def __init__(self, dim, w_param=None, **kwargs):
super(RNNLM, self).__init__(name=generate_name('hidden', 1),
user_type='kHidden')
if w_param == None:
# default: uniform
w_param = Parameter(name=generate_name('w'), init=init)
else:
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
def __init__(self, **kwargs):
super(UserLossRNNLM, self).__init__(name=generate_name('loss', 1),
user_type='kLoss')
self.layer.Extensions[loss_conf].nclass = kwargs['nclass']
self.layer.Extensions[loss_conf].vocab_size = kwargs['vocab_size']
setval(self.layer, param=Parameter(name=generate_name('w'),
init='uniform', scale=0.3).param)
setval(self.layer, param=Parameter(name=generate_name('w', 1),
init='uniform', scale=0.3).param)
def __init__(self, **kwargs):
super(UserLossRNNLM, self).__init__(name=generate_name('loss', 1),
user_type='kLoss')
self.layer.Extensions[loss_conf].nclass = kwargs['nclass']
self.layer.Extensions[loss_conf].vocab_size = kwargs['vocab_size']
setval(self.layer, param=Parameter(name=generate_name('w'),
init='uniform', scale=0.3).param)
setval(self.layer, param=Parameter(name=generate_name('w', 1),
init='uniform', scale=0.3).param)
def __init__(self, dim, w_param=None, **kwargs):
super(RNNLM, self).__init__(name=generate_name('hidden', 1),
user_type='kHidden')
if w_param == None:
# default: uniform
w_param = Parameter(name=generate_name('w'), init=init)
else:
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
def __init__(self, in_dim, out_dim, w_param=None, **kwargs):
super(Embedding, self).__init__(name=generate_name('embedding', 1),
user_type='kEmbedding')
fields = {'vocab_size': in_dim, 'word_dim': out_dim}
setval(self.layer.Extensions[embedding_conf], **fields)
if w_param == None:
# default: uniform
w_param = Parameter(name=generate_name('w'), init=init)
else:
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
def __init__(self, in_dim, out_dim, w_param=None, **kwargs):
super(Embedding, self).__init__(name=generate_name('embedding', 1),
user_type='kEmbedding')
fields = {'vocab_size': in_dim,
'word_dim': out_dim}
setval(self.layer.Extensions[embedding_conf], **fields)
if w_param == None:
# default: uniform
w_param = Parameter(name=generate_name('w'), init=init)
else:
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
def __init__(self, pool_size=None, stride=1, ignore_border=True, **kwargs):
'''
required
pool_size = (int|tuple) // size for pooling
optional
stride = (int) // size of striding
ignore_border = (bool) // flag for padding
**kwargs // fields for Layer class
'''
assert pool_size != None, 'pool_size is required'
if type(pool_size) == int:
pool_size = (pool_size, pool_size)
assert type(pool_size) == tuple and pool_size[0] == pool_size[1], \
'currently pool size should be square in Singa'
super(AvgPooling2D, self).__init__(name=generate_name('pool'),
type=kCPooling,
**kwargs)
self.layer.pooling_conf.pool = PoolingProto().AVG
fields = {
'pool': PoolingProto().AVG,
'kernel': pool_size[0],
'stride': stride,
'pad': 0 if ignore_border else 1
}
setval(self.layer.pooling_conf, **fields)
def __init__(self, pool_size=None,
stride=1, ignore_border=True, **kwargs):
'''
Max Pooling layer
required
pool_size = (int|tuple) // the size for pooling
optional
stride = (int) // the size of striding
ignore_border = (bool) // flag for padding
**kwargs // fields for Layer class
'''
assert pool_size != None, 'pool_size is required'
if type(pool_size) == int:
pool_size = (pool_size, pool_size)
assert type(pool_size) == tuple and pool_size[0] == pool_size[1], \
'pool size should be square in Singa'
super(MaxPooling2D, self).__init__(name=generate_name('pool'),
type=kCPooling, **kwargs)
fields = {'pool' : PoolingProto().MAX,
'kernel' : pool_size[0],
'stride' : stride,
'pad' : 0 if ignore_border else 1}
setval(self.layer.pooling_conf, **fields)
def __init__(self):
'''
'''
self.name = 'accuracy'
self.layer_type = enumLayerType(self.name)
super(Accuracy, self).__init__(name=generate_name(self.name),
type=self.layer_type)
def __init__(self):
'''
'''
self.name = 'accuracy'
self.layer_type = enumLayerType(self.name)
super(Accuracy, self).__init__(name=generate_name(self.name),
type=self.layer_type)
def set_param_field(param, pname, changename=False, withnumber=True, **kwargs):
'''
param = (ParamProto)
pname = (string) // 'w' for wiehgt, or 'b' for bias
changename = (bool) // update parameter name if True
withnumber = (bool) // add layer number if True
**kwargs
w_lr = (float) // learning rate multiplier for weight, used to
// scale the learning rate when updating parameters.
w_wd = (float) // weight decay multiplier for weight, used to
// scale the weight decay when updating parameters.
b_lr = (float) // learning rate multiplier for bias
b_wd = (float) // weight decay multiplier for bias
'''
assert pname == 'w' or pname == 'b', 'pname should be w or b'
lr_ = param.lr_scale
wd_ = param.wd_scale
initkv = {}
if pname == 'w':
if 'w_lr' in kwargs:
lr_ = kwargs['w_lr']
del kwargs['w_lr']
if 'w_wd' in kwargs:
wd_ = kwargs['w_wd']
del kwargs['w_wd']
for key, val in kwargs.items():
if key.startswith('w_'):
initkv[key[2:]] = val
elif pname == 'b':
if 'b_lr' in kwargs:
lr_ = kwargs['b_lr']
del kwargs['b_lr']
if 'b_wd' in kwargs:
wd_ = kwargs['b_wd']
del kwargs['b_wd']
for key, val in kwargs.items():
if key.startswith('b_'):
initkv[key[2:]] = val
field = {'lr_scale': lr_, 'wd_scale': wd_}
# Set/update parameter fields
if param.name.startswith('param') or changename == True:
if 'level' in kwargs: # parameter name for RBM
pname += str(kwargs['level'])
setval(param,
name=generate_name(pname, withnumber=withnumber),
**field)
else:
setval(param, **field)
# Set/update parameter init fields
setval(param.init, **initkv)
def __init__(self, lossname, topk=1, **kwargs):
'''
required
lossname = (string) // softmaxloss, euclideanloss
'''
self.layer_type = enumLayerType(lossname)
super(Loss, self).__init__(name=generate_name(lossname),
type=self.layer_type, **kwargs)
if lossname == 'softmaxloss':
self.layer.softmaxloss_conf.topk = topk
def __init__(self,
load,
phase='train',
checkpoint=None,
conf=None,
**kwargs):
'''
required
load = (string) // type of data
optional
phase = (string) // phase of data layer
checkpoint = (string) // checkpoint path
conf = (Store) // Store object
**kwargs (KEY=VALUE)
partition_dim = (int) // partition dimension for net
'''
assert load != None, 'data type should be specified'
if load == 'kData':
super(Data, self).__init__(name=generate_name('data'),
user_type=load,
**kwargs)
else:
self.layer_type = enumLayerType(load)
super(Data, self).__init__(name=generate_name('data'),
type=self.layer_type,
**kwargs)
self.is_datalayer = True
# include/exclude
setval(self.layer, include=enumPhase(phase))
#setval(self.layer, exclude=kTest if phase=='train' else kTrain)
if conf == None:
if load == 'kData':
setval(self.layer.Extensions[data_conf], **kwargs)
else:
setval(self.layer.store_conf, **kwargs)
else:
setval(self.layer, store_conf=conf.proto)
self.checkpoint = checkpoint # checkpoint for training data
def __init__(self, ratio=0.5):
'''
required
ratio = (float) // ratio of drop out nodes
'''
self.name = 'dropout'
self.layer_type = enumLayerType(self.name)
super(Dropout, self).__init__(name=generate_name(self.name),
type=self.layer_type)
self.layer.dropout_conf.dropout_ratio = ratio
def __init__(self, ratio=0.5):
'''
required
ratio = (float) // ratio of drop out nodes
'''
self.name = 'dropout'
self.layer_type = enumLayerType(self.name)
super(Dropout, self).__init__(name=generate_name(self.name),
type=self.layer_type)
self.layer.dropout_conf.dropout_ratio = ratio
def __init__(self, lossname, topk=1, **kwargs):
'''
required
lossname = (string) // softmaxloss, euclideanloss
'''
self.layer_type = enumLayerType(lossname)
super(Loss, self).__init__(name=generate_name(lossname),
type=self.layer_type,
**kwargs)
if lossname == 'softmaxloss':
self.layer.softmaxloss_conf.topk = topk
def set_param_field(param, pname, changename=False, withnumber=True, **kwargs):
'''
param = (ParamProto)
pname = (string) // 'w' for wiehgt, or 'b' for bias
changename = (bool) // update parameter name if True
withnumber = (bool) // add layer number if True
**kwargs
w_lr = (float) // learning rate multiplier for weight, used to
// scale the learning rate when updating parameters.
w_wd = (float) // weight decay multiplier for weight, used to
// scale the weight decay when updating parameters.
b_lr = (float) // learning rate multiplier for bias
b_wd = (float) // weight decay multiplier for bias
'''
assert pname == 'w' or pname == 'b', 'pname should be w or b'
lr_ = param.lr_scale
wd_ = param.wd_scale
initkv = {}
if pname == 'w':
if 'w_lr' in kwargs:
lr_ = kwargs['w_lr']
del kwargs['w_lr']
if 'w_wd' in kwargs:
wd_ = kwargs['w_wd']
del kwargs['w_wd']
for key, val in kwargs.items():
if key.startswith('w_'):
initkv[key[2:]] = val
elif pname == 'b':
if 'b_lr' in kwargs:
lr_ = kwargs['b_lr']
del kwargs['b_lr']
if 'b_wd' in kwargs:
wd_ = kwargs['b_wd']
del kwargs['b_wd']
for key, val in kwargs.items():
if key.startswith('b_'):
initkv[key[2:]] = val
field = {'lr_scale' : lr_, 'wd_scale' : wd_}
# Set/update parameter fields
if param.name.startswith('param') or changename == True:
if 'level' in kwargs: # parameter name for RBM
pname += str(kwargs['level'])
setval(param, name=generate_name(pname, withnumber=withnumber), **field)
else:
setval(param, **field)
# Set/update parameter init fields
setval(param.init, **initkv)
def __init__(self, meanfile=None, **kwargs):
'''
required
meanfile = (string) // path to meanfile (depreciated)
'''
assert meanfile != None, 'meanfile should be specified'
self.name = 'rgb'
self.layer_type = kRGBImage
super(RGB, self).__init__(name=generate_name(self.name),
type=self.layer_type)
self.layer.rgbimage_conf.meanfile = meanfile
def __init__(self, meanfile=None, **kwargs):
'''
required
meanfile = (string) // path to meanfile (depreciated)
'''
assert meanfile != None, 'meanfile should be specified'
self.name = 'rgb'
self.layer_type = kRGBImage
super(RGB, self).__init__(name=generate_name(self.name),
type=self.layer_type)
self.layer.rgbimage_conf.meanfile = meanfile
def __init__(self, size=0, **kwargs):
'''
required
size = (int) // local size
'''
super(LRN2D, self).__init__(name=generate_name('norm'), type=kLRN)
# required
assert size != 0, 'local size should be set'
self.layer.lrn_conf.local_size = size
init_values = get_init_values('lrn2d', **kwargs)
setval(self.layer.lrn_conf, **init_values)
def __init__(self, size=0, **kwargs):
'''
required
size = (int) // local size
'''
super(LRN2D, self).__init__(name=generate_name('norm'), type=kLRN)
# required
assert size != 0, 'local size should be set'
self.layer.lrn_conf.local_size = size
init_values = get_init_values('lrn2d', **kwargs)
setval(self.layer.lrn_conf, **init_values)
def __init__(self, **kwargs):
'''
**kwargs (KEY=VALUE)
partition_dim = (int) // partition dimension for net
'''
self.layer = Message('Layer', **kwargs).proto
# required
if not 'name' in kwargs:
setval(self.layer, name=generate_name('layer', 1))
# srclayers are set in Model.build()
self.is_datalayer = False
def __init__(self, **kwargs):
'''
**kwargs (KEY=VALUE)
partition_dim = (int) // partition dimension for net
'''
self.layer = Message('Layer', **kwargs).proto
# required
if not 'name' in kwargs:
setval(self.layer, name=generate_name('layer', 1))
# srclayers are set in Model.build()
self.is_datalayer = False
def __init__(self, activation='stanh', topk=1):
'''
required
activation = (string)
optional
topk = (int) // the number of results
'''
self.name = activation
if activation == 'tanh': activation = 'stanh' # <-- better way to set?
self.layer_type = enumLayerType(activation)
super(Activation, self).__init__(name=generate_name(self.name),
type=self.layer_type)
if activation == 'softmaxloss':
self.layer.softmaxloss_conf.topk = topk
def __init__(self, load, phase='train', checkpoint=None,
conf=None, **kwargs):
'''
required
load = (string) // type of data
optional
phase = (string) // phase of data layer
checkpoint = (string) // checkpoint path
conf = (Store) // Store object
**kwargs (KEY=VALUE)
partition_dim = (int) // partition dimension for net
'''
assert load != None, 'data type should be specified'
if load == 'kData':
super(Data, self).__init__(name=generate_name('data'),
user_type=load)
else:
self.layer_type = enumLayerType(load)
super(Data, self).__init__(name=generate_name('data'),
type=self.layer_type)
self.is_datalayer = True
# include/exclude
setval(self.layer, include=enumPhase(phase))
#setval(self.layer, exclude=kTest if phase=='train' else kTrain)
if conf == None:
if load == 'kData':
setval(self.layer.Extensions[data_conf], **kwargs)
else:
setval(self.layer.store_conf, **kwargs)
else:
setval(self.layer, store_conf=conf.proto)
self.checkpoint = checkpoint # checkpoint for training data
def __init__(self, activation='stanh', topk=1):
'''
required
activation = (string)
optional
topk = (int) // the number of results
'''
self.name = activation
if activation == 'tanh': activation = 'stanh' # <-- better way to set?
self.layer_type = enumLayerType(activation)
super(Activation, self).__init__(name=generate_name(self.name),
type=self.layer_type)
if activation == 'softmaxloss':
self.layer.softmaxloss_conf.topk = topk
def __init__(self, nb_filter=0, kernel=0, stride=1, pad=0,
init=None, w_param=None, b_param=None,
activation=None, **kwargs):
'''
required
nb_filter = (int) // the number of filters
kernel = (int) // the size of filter
optional
stride = (int) // the size of stride
pad = (int) // the size of padding
init = (string) // 'unirom', 'gaussian', 'constant'
w_param = (Parameter) // Parameter object for weight
b_param = (Parameter) // Parameter object for bias
**kwargs (KEY=VALUE)
w_lr = (float) // learning rate multiplier for weight, used to
// scale the learning rate when updating parameters.
w_wd = (float) // weight decay multiplier for weight, used to
// scale the weight decay when updating parameters.
b_lr = (float) // learning rate multiplier for bias
b_wd = (float) // weight decay multiplier for bias
'''
assert nb_filter > 0 and kernel > 0, 'should be set as positive int'
super(Convolution2D, self).__init__(name=generate_name('conv', 1),
type=kCConvolution)
fields = {'num_filters' : nb_filter,
'kernel' : kernel,
'stride' : stride,
'pad' : pad}
setval(self.layer.convolution_conf, **fields)
# parameter w
if w_param == None:
self.init = 'gaussian' if init == None else init
w_param = Parameter(init=self.init)
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
# parameter b
if b_param == None:
self.init = 'constant' if init == None else init
b_param = Parameter(init=self.init) # default: constant
set_param_field(b_param.param, 'b', True, **kwargs)
setval(self.layer, param=b_param.param)
# following layers: e.g., activation, dropout, etc.
if activation:
self.mask = Activation(activation=activation).layer
def __init__(self, **kwargs):
'''
optional
**kwargs
name = (string) // parameter name
lr = (float) // learning rate multiplier
wd = (float) // weight decay multiplier
init = (string) // init type {'constant','uniform','gaussian'}
value = (int) // value for 'constant'
scale = (float) // [low=-scale, high=scale] for 'uniform'
low = (float) // low value for 'uniform'
high = (float) // high value for 'uniform'
mean = (float) // mean for 'gaussian'
std = (float) // std for 'gaussian'
'''
fields = {
'lr_scale': kwargs['lr'] if 'lr' in kwargs else 1,
'wd_scale': kwargs['wd'] if 'wd' in kwargs else 1
}
self.param = Message('Param', **fields).proto
if not 'name' in kwargs:
setval(self.param, name=generate_name('param', 1))
else:
pname = kwargs['name']
# parameter name for RBM
if 'level' in kwargs:
pname += str(kwargs['level'])
if pname[0] == 'b':
pname += '2'
setval(self.param, name=pname)
if 'share_from' in kwargs:
setval(self.param, share_from=kwargs['share_from'])
if 'init' in kwargs:
init_values = get_init_values(kwargs['init'], **kwargs)
if not kwargs['init'] == 'none':
pgen = Message('ParamGen',
type=enumInitMethod(kwargs['init']),
**init_values)
del kwargs['init']
setval(self.param, init=pgen.proto)
else: # default: uniform
pgen = Message('ParamGen', type=enumInitMethod('uniform'))
setval(self.param, init=pgen.proto)
def __init__(self, **kwargs):
'''
optional
**kwargs
name = (string) // parameter name
lr = (float) // learning rate multiplier
wd = (float) // weight decay multiplier
init = (string) // init type {'constant','uniform','gaussian'}
value = (int) // value for 'constant'
scale = (float) // [low=-scale, high=scale] for 'uniform'
low = (float) // low value for 'uniform'
high = (float) // high value for 'uniform'
mean = (float) // mean for 'gaussian'
std = (float) // std for 'gaussian'
'''
fields = {'lr_scale' : kwargs['lr'] if 'lr' in kwargs else 1,
'wd_scale' : kwargs['wd'] if 'wd' in kwargs else 1
}
self.param = Message('Param', **fields).proto
if not 'name' in kwargs:
setval(self.param, name=generate_name('param', 1))
else:
pname = kwargs['name']
# parameter name for RBM
if 'level' in kwargs:
pname += str(kwargs['level'])
if pname[0] == 'b':
pname += '2'
setval(self.param, name=pname)
if 'share_from' in kwargs:
setval(self.param, share_from=kwargs['share_from'])
if 'init' in kwargs:
init_values = get_init_values(kwargs['init'], **kwargs)
if not kwargs['init'] == 'none':
pgen = Message('ParamGen', type=enumInitMethod(kwargs['init']),
**init_values)
del kwargs['init']
setval(self.param, init=pgen.proto)
else: # default: uniform
pgen = Message('ParamGen', type=enumInitMethod('uniform'))
setval(self.param, init=pgen.proto)
def __init__(self, **kwargs):
'''
**kwargs (KEY=VALUE)
partition_dim = (int) // partition dimension for net
'''
self.layer = Message('Layer', **kwargs).proto
# required field
if not 'name' in kwargs:
setval(self.layer, name=generate_name('layer', 1))
# layer connectivity is set in Model.build()
self.is_datalayer = False
self.singalayer = None
self.srclayers = []
# set src for Rafiki
if 'src' in kwargs:
self.src = kwargs['src']
else:
self.src = None
def __init__(self, **kwargs):
'''
**kwargs (KEY=VALUE)
partition_dim = (int) // partition dimension for net
'''
self.layer = Message('Layer', **kwargs).proto
# required field
if not 'name' in kwargs:
setval(self.layer, name=generate_name('layer', 1))
# layer connectivity is set in Model.build()
self.is_datalayer = False
self.singalayer = None
self.srclayers = []
# set src for Rafiki
if 'src' in kwargs:
self.src = kwargs['src']
else:
self.src = None
def __init__(self, activation='stanh', **kwargs):
'''
required
activation = (string) // relu, sigmoid, tanh, stanh, softmax.
'''
if activation == 'tanh':
print 'Warning: Tanh layer is not supported for CPU'
self.name = activation
self.layer_type = kActivation
if activation == 'stanh':
self.layer_type = kSTanh
elif activation == 'softmax':
self.layer_type = kSoftmax
super(Activation, self).__init__(name=generate_name(self.name),
type=self.layer_type, **kwargs)
if activation == 'relu':
self.layer.activation_conf.type = RELU
elif activation == 'sigmoid':
self.layer.activation_conf.type = SIGMOID
elif activation == 'tanh':
self.layer.activation_conf.type = TANH # for GPU
def __init__(self, activation='stanh', **kwargs):
'''
required
activation = (string) // relu, sigmoid, tanh, stanh, softmax.
'''
if activation == 'tanh':
print 'Warning: Tanh layer is not supported for CPU'
self.name = activation
self.layer_type = kActivation
if activation == 'stanh':
self.layer_type = kSTanh
elif activation == 'softmax':
self.layer_type = kSoftmax
super(Activation, self).__init__(name=generate_name(self.name),
type=self.layer_type,
**kwargs)
if activation == 'relu':
self.layer.activation_conf.type = RELU
elif activation == 'sigmoid':
self.layer.activation_conf.type = SIGMOID
elif activation == 'tanh':
self.layer.activation_conf.type = TANH # for GPU
def __init__(self, nb_filter=0, kernel=0, stride=1, pad=0,
init=None, w_param=None, b_param=None,
activation=None, **kwargs):
'''
required
nb_filter = (int) // the number of filters
kernel = (int/tuple) // the size of filter
optional
stride = (int/tuple) // the size of stride
pad = (int/tuple) // the size of padding
init = (string) // 'uniform', 'gaussian', 'constant'
w_param = (Parameter) // Parameter object for weight
b_param = (Parameter) // Parameter object for bias
**kwargs (KEY=VALUE)
w_lr = (float) // learning rate multiplier for weight, used to
// scale the learning rate when updating parameters.
w_wd = (float) // weight decay multiplier for weight, used to
// scale the weight decay when updating parameters.
b_lr = (float) // learning rate multiplier for bias
b_wd = (float) // weight decay multiplier for bias
'''
assert nb_filter > 0, 'nb_filter should be set as positive int'
super(Convolution2D, self).__init__(name=generate_name('conv', 1),
type=kCConvolution, **kwargs)
fields = {"num_filters":nb_filter}
# for kernel
if type(kernel) == int:
fields['kernel'] = kernel
else:
fields['kernel_x'] = kernel[0]
fields['kernel_y'] = kernel[1]
# for stride
if type(stride) == int:
fields['stride'] = stride
else:
fields['stride_x'] = stride[0]
fields['stride_y'] = stride[1]
# for pad
if type(pad) == int:
fields['pad'] = pad
else:
fields['pad_x'] = pad[0]
fields['pad_y'] = pad[1]
setval(self.layer.convolution_conf, **fields)
# parameter w
if w_param == None:
self.init = 'gaussian' if init == None else init
w_param = Parameter(init=self.init)
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
# parameter b
if b_param == None:
self.init = 'constant' if init == None else init
b_param = Parameter(init=self.init) # default: constant
set_param_field(b_param.param, 'b', True, **kwargs)
setval(self.layer, param=b_param.param)
# following layers: e.g., activation, dropout, etc.
if activation:
self.mask = Activation(activation=activation).layer
def __init__(self, out_dim=None, w_param=None, b_param=None,
sampling=None, **kwargs):
'''
Generate a set of layers (like MLP) according to the number of elements
in out_dim, and on top of it, two layers RBMVis and RBMHid with
bidirectional connection
The layers are expanded and added in Energy.add()
required
out_dim = (int) or (int list) // the number of hidden nodes
optional
w_param = (Parameter) // Parameter object for weight
b_param = (Parameter) // Parameter object for bias
sampling = (string)
'''
assert out_dim > 0, 'out_dim should be set'
self.out_dim = [out_dim] if type(out_dim) == int else out_dim
self.name = kwargs['name'] if 'name' in kwargs else 'RBMVis'
self.layer_type = kwargs['type'] if 'type' in kwargs else kRBMVis
super(RBM, self).__init__(name=generate_name(self.name,
withnumber=False), type=self.layer_type)
setval(self.layer.rbm_conf, hdim=self.out_dim[-1])
if self.layer_type == kRBMHid and sampling != None:
if sampling == 'gaussian':
setval(self.layer.rbm_conf, gaussian=True)
# parameter w
if w_param == None:
w_param = Parameter(init='gaussian', **kwargs)
set_param_field(w_param.param, 'w', withnumber=False,
level=len(self.out_dim), **kwargs)
else:
if self.layer_type == kRBMHid:
del kwargs['name']
else:
set_param_field(w_param.param, 'w', withnumber=False,
level=len(self.out_dim), **kwargs)
setval(self.layer, param=w_param.param)
# parameter b
if b_param == None:
b_param = Parameter(init='constant', **kwargs)
set_param_field(b_param.param, 'b', withnumber=False,
level=len(self.out_dim), **kwargs)
else:
if self.layer_type == kRBMHid:
pass
else:
set_param_field(b_param.param, 'b', withnumber=False,
level=len(self.out_dim), **kwargs)
setval(self.layer, param=b_param.param)
if self.layer_type == kRBMVis:
wname = w_param.param.name
parw = Parameter(name=wname+"_", init='none', share_from=wname)
bname = b_param.param.name
parb = Parameter(name=bname+"2", wd=0, init='constant')
self.bidirect = RBM(self.out_dim, name='RBMHid', type=kRBMHid,
w_param=parw, b_param=parb, sampling=sampling).layer
def __init__(self,
out_dim=None,
w_param=None,
b_param=None,
sampling=None,
**kwargs):
'''
Generate a set of layers (like MLP) according to the number of elements
in out_dim, and on top of it, two layers RBMVis and RBMHid with
bidirectional connection
The layers are expanded and added in Energy.add()
required
out_dim = (int) or (int list) // the number of hidden nodes
optional
w_param = (Parameter) // Parameter object for weight
b_param = (Parameter) // Parameter object for bias
sampling = (string)
'''
assert out_dim > 0, 'out_dim should be set'
self.out_dim = [out_dim] if type(out_dim) == int else out_dim
self.name = kwargs['name'] if 'name' in kwargs else 'RBMVis'
self.layer_type = kwargs['type'] if 'type' in kwargs else kRBMVis
super(RBM, self).__init__(name=generate_name(self.name,
withnumber=False),
type=self.layer_type)
setval(self.layer.rbm_conf, hdim=self.out_dim[-1])
if self.layer_type == kRBMHid and sampling != None:
if sampling == 'gaussian':
setval(self.layer.rbm_conf, gaussian=True)
# parameter w
if w_param == None:
w_param = Parameter(init='gaussian', **kwargs)
set_param_field(w_param.param,
'w',
withnumber=False,
level=len(self.out_dim),
**kwargs)
else:
if self.layer_type == kRBMHid:
del kwargs['name']
else:
set_param_field(w_param.param,
'w',
withnumber=False,
level=len(self.out_dim),
**kwargs)
setval(self.layer, param=w_param.param)
# parameter b
if b_param == None:
b_param = Parameter(init='constant', **kwargs)
set_param_field(b_param.param,
'b',
withnumber=False,
level=len(self.out_dim),
**kwargs)
else:
if self.layer_type == kRBMHid:
pass
else:
set_param_field(b_param.param,
'b',
withnumber=False,
level=len(self.out_dim),
**kwargs)
setval(self.layer, param=b_param.param)
if self.layer_type == kRBMVis:
wname = w_param.param.name
parw = Parameter(name=wname + "_", init='none', share_from=wname)
bname = b_param.param.name
parb = Parameter(name=bname + "2", wd=0, init='constant')
self.bidirect = RBM(self.out_dim,
name='RBMHid',
type=kRBMHid,
w_param=parw,
b_param=parb,
sampling=sampling).layer
def __init__(self,
nb_filter=0,
kernel=0,
stride=1,
pad=0,
init=None,
w_param=None,
b_param=None,
activation=None,
**kwargs):
'''
required
nb_filter = (int) // the number of filters
kernel = (int/tuple) // the size of filter
optional
stride = (int/tuple) // the size of stride
pad = (int/tuple) // the size of padding
init = (string) // 'uniform', 'gaussian', 'constant'
w_param = (Parameter) // Parameter object for weight
b_param = (Parameter) // Parameter object for bias
**kwargs (KEY=VALUE)
w_lr = (float) // learning rate multiplier for weight, used to
// scale the learning rate when updating parameters.
w_wd = (float) // weight decay multiplier for weight, used to
// scale the weight decay when updating parameters.
b_lr = (float) // learning rate multiplier for bias
b_wd = (float) // weight decay multiplier for bias
'''
assert nb_filter > 0, 'nb_filter should be set as positive int'
super(Convolution2D, self).__init__(name=generate_name('conv', 1),
type=kCConvolution)
fields = {}
# for kernel
if type(kernel) == int:
fields['kernel'] = kernel
else:
fields['kernel_x'] = kernel[0]
fields['kernel_y'] = kernel[1]
# for stride
if type(stride) == int:
fields['stride'] = stride
else:
fields['stride_x'] = stride[0]
fields['stride_y'] = stride[1]
# for pad
if type(pad) == int:
fields['pad'] = pad
else:
fields['pad_x'] = pad[0]
fields['pad_y'] = pad[1]
setval(self.layer.convolution_conf, **fields)
# parameter w
if w_param == None:
self.init = 'gaussian' if init == None else init
w_param = Parameter(init=self.init)
set_param_field(w_param.param, 'w', True, **kwargs)
setval(self.layer, param=w_param.param)
# parameter b
if b_param == None:
self.init = 'constant' if init == None else init
b_param = Parameter(init=self.init) # default: constant
set_param_field(b_param.param, 'b', True, **kwargs)
setval(self.layer, param=b_param.param)
# following layers: e.g., activation, dropout, etc.
if activation:
self.mask = Activation(activation=activation).layer
