wrong loss returned by ctc_cost()

[david-leon]

@mcf06 I compared the returned average loss by ctc_cost() and my own theano-based version, and the results are never consistent. The following are the result examples, loss1 is the result of my own theano version, and loss2 is the result from ctc_cost()

loss1 = 33.0587, loss2 = 32.9557
loss1 = 30.6853, loss2 = 32.0488
loss1 = 31.6061, loss2 = 33.2019
loss1 = 31.8931, loss2 = 31.1848
loss1 = 31.2144, loss2 = 31.2196
loss1 = 33.4684, loss2 = 35.6380
loss1 = 29.9425, loss2 = 32.2495
loss1 = 30.3169, loss2 = 32.5860
loss1 = 30.8735, loss2 = 31.2624
loss1 = 32.9855, loss2 = 32.8420

The way I use ctc_cost() is as

import theano.tensor as tensor
sm_tensor  = tensor.ftensor3()    # activation output, before softmax
seq_tensor_i = tensor.imatrix()   # target label sequence
loss_tensor2 = ctc_cost(sm_tensor.dimshuffle(1, 0, 2), seq_tensor_i)
f2 = theano.function([seq_tensor_i, sm_tensor], loss_tensor2)

B = 1                  # batch size
T = 20                 # time steps
Np1 = 11               # class number plus 1
L = 8                  # label length
x = np.random.rand(B, T, Np1).astype('float32')
y = np.random.randint(0, Np1-1, size=(B, L)) + 1
loss2 = np.mean(f2(y.astype('int32'), x))

Is there any place I've done wrongly?