I have tried to compare my new backbone network with SOTA models such as DenseNet and ResNet. For fair comparison, all models need to be trained by same training protocols such as total training epoch, data augmentation, etc.
So,
I want to know the training protocol to get the reported accuracies of torch.models such as ResNet and DenseNet.
For example, I wonder whether those models were trained by default settings i.e., 90 epoch, initial learning rate of 0.1, learning rate schedule, same data augmentation( flip and normalization, not color augmentation,etc) that code provides, etc.
Sorry, it is unclear what is the question. Did you just answer yourself? default settings link seems to answer it. Unless you are asking for confirmation that those were the parameters used to train the models.
Sorry, my question was ambiguous. I wonder whether those models were trained by default settings or If ResNet or DenseNet used their own data augmentation such as color augmentation.
Thanks! I thought that was the case, just I wasnt entirely sure. Unfortunately we will have to wait for a tochvision dev to answer the question, as there is no way to be certain.