CODEWITHSUNDEEP
pythonindexingtensorflowtensor
moreo
moreo

Reputation: 81

Tensorflow - use a tensor as an index

I would like to use the backward cumulative sum function: 

def _backwards_cumsum(x, length, batch_size):

upper_triangular_ones = np.float32(np.triu(np.ones((length, length))))
repeated_tri = np.float32(np.kron(np.eye(batch_size), upper_triangular_ones))
return tf.matmul(repeated_tri,
                  tf.reshape(x, [length, 1]))

However length is a placeholder:

length = tf.placeholder("int32" ,name = 'xx')

So every time it gets a new value and then the calculation of _backwards_cumsum begins.

Once trying to run the function, I got an error:

TypeError: 'Tensor' object cannot be interpreted as an index

The full traceback:

{
TypeError                                 Traceback (most recent call last)
<ipython-input-561-970ae9e96aa1> in <module>()
----> 1 rewards = _backwards_cumsum(tf.reshape(tf.reshape(decays,[-1,1]) * tf.sigmoid(disc_pred_gen_ph), [-1]), _maxx, batch_size)

<ipython-input-546-5c6928fac357> in _backwards_cumsum(x, length, batch_size)
      1 def _backwards_cumsum(x, length, batch_size):
      2 
----> 3     upper_triangular_ones = np.float32(np.triu(np.ones((length, length))))
      4     repeated_tri = np.float32(np.kron(np.eye(batch_size), upper_triangular_ones))
      5     return tf.matmul(repeated_tri,

/Users/onivron/anaconda/envs/tensorflow/lib/python2.7/site-packages/numpy/core/numeric.pyc in ones(shape, dtype, order)
    190 
    191     """
--> 192     a = empty(shape, dtype, order)
    193     multiarray.copyto(a, 1, casting='unsafe')
    194     return a

Where _maxx is the same as length placeholder above.

Any workaround it?

Upvotes: 0

Views: 884

Answers (1)

Ishant Mrinal
Ishant Mrinal

Reputation: 4918

The error is related to tensor object that you are unknowingly using for numpy array:length. The best way to use numpy functionality within tensorflow is to use tf.py_func. 

# Define a new function that only depends on numpy/any non tensorflow graph object

def get_repeated_tri(length, batch_size):
    upper_triangular_ones = np.float32(np.triu(np.ones((length, length))))
    repeated_tri = np.float32(np.kron(np.eye(batch_size), upper_triangular_ones))
    return repeated_tri
# Here length and batch size must be non tensor object
repeated_tri = tf.py_func(get_repeated_tri, [length, batch_size], tf.int32)
# there're some size mismacthes also in your code `tf.matmul`
def _backwards_cumsum(repeated_tri, x, length_, batch_size):
    return tf.matmul(repeated_tri, tf.reshape(x, [length_*batch_size, -1]))
length_ = tf.placeholder(tf.int32, name='length')
# also define length, batch_size as nump constants
# x as tensorflow tensor
some_tensor_out= _backwards_cumsum(repeated_tri, x, length_, batch_size)

some_tensor_out_ = sess.run(some_tensor_out, {length_:length})

Upvotes: 1

Related Questions