Don't downcast shape dimensions to int32 in random. about theano HOT 12 CLOSED

Use int64 in that case. This is what we use for the infer_shape.

For safety, we should use run Theano tests on windows for that PR.

from theano.

For safety, test on windows and with a 32-bit python

from theano.

This issue should be closed as we don't do this anymore.

from theano.

Why did you close this? We still do this. Just grep the file tensor/raw_random.py for int32.

from theano.

@nouiz is right, we are still doing that for the provided shape argument (not for the infer_shape mechanism).

from theano.

The cast there https://github.com/Theano/Theano/blob/master/theano/tensor/raw_random.py#L187 is only done when shape == ().

Unless you mean something else.

from theano.

• Theano/theano/tensor/raw_random.py

  Line 397 in 575c7a9

  return ndim, tensor.cast(v_shape, 'int32'), tuple(bcast)

• Theano/theano/tensor/raw_random.py

  Line 521 in 575c7a9

  n = tensor.cast(n, 'int32')

from theano.

Oh, ok.

from theano.

Actually, we also do that in theano/sandbox/rng_mrg.py (l. 334, l. 542, l. 784), and we assert that the dtype is int32 in the C code (l. 425, l. 674, l. 932).
I'm not sure what we should do for that case.

from theano.

So, for rng_mrg.py, here is what we should do, at all 3 places:

• In the graph, make sure shapes are int64.
• In the C code, expect int64, and check the values when setting odims[i] to make sure it fits in int32, raise a ValueError if it does not.

@yaoli this is your ticket.

from theano.

I'd appreciate if someone has a chat with me of this problem as I'm not sure if I understand it correctly.

from theano.

@yaoli We can chat on google hangout

from theano.