|
356
|
1 """
|
|
|
2 One hot Op
|
|
|
3 """
|
|
|
4
|
|
|
5 #from theano import tensor
|
|
|
6 from theano.tensor import as_tensor, Tensor
|
|
|
7 #from theano import scalar
|
|
|
8 from theano.scalar import as_scalar
|
|
|
9 from theano.gof import op
|
|
|
10 from theano.gof.graph import Apply
|
|
|
11
|
|
|
12 import numpy
|
|
|
13
|
|
|
14 class OneHot(op.Op):
|
|
|
15 """
|
|
|
16 Construct a one-hot vector, x out of y.
|
|
|
17
|
|
|
18 @todo: Document inputs and outputs
|
|
|
19 @todo: Use 'bool' as output dtype? Or, at least 'int64' ? Not float64!
|
|
|
20 @todo: Use 'bool' as output dtype, not 'int64' ?
|
|
|
21 @todo: Allow this to operate on column vectors (Tensor)
|
|
|
22 @todo: Describe better.
|
|
|
23 @todo: What type is y?
|
|
|
24 @todo: What about operating on L{Scalar}s?
|
|
|
25 """
|
|
|
26
|
|
|
27 def make_node(self, x, y):
|
|
|
28 """
|
|
|
29 @type x: Vector L{Tensor} of integers
|
|
|
30 @param x: The entries of the one-hot vector to be one.
|
|
|
31 @type y: Integer L{Scalar}
|
|
|
32 @param y: The length (#columns) of the one-hot vectors.
|
|
|
33 @return: A L{Tensor} of one-hot vectors
|
|
|
34
|
|
|
35 @precondition: x < y for all entries of x
|
|
|
36 @todo: Check that x and y are int types
|
|
|
37 """
|
|
|
38 #x = tensor.as_tensor(x)
|
|
|
39 #y = scalar.as_scalar(y)
|
|
|
40 x = as_tensor(x)
|
|
|
41 y = as_scalar(y)
|
|
|
42 #assert x.dtype[0:3] == "int"
|
|
|
43 #assert y.dtype[0:3] == "int"
|
|
|
44 inputs = [x, y]
|
|
|
45 ##outputs = [tensor.Tensor("int64", broadcastable=[False, False])]
|
|
|
46 #outputs = [tensor.Tensor("float64", broadcastable=[False, False])]
|
|
|
47 #outputs = [Tensor("int64", broadcastable=[False, False])]
|
|
|
48 outputs = [Tensor("float64", broadcastable=[False, False]).make_result()]
|
|
|
49 node = Apply(op = self, inputs = inputs, outputs = outputs)
|
|
|
50 return node
|
|
|
51
|
|
|
52 def perform(self, node, (x, y), (out, )):
|
|
|
53 assert x.dtype == "int64"
|
|
|
54 assert type(y) == numpy.int64
|
|
|
55 assert x.ndim == 1
|
|
|
56 #out = numpy.zeros((x.shape[0], y), dtype="int64")
|
|
|
57 out[0] = numpy.zeros((x.shape[0], y), dtype="float64")
|
|
|
58 for c in range(x.shape[0]):
|
|
|
59 assert x[c] < y
|
|
|
60 out[0][c, x[c]] = 1
|
|
|
61
|
|
|
62 def grad(self, (x, y), (out_gradient, )):
|
|
|
63 return None, None
|
|
|
64 one_hot = OneHot()
|
