How to delete layer in pretrained model?

[ptrblck]

InceptionV3 uses functional calls in its forward, which will be missing, if you wrap all submodules in an nn.Sequential container.
In particular this flatten operation will be missing.

@jimmase

How can I output the last convolution layer output of Inception V3 with the shape [2048,5,5] before the pooling layer averages it to [2048,1,1]?

You could use forward hooks as described here.