Excessively long initialization time for Inception3

[os-gabe]

Creating an instance of the Inception3 class via torchvision.models.inception_v3(pretrained=False) or torchvision.models.Inception3() takes several minutes to complete.

It appears that all of the time is being spent here here.

I believe this may be related to some change in scipy perhaps as I do not think this model used to take so long to instantiate.

[os-gabe]

I confirmed that rolling scipy from 1.4.1 back to 1.3.1 causes the issue to go away.

So this is not strictly a torchvision bug but it does mean that Inception3 becomes somewhat unusable for people who install the current scipy version from pip.

[fmassa]

@os-gabe one alternative, given that we don't really train inception from scratch, would be to remove the custom initialization based on scipy.

That would be a breaking change though.

[os-gabe]

Here is the scipy issue for reference scipy/scipy#11299

[os-gabe]

@fmassa what if the Inception3 constructor took in an keyword argument to skip the random initialization? If the flag defaulted to False wouldn't this maintain backwards compatibility? This seems like a fairly nice solution in general since the random initialization is not needed when we just plan to load weights anyway.

[fmassa]

Yeah, we could do something similar to what we do in VGG

vision/torchvision/models/vgg.py

Line 26 in e130c6c

def __init__(self, features, num_classes=1000, init_weights=True):

with init_weights. Can you send a PR for inception?

[fmassa]

Fixed with #1832

[mingjunliu2]

Is this released?

[os-gabe]

There hasn't been a release since this change was merged. I'm curious to know when the next release is planned for.

[mingjunliu2]

I had very bad time to debug problem related to this issue.

[fmassa]

Next release for torchvision will happen in a couple of weeks. But the underlying problem is with scipy, so the "fix" we use in torchvision is to skip the default initialization if the user specifies it.