def __init__(self,
inputs,
bs,
max_time,
classes,
feature_dim,
hidden_size,
method='max',
seed=12345):
self._inputs = inputs
self.method = method
self.batch_size = bs
self.classes = classes
self.max_time = max_time
self.feature_dim = feature_dim
self.dropout = True
self.hidden = HiddenLayer(input_size=feature_dim,
hidden_size=hidden_size,
batch_size=bs,
name='hidden',
dropout=0.5,
activation=act.LeakyRelu())
self.softmax = SoftmaxLayer(input_size=hidden_size,
classes=self.classes,
batch_size=bs,
name='softmax',
dropout=0.5)
def __init__(self,
inputs,
bs,
max_time,
class_num,
feature_dim,
hidden_size,
method='max',
seed=12345):
self._inputs = inputs
self.method = method
self.name = 'baseline_' + str(class_num)
self.batch_size = bs
self.class_num = theano.shared(class_num)
self.max_time = max_time
self.feature_dim = feature_dim
self.dropout = True
self.hidden = HiddenLayer(input_size=feature_dim,
hidden_size=hidden_size,
batch_size=bs,
name='hidden',
dropout=0.5,
activation=act.LeakyRelu())
self.classify = HiddenLayer(input_size=hidden_size,
hidden_size=1,
batch_size=bs,
name='classify',
dropout=0.0,
activation=act.sigmoid)
def __init__(self,
inputs,
bs,
max_time,
class_num,
feature_dim,
hidden_size,
levels,
N=1,
pool=None,
seed=12345):
self._inputs = inputs
self.N = N
self.batch_size = bs
self.name = 'learned_' + str(class_num)
self.class_num = theano.shared(class_num)
self.max_time = max_time
self.filters = levels
self.feature_dim = feature_dim
self.pool = pool
self.dropout = True
self.temporal_pyramid = []
for f in range(self.filters):
tf = TemporalAttentionLayer(batch_size=bs,
N=N,
channels=feature_dim,
name='af-' + str(f))
self.temporal_pyramid.append(tf)
input_size = feature_dim * len(self.temporal_pyramid) #*N
self.hidden = HiddenLayer(input_size=input_size,
hidden_size=hidden_size,
activation=act.LeakyRelu(),
batch_size=bs,
name='hidden',
dropout=0.5)
self.classify = HiddenLayer(input_size=hidden_size,
hidden_size=1,
batch_size=bs,
name='classify',
dropout=0.0,
activation=act.sigmoid)
def __init__(self, inputs, bs, max_time, classes, feature_dim, hidden_size, levels, N=1, pool=None, seed=12345):
self._inputs = inputs
self.N = N
self.batch_size = bs
self.classes = classes
self.max_time = max_time
self.levels = levels
self.feature_dim = feature_dim
self.pool = pool
self.dropout = True
# create a pyramid of filters
self.temporal_pyramid = []
for l in range(self.levels):
for f in range(2**l):
tf = TemporalAttentionLayer(batch_size=bs, N=N, channels=feature_dim,
name='temporal-attention-layer-'+str(l)+'-filter-'+str(f))
tf.test = True
tf.d = theano.shared(value=np.asarray([1./2**(l+1)]).astype('float32'), name='d', borrow=True,
broadcastable=[True])
tf.g = theano.shared(value=np.asarray([((1./2**l)+(2*f/2.**l))]).astype('float32'), name='g', borrow=True,
broadcastable=[True])
tf.sigma = theano.shared(value=np.asarray([5.0]).astype('float32'), name='sigma', borrow=True,
broadcastable=[True])
self.temporal_pyramid.append(tf)
input_size = feature_dim*N*(len(self.temporal_pyramid) if pool == None else 1)
self.hidden = HiddenLayer(input_size=input_size, hidden_size=hidden_size, activation=act.LeakyRelu(),
batch_size=bs, name='hidden', dropout=0.5)
self.softmax = SoftmaxLayer(input_size=hidden_size, classes=self.classes,
batch_size=bs, name='softmax', dropout=0.5)