Display order of properties of a tensor object in debugging

[Jesbong]

Can I adjust the display order of properties of a tensor object when debugging Python in vscode? Because in deep learning, we focus more on the shape of a tensor rather than specific values. Now we want to see the value of a tensor by pulling it down a long distance. This has reduced my debugging efficiency.

[int19h]

We don't officially support this, but I believe you could author a pydevd extension to customize the representation for your type - have a look at https://github.com/microsoft/debugpy/tree/main/src/debugpy/_vendored/pydevd/pydevd_plugins/extensions. Note however that we will be moving away from pydevd in the future.

However, given that ML/AI is a particularly important scenario these days, I'd love to hear more about the nature of customizations in question, so that we can properly address this in our new debugger.

[Jesbong]

The shape(dimension) of a tensor in vscode is far away from the top. Frequent need to expand and view results in low efficiency.

And pycharm can see the type and shape without opening the variable, as shown in the figure. We are very concerned about the dimensions of tensors in ML/AI, it's so convenient.

[Jesbong]

I believe the following issue are all about the same thing. The shape of numpy and tensor is so important for debuger in ML/AI.

#1191
pytorch/pytorch#35071
pytorch/pytorch#63855
microsoft/vscode-python#14244

[54nzg]

The shape of a variable cannot be easily viewed.Whether mouse hover will be supported in the future to see the shape of the variable.

[makaspacex]

It is obvious that everyone is very satisfied with the debugging of pycharm. If there are no patents or other restrictions, you can copy the display stype and information of pycharm. If there are restrictions, ignore my comments pls.

[makaspacex]

In addition to tensor and numpy, pycharm also has good type display support for list, tuple, dict, etc.

[haduoken]

I tempory use this
https://stackoverflow.com/questions/70704619/is-it-possible-to-show-variable-shapes-and-lengths-in-vscode-python-debugger-jus

but I strongly suggest vscode to support it natively

[ChaXxl]

I strongly suggest VS Code can show the len/shape of an countable object when debug, which PyCharm can do this, thanks to your work~~

[xuanhua]

any progress on this issue ? vscode is so unefficient for debugging deep learning program

[qige96]

any progress on this issue ? vscode is so unefficient for debugging deep learning program

I made a quick attempt to follow the advice

We don't officially support this, but I believe you could author a pydevd extension to customize the representation for your type - have a look at https://github.com/microsoft/debugpy/tree/main/src/debugpy/_vendored/pydevd/pydevd_plugins/extensions. Note however that we will be moving away from pydevd in the future.

Here is how I did: find the local directory where my pydevd extensions sit, in my case is
C:\Users\ruiqi\.vscode\extensions\ms-python.debugpy-2024.10.0-win32-x64\bundled\libs\debugpy\_vendored\pydevd\pydevd_plugins\extensions\;
in the types sub-directory, create a new file called pydevd_plugin_pytorch_tensor_str.py

'''
A simple example to show tensor shape on debugger
'''
from _pydevd_bundle.pydevd_extension_api import StrPresentationProvider
from .pydevd_helpers import find_mod_attr


class PyTorchTensorShapeStr:
    def can_provide(self, type_object, type_name):
        torch_tensor = find_mod_attr('torch', 'Tensor')
        return torch_tensor is not None and issubclass(type_object, torch_tensor)

    def get_str(self, val):
        return f'shape: {val.shape}, value: {val}'

import sys

if not sys.platform.startswith("java"):
    StrPresentationProvider.register(PyTorchTensorShapeStr)

Then, reload my vscode, and I can see the shapes

---

And if you would like to see the length/shape at any possible cases, you may leverage the collections.abc.Sized type to generalise the code, like:

from _pydevd_bundle.pydevd_extension_api import StrPresentationProvider
from .pydevd_helpers import find_mod_attr


class SizedShapeStr:
    '''Displays the size of a Sized object before displaying its value.
    '''
    def can_provide(self, type_object, type_name):
        sized_obj = find_mod_attr('collections.abc', 'Sized')
        return sized_obj is not None and issubclass(type_object, sized_obj)

    def get_str(self, val):
        if hasattr(val, 'shape'):
            return f'shape: {val.shape}, value: {val}'
        return f'len: {len(val)}, value: {val}'

import sys

if not sys.platform.startswith("java"):
    StrPresentationProvider.register(SizedShapeStr)

---

Hi @ChaXxl @xuanhua @54nzg @Jesbong, I made a pull request #1661 ,which is waiting to be reviewed. In case it is not approved, perhaps you make it by yourself.

Hope this may help you.