Update our cpp export tutorial

[Krovatkin]

No description provided.

[netlify]

Deploy preview for pytorch-tutorials-preview failed.

Built with commit 39aaf86

https://app.netlify.com/sites/pytorch-tutorials-preview/deploys/5d2f6862635013000a782293

[netlify]

Deploy preview for pytorch-tutorials-preview failed.

Built with commit 4294700

https://app.netlify.com/sites/pytorch-tutorials-preview/deploys/5d2f68d16b0ce000096c4cd7

[netlify]

Deploy preview for pytorch-tutorials-preview ready!

Built with commit d028e91

https://deploy-preview-560--pytorch-tutorials-preview.netlify.com

[driazati]

This flag is gone in nightly and recursive mode is always on

[Krovatkin]

oopsie, I'll fix this.

[brianjo]

@Krovatkin Nick, can you rebase this? Thanks!

[Krovatkin]

@brianjo done!

[suo]

@Krovatkin is this ready for merge?

[Krovatkin]

@suo looks it like it's out of date again. Lemme rebase it again

[SethHWeidman]

@brianjo @Krovatkin running the code in the tutorial here did not work. See here.

[Krovatkin]

@SethHWeidman I'll take a look!

[Krovatkin]

@SethHWeidman looks like you are using 1.1.0. This tutorial was updated for 1.2 or 1.1.1 whichever it will be called.
I tried the scripting part of the tutorial against our nightly and it seems to work: https://colab.research.google.com/drive/1BRE0hbKoW9jYwmNtxh4X93ol1xNB5304
I couldn't test tracing resnet since it seems to require GPU and I couldn't find nightly for GPU; I'll double check that it works in a little bit.

[SethHWeidman]

@Krovatkin cool - that is why I noted that I used 1.1.0 🙂. I will check that this works with nightly as well!

[SethHWeidman]

@Krovatkin still getting an error using 1.2.0. See here for how I exported the model into C++ (doesn't explicitly say how to do so in the tutorial).

When I save the model and try to load into C++, as in the tutorial, I get an error; it simply says "error loading the model". (Step 3 of the tutorial)

Also, the tutorial contains this line: traced_script_module.save("model.pt"), which doesn't make sense since we haven't defined an object called traced_script_module.

This is all on torch 1.2.0 by the way. Can you please try testing the C++ load sections and see if they work?

[Krovatkin]

I'll take a look!

…

On Thu, Aug 8, 2019, 15:05 Seth Weidman ***@***.***> wrote: @Krovatkin <https://github.com/Krovatkin> still getting an error using 1.2.0. See here <https://github.com/SethHWeidman/tutorials/blob/exploratory/advanced_source/cpp_test/CPP%20test.ipynb> for how I exported the model into C++ (doesn't explicitly say how to do so in the tutorial). When I save the model and try to load into C++, as in the tutorial, I get an error; it simply says "error loading the model". (Step 3 of the tutorial) Also, the tutorial contains this line: traced_script_module.save("model.pt "), which doesn't make sense since we haven't defined an object called traced_script_module. This is all on torch 1.2.0 by the way. Can you please try testing the C++ load sections and see if they work? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#560?email_source=notifications&email_token=ABGZY4WHEEM4QVM4VVB6HSDQDSKEJA5CNFSM4IETHWKKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD35BAZI#issuecomment-519704677>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABGZY4WRYMWB3VUOEO35H5LQDSKEJANCNFSM4IETHWKA> .

[SethHWeidman]

Thanks @Krovatkin.

Update:

• Reading the model into C++ works - you just have to point the command to the path of model.pt. Perhaps we could say "run the command ./example-app <PATH-TO-model.pt>" in the tutorial so people don't get confused if they're just following along.

• This is a minor point; the major point is: the second part, where we try to pass a Tensor of 1s through our model, generated the following error when I ran ./example-app ../model.pt from within build:

terminate called after throwing an instance of 'std::runtime_error'
  what():  matrix and vector expected, got 2D, 4D at /pytorch/aten/src/TH/generic/THTensorMath.cpp:626
The above operation failed in interpreter, with the following stack trace:
at code/model.py:6:14
op_version_set = 1
def forward(self,
    input: Tensor) -> Tensor:
  _0 = torch.gt(torch.sum(input, dtype=None), 0)
  if bool(_0):
    output = torch.mv(self.weight, input)
             ~~~~~~~~ <--- HERE
  else:
    output = torch.add(self.weight, input, alpha=1)
  return output
Compiled from code at <ipython-input-4-0d89e318784e>:8:20
    def forward(self, input):
        if input.sum() > 0:
          output = self.weight.mv(input)
                   ~~~~~~~~~~~~~~ <--- HERE
        else:
          output = self.weight + input
        return output

Aborted (core dumped)

See my .cpp file here.

Hope this helps.

[Krovatkin]

@SethHWeidman

Reading the model into C++ works - you just have to point the command to the path of model.pt. Perhaps we could say "run the command ./example-app <PATH-TO-model.pt>" in the tutorial so people don't get confused if they're just following along.

I'll make the correction!

where we try to pass a Tensor of 1s through our model, generated the following error when I ran ./example-app ../model.pt from within build:

What's input's shape you are passing in? mv expects a vector of shape (m) where m should match m of weight's shape (n, m)
I think the confusion might stem from the fact that the tutorial assumes that you are serializing/loading resnet18.

[SethHWeidman]

Hey @Krovatkin, I see what you mean, you're right: the problem is that we save two models during the tutorial, and unless people are paying close attention - which, I acknowledge, they should be! - they might try to run example-app with the wrong one.

How about we call the models different things to clarify - see here. I think just calling the ResNet model traced_resnet_model.pt could help quite a bit.

Tutorial is great by the way - very clear and to-the-point. Thanks for looking into this 🙂

[Krovatkin]

@SethHWeidman , Thank you very much for your very detailed feedback! I'll tried to address all the points you brought up in #609 . Please let me know if I missed anything.