Error messages to improve

[apaszke]

• THNN errors (say exactly which function has failed) - depends on the C API
• double backward without save_variables=True
• accept int as real for float types
• constructing variables with non-tensor objects
• torch.cat prints TODO: torch.cat(torch.Tensor(128), torch.Tensor(128))
• CUDA OOM when constructing tensors says the arguments were invalid
• out of range errors
• inconsistent tensor size
• use of variables in torch.xxx APIs expecting tensors
• Convolution with invalid input sizes
• torch.cat (on variables) when given a.cat((b), 1) # a and b are 2D
• torch.cat when given different tensor types. torch.cat((a,b)) where a is LongTensor and b is FloatTensor for example.
• when autograd.Function.forward returns something other than a tensor or tuple(tensor)
• when MyFunction.backward() takes the incorrect number of arguments
• addbmm output tensor of incorrect size at /Users/soumith/code/pytorch/torch/lib/TH/generic/THTensorMath.c:1040

RuntimeError: out of range at /Users/soumith/anaconda/conda-bld/pytorch-0.1.4_1475478983079/work/torch/lib/TH/generic/THTensor.c:350

[soumith]

reported by @y0ast

When calling the Linear module with a Tensor with higher than 2D the error is: RuntimeError: matrices expected, got 3D, 2D tensors. It was quite confusing to me, maybe we can change to RuntimeError: matrices expected, got 3D tensors and a matrix or even better RuntimeError: 2D tensors expected, got 3D, 2D tensors

[Atcold]

The following error is hard to parse.

TypeError: FloatClassNLLCriterion_updateOutput received an invalid combination of arguments - got (int, torch.FloatTensor, torch.FloatTensor, torch.FloatTensor, bool, NoneType, torch.FloatTensor), but expected (int state, torch.FloatTensor input, torch.LongTensor target, torch.FloatTensor output, bool sizeAverage, [torch.FloatTensor weights or None], torch.FloatTensor total_weight)

It would be nicer to see something like "target is expected to be a LongTensor, got FloatTensor instead". Moreover, I couldn't figure out where the error has been thrown (which assert does this).

[lwolfsonkin]

Consider the following error:

In [64]: Em
Out[64]: 
Variable containing:
 0.3989  0.1153  0.2762
 0.3676  0.6075  0.9249
 0.8584  0.1807  0.9608
 0.7265  0.8045  0.1511
 0.9004  0.1337  0.0259
 0.7794  0.2576  0.1897
 0.6640  0.4450  0.0762
 0.0892  0.0540  0.7068
[torch.FloatTensor of size 8x3]

In [65]: Em[0]
Out[65]: 
Variable containing:
 0.3989
 0.1153
 0.2762
[torch.FloatTensor of size 3]

In [66]: Em[0].t()
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-66-ddcc6c77b14a> in <module>()
----> 1 Em[0].t()

/usr/local/lib/python3.6/site-packages/torch/autograd/variable.py in t(self)
    663 
    664     def t(self):
--> 665         return Transpose(0, 1)(self)
    666 
    667     def transpose(self, dim1, dim2):

/usr/local/lib/python3.6/site-packages/torch/autograd/_functions/tensor.py in forward(self, i)
     77 
     78     def forward(self, i):
---> 79         result = i.transpose(*self.dims)
     80         self.mark_shared_storage((i, result))
     81         return result

RuntimeError: out of range at /Users/soumith/code/pytorch-builder/wheel/pytorch-src/torch/lib/TH/generic/THTensor.c:408

Trying to do transpose on a vector (1 dimensional) gives an unintelligible error when that vector is wrapped in Variable.

RuntimeError: out of range at /Users/soumith/code/pytorch-builder/wheel/pytorch-src/torch/lib/TH/generic/THTensor.c:408

The equivalent operation on a Tensor returns:

RuntimeError: t() expects a 2D tensor, but self is 1D

This is an example of the last member of the above checklist.

[soumith]

std::bad_cast when running https://gist.github.com/panovr/2977d9f26866b05583b0c40d88a315bf instead of a proper error message.
As posted here: https://discuss.pytorch.org/t/runtimeerror-std-bad-cast-trying-to-fine-tune-resnet18/1592

[hughperkins]

Just generally, I think that all c errors should be caught, and converted to something more meaningful, at the python level. The c errors are good for dev debugging of the c libraries, but are really a pain when trying to write actual models. Another error, which might or might not be included above (hard to tell):

Traceback (most recent call last):
  File "seq2seq_attention.py", line 116, in <module>
    state, enc_loss = encode(input_encoded, state)
  File "seq2seq_attention.py", line 105, in encode
    enc_loss += criterion(pred_c, autograd.Variable(torch.LongTensor([input_c_encoded.item()])))
  File "/Users/hugh2/conda3/envs/pytorch/lib/python3.6/site-packages/torch/nn/modules/module.py", line 206, in __call__
    result = self.forward(*input, **kwargs)
  File "/Users/hugh2/conda3/envs/pytorch/lib/python3.6/site-packages/torch/nn/modules/loss.py", line 36, in forward
    return backend_fn(self.size_average, weight=self.weight)(input, target)
  File "/Users/hugh2/conda3/envs/pytorch/lib/python3.6/site-packages/torch/nn/_functions/thnn/auto.py", line 41, in forward
    output, *self.additional_args)
RuntimeError: Assertion `THIndexTensor_(size)(target, 0) == batch_size' failed.  at /Users/soumith/miniconda2/conda-bld/pytorch_1493757886437/work/torch/lib/THNN/generic/ClassNLLCriterion.c:50

There are a few issues really:

• it's hard to search through the stack trace to figure out which line of one's one code is the last thing before it goes into torch code
  • solution: cut the stack trace off wherever it passes into the torch library (in this case, cut everything from module.py line 206 forards off
  • provide an option to display this stuff anywyay, like torch.StackTraces.show_all(True), or whatever (no opinoin on the exact command, it's pytorch-dev only mostly anyway)
• it's hard to figure out which tensors are at fault. eg you pass in three tensors to a torch python method, which ones are mismatching etc?
  • solution: somehow map the error message to the correct tensor. I realize this is generally really hard/impossible. But I'm at least making the observation that it would be nice, if someone can figure out how to make it possible :-)
• doesnt show the actual and expected sizes of the tensors going in. Like, 'When you called foobar(spam, eggs, chips), len(spam[:,0,:]) was 10, and len(eggs(0, :, :) was 5, and they should be the same', or something
  • solution: at least print out the sizes of the relevant tensors etc, even from the c code would be nice, rather than just calling assert

One challenge is that it's hard to tell where is the entry point into the pytorch library from someones code: an entrypoint in one persons code might be used by the entry point funcgtion in another persons code. and putting exception stuff all over the place probably wont be graet for performance. One option that occurs to me, just after finishing writing the above, could be to make some kind of 'prettifier', that takes all the stack trace output, and automatically beuatifies it a bit. So, a user can just call print(prettify(some_exception)) or similar

• obviously this would need all required information to be present somehow/somewhere in the stack trace. eg the c code should print out the tensor dimensions etc, in some easy-to-parse format, somewhere
• however it does have a few advantages:
  • could be a separate project from core, could even be pluggable, so doesnt involve changing core code (beyond making sure core code dumps all required information somewhere, during an exception)
  • wont have any performance impact etc, unless someone explicitly calls prettify

Note that I'm not really thinking of actually contributing to any of these fixes :-P . Just logging it here for completeness. I might occasionally add one single error message or something, though it'd be nice to somehow decide on /document a standard framework/pattern for doing this first perhaps.

[hughperkins]

Its in the list above, but I think it could be easier to debug inconsistent tensor size if it stated the size of the two tensors, to save having to add prints to the program to get this information:

RuntimeError: inconsistent tensor size at /Users/soumith/miniconda2/conda-
bld/pytorch_1493757886437/work/torch/lib/TH/generic/THTensorMath.c:831

[soumith]

@hughperkins "inconsistent tensor size" is fixed in master to be much better.

[Stonesjtu]

std::bad_cast is still not fixed in the newest master. Is anyone working on this?

[soumith]

@Stonesjtu in what context did you get std::bad_cast? Happy to fix, but need some context.

[Stonesjtu]

env

version: 0.2.0+bcea678(build from source)
CUDA: 8.0
CUDNN: 6

bn = nn.BatchNorm1d(100)
input = Variable(torch.randn(10,100).cuda())

bn(input) # produce std::bad_cast

bn.cuda()
bn(input) # works fine

[soumith]

@alexholdenmiller reports:

shoot I had just tried torch.gather(source, 1, batch_of_indices), and got the error "Input tensor must have same dimensions as output tensor", which gave me the impression that gather wouldn't work.

Turns out this error meant to communicate the issue "Input tensor must have same number of dimensions as index tensor", as the working command was torch.gather(source, 1, batch_of_indices.view(-1, 1))

[Stonesjtu]

Something wrong when doing mixed advanced indexing. The error message is not helping in this case.

In [1]: import torch
from torch.nn        
In [2]: from torch.autograd import Variable

In [3]: va = Variable(torch.rand(5,5))

In [4]: index = torch.LongTensor([1,2,3])

In [5]: va[1,index]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-5-c87759241b4e> in <module>()
----> 1 va[1,index]

/slwork/users/kys10/anaconda2/lib/python2.7/site-packages/torch/autograd/variable.pyc in __getitem__(self, key)
     73                 return IndexSelect.apply(self, 0, key)
     74             # else fall through and raise an error in Index
---> 75         return Index.apply(self, key)
     76 
     77     def __setitem__(self, key, value):

/slwork/users/kys10/anaconda2/lib/python2.7/site-packages/torch/autograd/_functions/tensor.pyc in forward(ctx, i, index)
     14         ctx.input_size = i.size()
     15         ctx.index = index
---> 16         result = i.index(ctx.index)
     17         ctx.advanced_indexing = i._check_advanced_indexing(index)
     18         if not ctx.advanced_indexing:

TypeError: indexing a tensor with an object of type torch.LongTensor. The only supported types are integers, slices, numpy scalars and torch.LongTensor or torch.ByteTensor as the only argument.

[jdily]

Hi, I am under v0.2, and have the same issue mentioned as @alexholdenmiller
and i already try to use the fix as follow,
current_Q_values = Q(obs_batch).gather(1, act_batch.view(-1,1))
but still got the following error :
Index tensor must have same dimensions as input tensor at /home/jdily/Desktop/project/lib/pytorch/torch/lib/THC/generic/THCTensorScatterGather.cu:111

thx

[shkr]

I just received this error. It is very confusing

size '[-1 x 400]' is invalid for input of with 952576 elements at /Users/soumith/code/builder/wheel/pytorch-src/torch/lib/TH/THStorage.c:

[idansc]

torch.mm on 2D and 3D tensor returns
RuntimeError: matrices expected, got 2D, 3D tensors at /Users/soumith/code/builder/wheel/pytorch-src/aten/src/TH/generic/THTensorMath.c:2028

I'd change term matrices to 2D tensors

[ezyang]

@idansc Sure, sounds reasonable. Are you interested in submitting a PR?