Ch 7 — CNNs — Under the Hood

Convolution math, output dimensions, parameter counting, 1×1 convolutions, depthwise separable, and FLOP analysis
Index← High Level
Under the Hood
-
Click play or press Space to begin. Click any node for deep-dive details...
Step- / 10
AConvolution MathCross-correlation · output dimensions · multi-channel convolution
1
calculate
Conv FormulaDot product
sliding window
2
straighten
Output SizeStride, padding
dimension formula
3
arrow_downward Parameter counting and computational cost
BParameters & ComputationWeight counting · FLOPs · memory footprint
tag
Param CountK\u00b2\u00b7C_in\u00b7C_out
per conv layer
4
speed
FLOPsMultiply-adds
per layer
5
arrow_downward Advanced convolution types: 1\u00d71 and depthwise separable
CAdvanced Convolutions1×1 convolutions · depthwise separable · dilated convolutions
filter_1
1×1 ConvChannel mixing
dimensionality control
6
grid_view
Depthwise Sep.Spatial + channel
factored convolution
7
arrow_downward ResNet and Inception block internals
DArchitecture InternalsResNet bottleneck · Inception module · EfficientNet scaling
alt_route
ResNet BlockBottleneck design
1\u00d71 \u2192 3\u00d73 \u2192 1\u00d71
8
account_tree
InceptionMulti-scale parallel
branches merged
9
arrow_downward Object detection and segmentation architectures
EDetection & SegmentationYOLO grid · anchor boxes · U-Net encoder-decoder
select_all
YOLOGrid cells, anchors
single-pass detection
10
gradient
U-NetEncoder-decoder
skip connections