-
The point 1
- The following picture is a flow chart of the entire resnet26. We need to remember that we are typing in a fixed size 224 * 224
resnet26
conv7x7
block1(conv1x1-64,conv3x3-64,conv1x1-256)
block2(conv1x1-64,conv3x3-64,conv1x1-256)
block3(conv1x1-128,conv3x3-128,conv1x1-512)
block4(conv1x1-128,conv3x3-128,conv1x1-512)
block5(conv1x1-256,conv3x3-256,conv1x1-1024)
block6(conv1x1-256,conv3x3-256,conv1x1-1024)
block7(conv1x1-512,conv3x3-512,conv1x1-2048)
block8(conv1x1-512,conv3x3-512,conv1x1-2048)
fc
-
The point 2
- The H * W that we use is achieved by the convolution of the stride = 2 to become H/2 * W/2
- But the soul code that triggers down-sampling is inchannels! = block.expansion * out_chans
-
The point 3
- The stride of the first block in our first layer is 1, and the stride of the first block in the other layers is 2
