[LibTorch] How to print the content of a tensor in gdb?

C++
#1

Hey guys,

I’m relatively new to C++ and LibTorch, but I would like to know how to inspect a tensor content in the debugger, since my workflow involves a lot of usage of the debugger.

In gdb I’m trying to do:

print x

$8 = {impl_ = {target_ = 0x14dfa20}}

print x.data()

Couldn't find method at::Tensor::data

I know torch::Tensor is a pointer, but I just can’t figure out how to dereference it

Cheers,
Miguel

3 Likes
#2

You could try something like the following:

With some modifications it could be used for visualizing tensors.

#3

how do you import that into gdb?

#4

You can use some p *x[0].data<float>() if it is a single dimension tensor, or p *x.data<float>() if it is a no-dim tensor, and so on.

Afshin

1 Like
#5

it is a 4 dimensional tensor, but I would like to know how to do it for a tensor of arbitrary size.

Basically what I’m looking for is an output similar to

std::cout << x << '\n';

if I have to implement a python extension, so be it… but I still don’t understand how to do it.

#6

I think you can see the content using cout in either run or debug consoles. Since you asked about gdb, I mentioned that approach. If you want to use gdb, you need to add more dimensions, e.g. p *x[i][j][k][l].data<float>().

1 Like
#7

that is not working for me

it either says Could not find operator[]. or No symbol "operator*" in current context.

I also couldn’t use cout directly in gdb, so far.

I’m new to C++, so I might be doing something wrong.

#8

you cannot use cout in gdb, you need to hard code it. The error of Could not find operator[] is your gdb error. Probably it does not mach torch.

#9

Can you try print x.toString()?

#10

That prints a memory address and the type of tensor

#11

print at::print(std::cout, x, 99), 99 being the line width.

Best regards

Thomas

1 Like
#12

I’m getting No symbol "print" in namespace "at".

I must be missing something…

#13

It’s defined in ATen/core/Formatting.h. As far as I can see it should be there for 1.0, 1.0.1, and master. You could try using it in your code to see what’s wrong.

Best regards

Thomas

1 Like
#14

i can use it while running the script, just not while attached to gdb

#15

Strange. It worked for me the other day. While the version without the ostream might not be available due to the inlining, I would have expected the one with to be exported at the least.

#16

Have you found a good solution?
In pytorch_1.3, try:

p self.impl_->sizes() 
# such as 4-dims
p self.impl_->sizes().Data[0]
...
p self.impl_->sizes().Data[3]

to inspect data size, and

p self.impl_->storage().storage_impl_->numel()   # num of elements
p  ((float *)self.impl_->storage().storage_impl_->data())[0]  # the first element

to inspect value of tensor.

#17

As of 1.5.0, Tensor.print() seems to work to print type and shape info.

gdb) call t.print()
[CPUByteType [1, 10, 1, 10]]

To print full contents, it seems the form std::cout << Tensor << std::endl must be present in debugged C++ file first. But you also have to call fflush(0) since gdb has trouble with std::endl in my test case.

A .gdbinit quick solution:

define ptensor
  call std::cout << $arg0
  call fflush(0)
end

(gdb) ptensor tensor
(1,1,.,.) =
  0  0  1  3  4  4  3  1  0  0
...

$4 = (std::ostream &) @0x55555597e740: <incomplete type>
[ CPUByteType{1,10,1,10} ]$5 = 0

But this prints the WHOLE tensor so only a partial solution.

1 Like
#18

Hi,
got it working in LLDB … but it has to be similar. Are you still interested?
Best regards