Resnet 50 v1.5 definition

vision
1

Like to see the exact network architecture of resnet v1.5.

I see two references:
https://github.com/tensorflow/models/tree/master/official/resnet
“where a stride 2 is used on the 3x3 conv rather than the first 1x1 in the bottleneck”

" Strided convolution : When using the bottleneck architecture, we use stride 2 in the 3x3 convolution, instead of the first 1x1 convolution."

I have two questions.

  1. From the table in the resnet paper: https://arxiv.org/pdf/1512.03385.pdf

    image
    image875×334 50.6 KB

    “Downsampling is performed by conv3 1, conv4 1, and conv5 1 with a stride of 2.”,
    In v1.5, these 1x1 convolutions of stride 2 are removed and the following 3x3 is made into stride 2? (That would make sense) So, just 3 1x1 convolution blocks are removed and the 3 3x3 convolutions are made into stride 2?

  2. Is this all the changes?

2

Answer 1: In ResNet v1.5, 1x1 convolution layers haven’t been removed but modified to have a stride of 1 instead of 2. Meanwhile, the original 3x3 convolution layers have been modified to have a stride of 2 instead of 1. In summary, the previous downsample responsibility has been transferred from the first 1x1 convolution layer to the 3x3 convolution layer of Bottlenecks.

Answer 2: All the changes are as above.

Besides, I’d like to know where the name “V1.5” comes from since I can’t find the source of it. Hope someone could offer a source link.