2

I am writing a simple code to compute one-hot encoding for a list of indices. Eg: [1,2,3] => [[0,1,0,0],[0,0,1,0],[0,0,0,1]]

I write a function to do the same for a single vector:

n_val =4
def encoding(x_t):
    z = T.zeros((x_t.shape[0], n_val))
    one_hot = T.set_subtensor(z[T.arange(x_t.shape[0]), x_t], 1)
    return one_hot

To repeat the same function over the rows of a matrix, I do the following,

x = T.imatrix()
[m],_ = theano.scan(fn = encoding, sequences = x)

Y = T.stacklists(m)
f= theano.function([x],Y)

I am expecting a 3D tensor with each slice corresponding to the one-hot encoding of the rows of the matrix.

I am getting the following error while compiling the function,

/Library/Python/2.7/site-packages/theano/tensor/var.pyc in __iter__(self)
594         except TypeError:
595             # This prevents accidental iteration via builtin.sum(self)
--> 596             raise TypeError(('TensorType does not support iteration. '
    597                              'Maybe you are using builtin.sum instead of '
598                              'theano.tensor.sum? (Maybe .max?)'))

TypeError: TensorType does not support iteration. Maybe you are using builtin.sum instead of theano.tensor.sum? (Maybe .max?)

Can someone please help me understand where I am going wrong and how I can modify the code to obtain what I need?

Thanks in advance.

Prasanna Parthasarathi
  • 51
  • 1
  • 4

1 Answers1

2

Here is the code that work 

# input a matrix, expect scan to work with each row of matrix
my_matrix = np.asarray([[1,2,3],[1,3,2],[1,1,1]])

x = T.imatrix()

def encoding(idx):
    z = theano.tensor.zeros((idx.shape[0], 4))
    one_hot = theano.tensor.set_subtensor(z[theano.tensor.arange(idx.shape[0]), idx], 1)
    return one_hot

m, update = theano.scan(fn=encoding,
                        sequences=x)


f = theano.function([x], m)

##########3
result = f(my_matrix)
print (result)
Haha TTpro
  • 5,137
  • 6
  • 45
  • 71