PyTorch-LBFGS icon indicating copy to clipboard operation
PyTorch-LBFGS copied to clipboard

Encountering RuntimeError in CUTEst Problem: INDEF

Open hjmshi opened this issue 6 years ago • 8 comments

Dheevatsa, have you seen any errors of the following sort?

Traceback (most recent call last):
  File "lbfgs_tests.py", line 114, in <module>
    obj, grad, lr, backtracks, clos_evals, grad_evals, fail = optimizer.step(p, grad, options=options)
  File "../functions/LBFGS.py", line 720, in step
    self._add_update(t - t_prev, d)
  File "../functions/LBFGS.py", line 251, in _add_update
    p.data.add_(step_size, update[offset:offset + numel].view_as(p.data))
RuntimeError: Overflow when unpacking long

I observed this when testing the L-BFGS implementation on the CUTEst problems. I found this from a quick Google search: https://github.com/pytorch/pytorch/issues/6339, but my version of PyTorch is 0.4.1...do you know why this might occur?

hjmshi avatar Sep 15 '18 21:09 hjmshi

Let me try at my end with my version of pytorch, what was the specific run command that you used - so that I can recreate the problem ?

dmudiger avatar Sep 18 '18 23:09 dmudiger

Have you installed PyCUTEst? If so, then you should be able to run python lbfgs_tests.py --problemName 'INDEF'.

I'm currently looking at CUTEst problems with Jorge. It may take some time but it will make sure that our implementation is fool-proof.

hjmshi avatar Sep 18 '18 23:09 hjmshi

Setting it up now, following this - https://gist.github.com/antonior92/9fd6ac9f44c164ee42555e3a7dc5a332 Anything i need to do differently ?

dmudiger avatar Sep 18 '18 23:09 dmudiger

Check out this: https://jfowkes.github.io/pycutest/_build/html/index.html.

hjmshi avatar Sep 18 '18 23:09 hjmshi

I was able to get CUTEst and pycutest both setup (on my mac), but still running into a segfault ?? pytorch) dheevatsa-mbp:examples dheevatsa$ python lbfgs_tests.py --problemName 'INDEF' Parameters available for problem INDEF: N = 10 (int) N = 50 (int) N = 100 (int) N = 1000 (int, original value) N = 5000 (int) [default] ALPHA = 0.5 (float, indefiniteness) [default] ALPHA = 1 (float) ALPHA = 10 (float) ALPHA = 100 (float) ALPHA = 1000 (float) End of parameters for problem INDEF {'objective': 'O', 'constraints': 'U', 'regular': True, 'degree': 2, 'origin': 'A', 'internal': False, 'n': None, 'm': 0} Segmentation fault: 11 (pytorch) dheevatsa-mbp:examples dheevatsa$

dmudiger avatar Sep 19 '18 01:09 dmudiger

seems like this is also the macOS version problem, what did you do to get around it ?

dmudiger avatar Sep 19 '18 01:09 dmudiger

I didn't, I wound up using Linux. Could you run it on a Linux machine?

Interesting though that you are encountering a similar issue...

hjmshi avatar Sep 19 '18 02:09 hjmshi

请问了解这个测试库的matlab接口吗?

lwn66 avatar Jun 08 '22 09:06 lwn66