The matching problem between model parameters and model structure

In the original paper ''HoverNet'', the authors mentioned that "Compared to standard Preact-ResNet50 implementation, we reduce the total down-sampling factor from 32 to 8 by using stride of 1 in the first convolution and removing the subsequent max-pooling operation." Meantime, I see you have updated the codes "Hover-Net for CoNIC Challenge" and provided the pretrained model parameters—— Pytorch ImageNet ResNet50 in README.md. But the model parameters are trained in original ResNet 50 rather than the structure in above paper and the code in "net_desc.py" still use stride=1 (Line 42). So I would like to ask whether it is necessary to change the stride to the original ResNet 50 model so as to match the parameters.

Thank you!

Hi !

We simply use the ImageNet weights out-of-the-box, even though we don't downsample by a factor of 32 like is done in the original ImageNet trained models. So if you wanted to use them in the same way that we did, you don't need to do anything extra.

Of course you are more than welcome to match the downsampling factor to see if this boosts performance :)

Simon