def _gen(n, comp_c_f, struct_f=bintree):
dims = 32
lbls = 5
unk = '<UNK>'
sents = tuple(tuple(s) for s in islice(sentgen(min_len=2), n))
vocab = set(chain.from_iterable(sents))
dic = OrderedDict()
for tok in vocab:
dic[tok] = random_uniform(dims)
dic[unk] = random_uniform(dims)
Source = keyed_source_vertex(dims, dic, missing_=unk)
Class = softmax_vertex(lbls, dims)
Comp = comp_c_f(dims, 2)
Model = net_model((
Source,
Comp,
Class,
))
def class_c():
return Class(target=onehot(lbls))
nets = []
for sent in sents:
net = struct_f(Comp, (Source(tok) for tok in sent))
decorate(net, class_c, internal_prob=0.2)
nets.append(net)
return (tuple(nets), Model, Comp)
def softmax():
dims = 2
lbls = 2
Source = static_source_vertex(dims)
Class = softmax_vertex(lbls, 2 * dims)
Model = net_model((Class, ))
a = Source(random_uniform(dims))
b = Source(random_uniform(dims))
c = Class(target=array((
1,
0,
)).reshape(-1, 1))
net = Net()
net.add_edge(a, c)
net.add_edge(b, c)
model = Model()
return (
model,
net,
)
def rnn():
dims = 2
lbls = 2
dic = OrderedDict((
(
'a',
random_uniform(dims),
),
(
'b',
random_uniform(dims),
),
(
'<UNK>',
random_uniform(dims),
),
))
Source = keyed_source_vertex(dims, dic, missing_='<UNK>')
Comp = rnn_vertex(dims, 2)
Class = softmax_vertex(lbls, dims)
Model = net_model((
Source,
Comp,
Class,
))
a = Source('a')
b = Source('b')
c = Comp()
d = Class(target=array((
1,
0,
)).reshape(-1, 1))
net = Net()
net.add_edge(a, c)
net.add_edge(b, c)
net.add_edge(c, d)
model = Model()
return (
model,
net,
)
def pickle_check():
dims = 32
vertex_classes = (
static_source_vertex(dims),
# Note: This is not how the KeyedSourceVertex is used.
keyed_source_vertex(dims, {'a': 'foo'}),
softmax_vertex(dims, dims),
rnn_vertex(dims, 2),
)
for vertex_class in vertex_classes:
dumps(vertex_class)
Model = net_model(vertex_classes)
dumps(Model)