Reduce the number of Conv layers in ResNet18 and 34? #395
Description
In the current ResNet model (examples/imagenet/models.py), there are three Conv layers in every ResidualBlock instance regardless of the number of layers. This is accurate for ResNet50 and above. ResNet18 and 34, however, are known to use two Conv layers per block instead. It would be fitting to bypass the 3rd Conv for models with less than 50 layers.