)
self.hidden_2 = HiddenLayer(
seed=seed + 3,
# input=self.hidden_1.output,
input=self.dropout_layer_2.output,
n_in=n_hidden_1,
n_out=n_hidden_2,
activation=relu_activation,
W=W_hidden_2)
self.dropout_layer_3 = DropoutLayer(input=self.hidden_2.output,
seed=seed + 4,
dropout=dropout)
self.hidden_3 = HiddenLayer(seed=seed + 5,
input=self.dropout_layer_3.output,
n_in=n_hidden_2,
n_out=n_hidden_3,
activation=relu_activation,
W=W_hidden_3)
self.dropout_layer_4 = DropoutLayer(input=self.hidden_3.output,
seed=seed + 6,
dropout=dropout)
self.hidden_4 = HiddenLayer(seed=seed + 7,
input=self.dropout_layer_4.output,
n_in=n_hidden_3,
n_out=n_hidden_4,
activation=relu_activation,
W=W_hidden_4)
self.dropout_layer_5 = DropoutLayer(input=self.hidden_4.output,
seed=seed + 8,
dropout=dropout)
self.linear_layer = HiddenLayer(
seed=seed + 9,
# input=self.hidden_1.output,
# input=self.hidden_2.output,
input=self.dropout_layer_5.output,
n_in=n_hidden_4,
n_out=n_out,
activation=identity_map,
W=W_out)
self.softmax_layer = SoftmaxLayer(input=self.linear_layer.output)
# keep track of model input
self.input = input
self.p_y_given_x = self.softmax_layer.p_y_given_x
self.y_pred = self.softmax_layer.y_pred
self.L1 = (abs(self.hidden_1.W).sum() + abs(self.hidden_2.W).sum() +
abs(self.hidden_3.W).sum() + abs(self.hidden_4.W).sum() +
abs(self.linear_layer.W).sum())
self.L2_sqr = (T.sum(self.hidden_1.W**2) + T.sum(self.hidden_2.W**2) +
T.sum(self.hidden_3.W**2) + T.sum(self.hidden_4.W**2) +
T.sum(self.linear_layer.W**2))
self.mean_log_likelihood = (self.softmax_layer.mean_log_likelihood)
self.errors = self.softmax_layer.errors
self.params = (self.hidden_1.params + self.hidden_2.params +
self.hidden_3.params + self.hidden_4.params +
self.linear_layer.params)
x_train, y_train = load_mnist(MNIST_TRAINING_X, MNIST_TRAINING_y)
x_train = x_train.reshape(MNIST_NUM_TRAINING, MNIST_NUM_FEATURES)
y_train = y_train.reshape(MNIST_NUM_TRAINING)
# initialize parameters randomly
HIDDEN_LAYER_SIZE = 500
W1 = 0.1 * np.random.randn(MNIST_NUM_FEATURES, HIDDEN_LAYER_SIZE)
b1 = np.zeros((1, HIDDEN_LAYER_SIZE))
W2 = 0.1 * np.random.randn(HIDDEN_LAYER_SIZE, MNIST_NUM_OUTPUT)
b2 = np.zeros((1, MNIST_NUM_OUTPUT))
learning_rate = 0.1 # step size of the gradient descent algorithm
reg_parameter = 0.001 # regularization strength
softmax = SoftmaxLayer()
hidden = TanhLayer()
num_iter = 12000
BATCH_SIZE = 100
for i in range(num_iter):
idx = np.random.choice(MNIST_NUM_TRAINING, BATCH_SIZE, replace=True)
x_batch = x_train[idx, :]
y_batch = y_train[idx]
pre_activation, hidden_output = hidden.forward_pass(x_batch, W1, b1)
hidden_layer_weights = [W1]
output_prob, loss = softmax.forward_pass(hidden_output, y_batch, W2, b2,
reg_parameter,
test_data = [test_x, test_y]
# final
net = Network([
ConvPoolLayer(filter_shape=(5, 5, 3, 9),
image_shape=(mini_batch_size, 64, 64, 3),
poolsize=2,
activation_fn=ReLU),
ConvPoolLayer(filter_shape=(5, 5, 9, 18),
image_shape=(mini_batch_size, 30, 30, 9),
poolsize=2,
activation_fn=ReLU),
ConvPoolLayer(filter_shape=(4, 4, 18, 36),
image_shape=(mini_batch_size, 13, 13, 18),
poolsize=2,
activation_fn=ReLU),
FullyConnectedLayer(n_in=900, n_out=225, activation_fn=ReLU),
FullyConnectedLayer(n_in=225, n_out=50, activation_fn=ReLU),
SoftmaxLayer(n_in=50, n_out=20, activation_fn=None)
], mini_batch_size)
print('start')
net.train_save(training_data,
13,
mini_batch_size,
0.001,
validation_data,
test_data,
test=False,
save=2)
def __init__(self, config):
self.config = config
batch_size = config['batch_size']
flag_datalayer = config['use_data_layer']
lib_conv = config['lib_conv']
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x = T.ftensor4('x')
y = T.ivector('y')
rand = T.fvector('rand')
print '... building the model'
self.layers = []
params = []
weight_types = []
if flag_datalayer:
data_layer = DataLayer(input=x, image_shape=(3, 256, 256,
batch_size),
cropsize=227, rand=rand, mirror=True,
flag_rand=config['rand_crop'])
layer1_input = data_layer.output
else:
layer1_input = x
convpool_layer1 = ConvPoolLayer(input=layer1_input,
image_shape=(3, 227, 227, batch_size),
filter_shape=(3, 11, 11, 96),
convstride=4, padsize=0, group=1,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=True,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(input=convpool_layer1.output,
image_shape=(96, 27, 27, batch_size),
filter_shape=(96, 5, 5, 256),
convstride=1, padsize=2, group=2,
poolsize=3, poolstride=2,
bias_init=0.1, lrn=True,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(input=convpool_layer2.output,
image_shape=(256, 13, 13, batch_size),
filter_shape=(256, 3, 3, 384),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=0,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(input=convpool_layer3.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 384),
convstride=1, padsize=1, group=2,
poolsize=1, poolstride=0,
bias_init=0.1, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(input=convpool_layer4.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 256),
convstride=1, padsize=1, group=2,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
fc_layer6_input = T.flatten(
convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=9216, n_out=4096)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output, n_in=4096, n_out=4096)
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output, n_in=4096, n_out=4096)
softmax_layer8 = SoftmaxLayer(
input=dropout_layer7.output, n_in=4096, n_out=1000)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
self.x = x
self.y = y
self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
def __init__(self, config):
self.config = config
batch_size = config.batch_size
lib_conv = config.lib_conv
group = (2 if config.grouping else 1)
LRN = (True if config.LRN else False)
print 'LRN, group', LRN, group
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
x = T.ftensor4('x')
y = T.lvector('y')
print '... building the model with ConvLib %s, LRN %s, grouping %i ' \
% (lib_conv, LRN, group)
self.layers = []
params = []
weight_types = []
layer1_input = x
convpool_layer1 = ConvPoolLayer(
input=layer1_input,
image_shape=((3, 224, 224, batch_size) if lib_conv == 'cudaconvnet'
else (batch_size, 3, 227, 227)),
filter_shape=((3, 11, 11, 96) if lib_conv == 'cudaconvnet'
else (96, 3, 11, 11)),
convstride=4,
padsize=(0 if lib_conv == 'cudaconvnet' else 3),
group=1,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=LRN,
lib_conv=lib_conv)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(
input=convpool_layer1.output,
image_shape=((96, 27, 27, batch_size) if lib_conv == 'cudaconvnet'
else (batch_size, 96, 27, 27)),
filter_shape=((96, 5, 5, 256) if lib_conv == 'cudaconvnet'
else (256, 96, 5, 5)),
convstride=1, padsize=2, group=group,
poolsize=3, poolstride=2,
bias_init=0.1, lrn=LRN,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(
input=convpool_layer2.output,
image_shape=((256, 13, 13, batch_size) if lib_conv == 'cudaconvnet'
else (batch_size, 256, 13, 13)),
filter_shape=((256, 3, 3, 384) if lib_conv == 'cudaconvnet'
else (384, 256, 3, 3)),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=0,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(
input=convpool_layer3.output,
image_shape=((384, 13, 13, batch_size) if lib_conv == 'cudaconvnet'
else (batch_size, 384, 13, 13)),
filter_shape=((384, 3, 3, 384) if lib_conv == 'cudaconvnet'
else (384, 384, 3, 3)),
convstride=1, padsize=1, group=group,
poolsize=1, poolstride=0,
bias_init=0.1, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(
input=convpool_layer4.output,
image_shape=((384, 13, 13, batch_size) if lib_conv == 'cudaconvnet'
else (batch_size, 384, 13, 13)),
filter_shape=((384, 3, 3, 256) if lib_conv == 'cudaconvnet'
else (256, 384, 3, 3)),
convstride=1, padsize=1, group=group,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
if lib_conv == 'cudaconvnet':
fc_layer6_input = T.flatten(
convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
else:
fc_layer6_input = convpool_layer5.output.flatten(2)
fc_layer6 = FCLayer(input=fc_layer6_input, n_in=9216, n_out=4096)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output)
fc_layer7 = FCLayer(input=dropout_layer6.output, n_in=4096, n_out=4096)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output)
softmax_layer8 = SoftmaxLayer(
input=dropout_layer7.output, n_in=4096, n_out=1000)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
self.x = x
self.y = y
# self.rand = rand
self.weight_types = weight_types
self.batch_size = batch_size
def __init__(self, config):
self.config = config
batch_size = config['batch_size']
batch_size = config['batch_size']
flag_datalayer = config['use_data_layer']
lib_conv = config['lib_conv']
layers = []
params = []
weight_types = []
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x1 = T.ftensor4('x1')
x2 = T.ftensor4('x2')
y = T.lvector('y') # The ground truth to be compared with will go here
rand1 = T.fvector('rand1')
rand2 = T.fvector('rand2')
print '... building the model'
if flag_datalayer:
data_layerA = DataLayer(input=x1,
image_shape=(3, 256, 256, batch_size),
cropsize=227,
rand=rand,
mirror=True,
flag_rand=config['rand_crop'])
layer1A_input = data_layerA.output
else:
layer1A_input = x1
if flag_datalayer:
data_layerB = DataLayer(input=x2,
image_shape=(3, 256, 256, batch_size),
cropsize=227,
rand=rand,
mirror=True,
flag_rand=config['rand_crop'])
layer1B_input = data_layerB.output
else:
layer1B_input = x2
fc_layer2_input = T.concatenate(
(T.flatten(layer1A_input.dimshuffle(3, 0, 1, 2),
2), T.flatten(layer1B_input.dimshuffle(3, 0, 1, 2), 2)),
axis=1)
fc_layer2 = FCLayer(input=fc_layer2_input, n_in=154587 * 2, n_out=4096)
layers.append(fc_layer2)
params += fc_layer2.params
weight_types += fc_layer2.weight_type
dropout_layer2 = DropoutLayer(fc_layer2.output, n_in=4096, n_out=4096)
fc_layer3 = FCLayer(input=dropout_layer2.output, n_in=4096, n_out=4096)
layers.append(fc_layer3)
params += fc_layer3.params
weight_types += fc_layer3.weight_type
dropout_layer3 = DropoutLayer(fc_layer3.output, n_in=4096, n_out=4096)
# Final softmax layer
softmax_layer3 = SoftmaxLayer(
input=dropout_layer3.output, n_in=4096,
n_out=2) # Only a single binary output is required!
layers.append(softmax_layer3)
params += softmax_layer3.params
weight_types += softmax_layer3.weight_type
# #################### NETWORK BUILT #######################
self.cost = softmax_layer3.negative_log_likelihood(y)
self.errors = softmax_layer3.errors(y)
self.errors_top_5 = softmax_layer3.errors_top_x(y, 5)
self.x1 = x1
self.x2 = x2
self.y = y
self.rand1 = rand1
self.rand2 = rand2
self.layers = layers
self.params = params
self.weight_types = weight_types
self.batch_size = batch_size
import numpy as np
from layers import SoftmaxLayer
from datareader import load_mnist
from constants import *
x_train, y_train = load_mnist(MNIST_TRAINING_X , MNIST_TRAINING_y)
x_train = x_train.reshape(MNIST_NUM_TRAINING, MNIST_NUM_FEATURES)
y_train = y_train.reshape(MNIST_NUM_TRAINING)
# initialize parameters randomly
W = 0.001 * np.random.randn(MNIST_NUM_FEATURES, MNIST_NUM_OUTPUT)
b = np.zeros((1, MNIST_NUM_OUTPUT))
learning_rate = 0.1 # step size of the gradient descent algorithm
reg_parameter = 0.01 # regularization strength
softmax = SoftmaxLayer()
num_iter = 1000
BATCH_SIZE = 500
for i in range(num_iter):
idx = np.random.choice(MNIST_NUM_TRAINING, BATCH_SIZE, replace=True)
x_batch = x_train[idx, :]
y_batch = y_train[idx]
output_prob, loss = softmax.forward_pass(x_batch, y_batch, W, b, reg_parameter, [])
if i % 50 == 0:
print('iteration: {:3d} loss: {:3e}'.format(i, loss))
gradW, gradB, _ = softmax.backward_pass(output_prob, x_batch, y_batch, W, b, reg_parameter)
W = W - learning_rate * gradW
def __init__(self, config):
ModelBase.__init__(self)
self.config = config
self.verbose = self.config['verbose']
self.name = 'alexnet'
batch_size = config['batch_size']
flag_datalayer = config['use_data_layer']
lib_conv = config['lib_conv']
n_softmax_out=config['n_softmax_out']
# ##################### BUILD NETWORK ##########################
# allocate symbolic variables for the data
# 'rand' is a random array used for random cropping/mirroring of data
x = T.ftensor4('x')
y = T.lvector('y')
rand = T.fvector('rand')
lr = T.scalar('lr')
if self.verbose: print 'AlexNet 2/16'
self.layers = []
params = []
weight_types = []
if flag_datalayer:
data_layer = DataLayer(input=x, image_shape=(3, 256, 256,
batch_size),
cropsize=227, rand=rand, mirror=True,
flag_rand=config['rand_crop'])
layer1_input = data_layer.output
else:
layer1_input = x
convpool_layer1 = ConvPoolLayer(input=layer1_input,
image_shape=(3, 227, 227, batch_size),
filter_shape=(3, 11, 11, 96),
convstride=4, padsize=0, group=1,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=True,
lib_conv=lib_conv,
verbose = self.verbose
)
self.layers.append(convpool_layer1)
params += convpool_layer1.params
weight_types += convpool_layer1.weight_type
convpool_layer2 = ConvPoolLayer(input=convpool_layer1.output,
image_shape=(96, 27, 27, batch_size),
filter_shape=(96, 5, 5, 256),
convstride=1, padsize=2, group=2,
poolsize=3, poolstride=2,
bias_init=0.1, lrn=True,
lib_conv=lib_conv,
verbose = self.verbose
)
self.layers.append(convpool_layer2)
params += convpool_layer2.params
weight_types += convpool_layer2.weight_type
convpool_layer3 = ConvPoolLayer(input=convpool_layer2.output,
image_shape=(256, 13, 13, batch_size),
filter_shape=(256, 3, 3, 384),
convstride=1, padsize=1, group=1,
poolsize=1, poolstride=0,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
verbose = self.verbose
)
self.layers.append(convpool_layer3)
params += convpool_layer3.params
weight_types += convpool_layer3.weight_type
convpool_layer4 = ConvPoolLayer(input=convpool_layer3.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 384),
convstride=1, padsize=1, group=2,
poolsize=1, poolstride=0,
bias_init=0.1, lrn=False,
lib_conv=lib_conv,
verbose = self.verbose
)
self.layers.append(convpool_layer4)
params += convpool_layer4.params
weight_types += convpool_layer4.weight_type
convpool_layer5 = ConvPoolLayer(input=convpool_layer4.output,
image_shape=(384, 13, 13, batch_size),
filter_shape=(384, 3, 3, 256),
convstride=1, padsize=1, group=2,
poolsize=3, poolstride=2,
bias_init=0.0, lrn=False,
lib_conv=lib_conv,
verbose = self.verbose
)
self.layers.append(convpool_layer5)
params += convpool_layer5.params
weight_types += convpool_layer5.weight_type
fc_layer6_input = T.flatten(
convpool_layer5.output.dimshuffle(3, 0, 1, 2), 2)
fc_layer6 = FCLayer(input=fc_layer6_input,
n_in=9216,
n_out=4096,
verbose = self.verbose
)
self.layers.append(fc_layer6)
params += fc_layer6.params
weight_types += fc_layer6.weight_type
dropout_layer6 = DropoutLayer(fc_layer6.output,
n_in=4096,
n_out=4096,
verbose = self.verbose)
fc_layer7 = FCLayer(input=dropout_layer6.output,
n_in=4096,
n_out=4096,
verbose = self.verbose
)
self.layers.append(fc_layer7)
params += fc_layer7.params
weight_types += fc_layer7.weight_type
dropout_layer7 = DropoutLayer(fc_layer7.output,
n_in=4096,
n_out=4096,
verbose = self.verbose)
softmax_layer8 = SoftmaxLayer(input=dropout_layer7.output,
n_in=4096,
n_out=n_softmax_out,
verbose = self.verbose)
self.layers.append(softmax_layer8)
params += softmax_layer8.params
weight_types += softmax_layer8.weight_type
# #################### NETWORK BUILT #######################
self.p_y_given_x = softmax_layer8.p_y_given_x
self.y_pred = softmax_layer8.y_pred
self.cost = softmax_layer8.negative_log_likelihood(y)
self.errors = softmax_layer8.errors(y)
if n_softmax_out < 5:
self.errors_top_5 = softmax_layer8.errors_top_x(y, n_softmax_out)
else:
self.errors_top_5 = softmax_layer8.errors_top_x(y, 5)
self.params = params
# inputs
self.x = x
self.y = y
self.rand = rand
self.lr = lr
self.shared_x = theano.shared(np.zeros((3, config['input_width'],
config['input_height'],
config['file_batch_size']), # for loading large batch
dtype=theano.config.floatX),
borrow=True)
self.shared_y = theano.shared(np.zeros((config['file_batch_size'],),
dtype=int), borrow=True)
self.shared_lr = theano.shared(np.float32(config['learning_rate']))
# training related
self.base_lr = np.float32(config['learning_rate'])
self.step_idx = 0
self.mu = config['momentum'] # def: 0.9 # momentum
self.eta = config['weight_decay'] #0.0002 # weight decay
self.weight_types = weight_types
self.batch_size = batch_size
self.grads = T.grad(self.cost,self.params)
# shared variable for storing momentum before exchanging momentum(delta w)
self.vels = [theano.shared(param_i.get_value() * 0.)
for param_i in self.params]
# shared variable for accepting momentum during exchanging momentum(delta w)
self.vels2 = [theano.shared(param_i.get_value() * 0.)
for param_i in self.params]
self.train = None
self.get_vel = None
self.descent_vel = None
self.val = None
self.inference = None
def __init__(self, X, n_in, n_out, n_hidden_layers,
n_units_in, n_units_hidden,
M_lst=None, m_lst=None,
sigma_W_params_lst=None, sigma_b_params_lst=None,
sigma_W=1e-3, tune_sigma_W=True,
sigma_b=1e-6, tune_sigma_b=True,
l_W=1e-6, l_b=1e-6,
diag_noise=True, approx_cols=False,
divide_1st_layer_by_its_n_out=False,
b_out_deterministic=False, seed=None):
assert n_hidden_layers > 0, 'n_layers must be positive'
n_layers = n_hidden_layers + 1
M_lst = [None] * (n_layers) if M_lst is None else M_lst
m_lst = [None] * (n_layers) if m_lst is None else m_lst
if sigma_W_params_lst is None:
sigma_W_params_lst = [None] * (n_layers)
if sigma_b_params_lst is None:
sigma_b_params_lst = [None] * (n_layers)
assert \
len(M_lst) == len(m_lst) == len(sigma_W_params_lst) == \
len(sigma_b_params_lst) == n_layers, \
'length of all lists must be hte same and equal to ' \
'(n_layers + 1) where the +1 is for the output layer mapping'
# set seed to ensure each layer is init differently (cf. seed += 1)
seed = np.random.randint(int(1e6)) if seed is None else seed
np.random.seed(seed)
def activation(x):
return T.nnet.relu(x, alpha=0.1)
self.in_layer = GaussLayer(
input=X, n_in=n_in, n_out=n_units_in,
M=M_lst[0], m=m_lst[0],
sigma_W=sigma_W, tune_sigma_W=tune_sigma_W,
sigma_W_params=sigma_W_params_lst[0],
sigma_b=sigma_b, tune_sigma_b=tune_sigma_b,
sigma_b_params=sigma_b_params_lst[0],
l_W=l_W, l_b=l_b, diag_noise=diag_noise,
activation=activation, approx_cols=approx_cols,
seed=seed, name='h1'
)
self.layers = [self.in_layer]
seed += 1
# specific settings necessary for initialisation of deep GPs
if divide_1st_layer_by_its_n_out:
sqrt_n_out = T.constant(self.in_layer.n_out ** 0.5, dtype=floatX)
self.in_layer.output /= sqrt_n_out
# the first hidden layer was already set up above
for i in xrange(1, n_hidden_layers):
prev_layer = self.layers[-1]
layer = GaussLayer(
input=prev_layer.output,
n_in=prev_layer.n_out, n_out=n_units_hidden,
M=M_lst[i], m=m_lst[i],
sigma_W=sigma_W, tune_sigma_W=tune_sigma_W,
sigma_W_params=sigma_W_params_lst[i],
sigma_b=sigma_b, tune_sigma_b=tune_sigma_b,
sigma_b_params=sigma_b_params_lst[i],
l_W=l_W, l_b=l_b, diag_noise=diag_noise,
activation=activation, name='h' + str(i + 1),
approx_cols=approx_cols, seed=seed
)
self.layers += [layer]
seed += 1
# initialised separately because of the necessary linear activation
prev_layer = self.layers[-1]
self.out_layer = GaussLayer(
input=prev_layer.output, n_in=prev_layer.n_out, n_out=n_out,
M=M_lst[-1], m=m_lst[-1],
sigma_W=sigma_W, tune_sigma_W=tune_sigma_W,
sigma_W_params=sigma_W_params_lst[-1],
sigma_b=sigma_b, tune_sigma_b=tune_sigma_b,
sigma_b_params=sigma_b_params_lst[-1],
l_W=l_W, l_b=l_b, diag_noise=diag_noise,
b_is_deterministic=b_out_deterministic,
approx_cols=approx_cols, name='out', seed=seed
)
self.layers += [self.out_layer]
self.softmax = SoftmaxLayer(
input=self.out_layer.output, name='softmax'
)
self.params = reduce(
lambda x, y: x + y, [layer.grad_params for layer in self.layers]
)
self.input = X
self.p_y_given_x = self.softmax.p_y_given_x
self.y_pred = self.softmax.y_pred
self.mean_log_likelihood = self.softmax.mean_log_likelihood
self.errors = self.softmax.errors
# self.kl_W = T.sum([layer.kl_W() for layer in self.layers])
# self.kl_b = T.sum([layer.kl_b() for layer in self.layers])
# self.kl = self.kl_W + self.kl_b
self.effect_kl_W = T.sum([layer.effect_kl_W() for layer in self.layers])
self.effect_kl_b = T.sum([layer.effect_kl_b() for layer in self.layers])
self.effect_kl = self.effect_kl_W + self.effect_kl_b