a problem about slicing the tensor

[Moonet]

Hi,

I don't know why this bug happened, it seems like tensor cannot be sliced by M[:, slice_num*l:(slice_num+1)*l]

File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/sequence_labeling_experiment.py", line 300, in run_experiment sequencelabeler = SequenceLabeler(config) File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/sequence_labeler.py", line 42, in init char_output_tensor = recurrence.create_birnn(char_input_tensor, config["char_embedding_size"], char_mask_reshaped, config["char_recurrent_size"], return_combined=True, fn_create_parameter_matrix=self.create_parameter_matrix, name="char_birnn") File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/recurrence.py", line 12, in create_birnn recurrent_size, only_return_final=return_combined, go_backwards=False, fn_create_parameter_matrix=fn_create_parameter_matrix, name=name + "_forward") File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/recurrence.py", line 74, in create_lstm go_backwards=go_backwards) File "/Users/Xin/anaconda/lib/python3.5/site-packages/theano/scan_module/scan.py", line 773, in scan condition, outputs, updates = scan_utils.get_updates_and_outputs(fn(args)) File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/recurrence.py", line 34, in lstm_mask_step h_new, c_new = lstm_step(x, h_prev, c_prev, W_x, W_h, b, W_ci, W_cf, W_co) File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/recurrence.py", line 26, in lstm_step i = tensor.nnet.sigmoid(_slice(m_xhb, 0, 4) + c_prev W_ci) File "/Users/Xin/Documents/17_Spring/sequence-labeler-master/recurrence.py", line 47, in _slice return M[:, slice_numl:(slice_num+1)l] File "/Users/Xin/anaconda/lib/python3.5/site-packages/theano/tensor/var.py", line 519, in getitem theano.tensor.subtensor.Subtensor.convert(arg) File "/Users/Xin/anaconda/lib/python3.5/site-packages/theano/tensor/subtensor.py", line 370, in convert slice_a = Subtensor.convert(a, False) File "/Users/Xin/anaconda/lib/python3.5/site-packages/theano/tensor/subtensor.py", line 349, in convert raise TypeError("Expected an integer") TypeError: Expected an integer

[marekrei]

In your error message it says "slice_numl" instead of "slice_num*l". Have you edited the code?

[Moonet]

Well, sorry about that, because markdown style omits the , actually, my code has the

[marekrei]

Not sure what might be causing this then. Works fine on my machines. It even crashes at constructing the computation graph, before processing any data. Perhaps you're using an old version of Theano?

[Moonet]

I am using Theano 0.9.0. m_xhb's type I checked under debug mode is TensorType(float64,scalar) belonging to invalid_tensor_types in subtensor.py. I don't know why this happened

[marekrei]

Right. Seems that I'm the one using the old Theano. When I was developing this code, 0.8.2 was still the default version. I just tried with Theano 0.9 and it actually failed already on importing Lasagne. Apparently the stable version of Lasagne (0.1) doesn't support the latest stable version of Theano (0.9). That's unfortunate. I will find some time to update the code to work with Theano 0.9, possibly looking into options for removing the Lasagne dependency. Until then, using Theano 0.8.* is probably the best bet.

[Moonet]

Yes, thank you, and it seems like "lasagne 0.2.dev1" is compatible with Theano 0.9. Maybe you can try updating the development version and find out whether there is a bug like mine.

[Moonet]

I have tried running my codes under Theano 0.8.2 and Lasagne (0.1) and same bugs happened. Maybe it is my laptop's problem and I will try the code on another laptop. Anyway, thanks for your help.

[Moonet]

The code ran on my friend's laptop also has this bug, now I don't quite understand why.

[marekrei]

What version of the code are you using? Your error message says it comes from line 47 in recurrence.py, but that line is different from the version that's online. Can you try a clean unmodified version of the code, with Theano 0.8.2 and Lasagne 0.1? Preferably on a Linux/Ubuntu machine.

[Moonet]

I use the newest version and the difference happened is because I add two lines to print out some variable which I think it doesn't matter. Actually, I tried a clean version of the code on Mac OS system using python 3.5 and it didn't work.

[marekrei]

I haven't tested it with python3 really. Can you try if it works with python 2.7?

[chrisjbryant]

Been doing some testing of my own with Python3 and came across the same error. It seems the bug was caused by a difference in division operators between python2 and python3. Changing all / to // in the _slice function in recurrence.py seemed to solve the problem. See https://docs.scipy.org/doc/numpy/reference/generated/numpy.true_divide.html

[marekrei]

This should be solved now - the code can run in python3.