|
356
|
1 """
|
|
|
2 One hot Op
|
|
|
3 """
|
|
|
4
|
|
|
5 #from theano import tensor
|
|
|
6 from theano.tensor import as_tensor, Tensor
|
|
|
7 from theano.gof import op
|
|
|
8 from theano.gof.graph import Apply
|
|
|
9
|
|
|
10 import numpy
|
|
|
11
|
|
|
12 class OneHot(op.Op):
|
|
|
13 """
|
|
|
14 Construct a one-hot vector, x out of y.
|
|
|
15
|
|
|
16 @todo: Document inputs and outputs
|
|
|
17 @todo: Use 'bool' as output dtype? Or, at least 'int64' ? Not float64!
|
|
|
18 @todo: Use 'bool' as output dtype, not 'int64' ?
|
|
|
19 @todo: Allow this to operate on column vectors (Tensor)
|
|
|
20 @todo: Describe better.
|
|
|
21 """
|
|
|
22
|
|
|
23 def make_node(self, x, y):
|
|
|
24 """
|
|
|
25 @type x: Vector L{Tensor} of integers
|
|
|
26 @param x: The entries of the one-hot vector to be one.
|
|
|
27 @type y: Integer scalar L{Tensor}
|
|
|
28 @param y: The length (#columns) of the one-hot vectors.
|
|
|
29 @return: A L{Tensor} of one-hot vectors
|
|
|
30
|
|
|
31 @precondition: x < y for all entries of x
|
|
|
32 @todo: Check that x and y are int types
|
|
|
33 """
|
|
|
34 x = as_tensor(x)
|
|
|
35 y = as_tensor(y)
|
|
|
36 #assert x.dtype[0:3] == "int"
|
|
|
37 #assert y.dtype[0:3] == "int"
|
|
|
38 inputs = [x, y]
|
|
|
39 ##outputs = [tensor.Tensor("int64", broadcastable=[False, False])]
|
|
|
40 #outputs = [tensor.Tensor("float64", broadcastable=[False, False])]
|
|
|
41 #outputs = [Tensor("int64", broadcastable=[False, False])]
|
|
|
42 outputs = [Tensor("float64", broadcastable=[False, False]).make_result()]
|
|
|
43 node = Apply(op = self, inputs = inputs, outputs = outputs)
|
|
|
44 return node
|
|
|
45
|
|
|
46 def perform(self, node, (x, y), (out, )):
|
|
|
47 assert x.dtype == "int64" or x.dtype == "int32"
|
|
|
48 assert x.ndim == 1
|
|
|
49 assert y.dtype == "int64" or x.dtype == "int32"
|
|
|
50 assert y.ndim == 0
|
|
|
51 out[0] = numpy.zeros((x.shape[0], y), dtype="float64")
|
|
|
52 for c in range(x.shape[0]):
|
|
|
53 assert x[c] < y
|
|
|
54 out[0][c, x[c]] = 1
|
|
|
55
|
|
|
56 def grad(self, (x, y), (out_gradient, )):
|
|
|
57 return None, None
|
|
|
58 one_hot = OneHot()
|
