Fix warnings in unit-tests tests

[puhuk]

Fixes #6155

[vfdev-5]

@puhuk ci failures are related.

I think a fix could be something like:

diff --git a/torchvision/models/quantization/googlenet.py b/torchvision/models/quantization/googlenet.py
index 644df8ae4..682ba1833 100644
--- a/torchvision/models/quantization/googlenet.py
+++ b/torchvision/models/quantization/googlenet.py
@@ -74,6 +74,8 @@ class QuantizableInceptionAux(InceptionAux):
 class QuantizableGoogLeNet(GoogLeNet):
     # TODO https://github.com/pytorch/vision/pull/4232#pullrequestreview-730461659
     def __init__(self, *args: Any, **kwargs: Any) -> None:
+        if "init_weights" not in kwargs:
+            kwargs["init_weights"] = True
         super().__init__(  # type: ignore[misc]
             blocks=[QuantizableBasicConv2d, QuantizableInception, QuantizableInceptionAux], *args, **kwargs
         )
diff --git a/torchvision/models/quantization/inception.py b/torchvision/models/quantization/inception.py
index ba4b21d41..647c00472 100644
--- a/torchvision/models/quantization/inception.py
+++ b/torchvision/models/quantization/inception.py
@@ -142,6 +142,7 @@ class QuantizableInception3(inception_module.Inception3):
                 QuantizableInceptionE,
                 QuantizableInceptionAux,
             ],
+            init_weights=True,
         )
         self.quant = torch.ao.quantization.QuantStub()
         self.dequant = torch.ao.quantization.DeQuantStub()

[puhuk]

Hmm, why it raises typo error in test_models.py when I just revert to origin code

[vfdev-5]

@puhuk it is unrelated as main branch is broken : https://app.circleci.com/pipelines/github/pytorch/vision/18344/workflows/d0222980-4837-4c46-a9bf-d9cc4be80051/jobs/1484167

[puhuk]

@vfdev-5 Thank you for the feedback!

[vfdev-5]

@puhuk there are few similar warnings:

test/test_models.py::test_classification_model[cpu-googlenet]
  /root/project/torchvision/models/googlenet.py:47: FutureWarning: The default weight initialization of GoogleNet will be changed in future releases of torchvision. If you wish to keep the old behavior (which leads to long initialization times due to scipy/scipy#11299), please set init_weights=True.
    warnings.warn(

test/test_models.py::test_classification_model[cpu-inception_v3]
  /root/project/torchvision/models/inception.py:43: FutureWarning: The default weight initialization of inception_v3 will be changed in future releases of torchvision. If you wish to keep the old behavior (which leads to long initialization times due to scipy/scipy#11299), please set init_weights=True.
    warnings.warn(

Can you try to fix them as well ? Thanks

[vfdev-5]

This is not good. Check #2170 why they switched from True to None

[puhuk]

Hmm, I don't have good idea. Is there better method than I did here.
https://github.com/pytorch/vision/pull/6159/commits/16ddd6c0f23fe56e8029f336c5e3f3c69e9c534f

[datumbox]

I'm not sure why in order to fix the warning on unit-tests we modify the models themselves. My expectations without having all the details in my mind, is that the tests are the ones that need to be updated. What am I missing?

[vfdev-5]

At first I was thinking that QuantizableGoogLeNet does not properly set up its base class GoogLeNet. Thinking more about that, I think you are right, we have to try again to fix that from tests side

[vfdev-5]

@puhuk can you check if we could use _model_params when we set kwargs = {**defaults, **_model_params.get(model_name, {})}. Looks like _model_params can define specific values for special models:

vision/test/test_models.py

Lines 246 to 247 in 93b3e84

	_model_params = {
	"inception_v3": {"input_shape": (1, 3, 299, 299)},

[puhuk]

Ok, let me check and send PR again.

[datumbox]

I think the approach taken now should work. Our CI is toasted so we won't be able to test this thoroughly, so I recommend waiting for tomorrow until the breakage is fixed.

Edit: The failing linter is related. You need to add a space at "googlenet":{

[puhuk]

@datumbox
Thanks for the review, let me wait till tomorrow :)
And I fix the error by formatting with black.