When using your implementation, I noticed that the number of parameters doesn't match the original paper when used with DenseNet-BC_250_24 and DenseNet-BC_190_40, only DenseNet-BC_100_12 matches. This modification changes the first convolutional output channels to twice the growth rate so that the number of parameters matches those in the paper.
When using your implementation, I noticed that the number of parameters doesn't match the original paper when used with DenseNet-BC_250_24 and DenseNet-BC_190_40, only DenseNet-BC_100_12 matches. This modification changes the first convolutional output channels to twice the growth rate so that the number of parameters matches those in the paper.