All Stories

EfficientNet

EfficientNet's compound scaling rule, the FLOPs constraint that makes it work, the grid-searched constants, and the ImageNet accuracy it buys.

• cnn
• computer-vision
• coursera
• deep-learning
• Written in EN

Daniel Du
11 Apr 2023 · 5 min read

MobileNet Architecture

Why MobileNetV2 expands before it filters, what makes the residual inverted, and why the last layer of the block has no ReLU on it.

• cnn
• computer-vision
• coursera
• deep-learning
• Written in EN

Daniel Du
09 Apr 2023 · 3 min read

MobileNet

Depthwise separable convolution derived and costed: where the 1/N + 1/f-squared ratio comes from, and what an 8-9x saving actually looks like.

• cnn
• computer-vision
• coursera
• deep-learning
• Written in EN

Daniel Du
07 Apr 2023 · 4 min read

The Inception Network

How GoogLeNet stacks nine Inception modules into 6.8M parameters, what the auxiliary classifiers were really for, and what v2 and v3 changed.

• cnn
• computer-vision
• coursera
• deep-learning
• Written in EN

Daniel Du
05 Apr 2023 · 3 min read

Inception Network Motivation

Why the Inception module refuses to choose a filter size, and the 120M to 12.4M bottleneck arithmetic that makes running all of them affordable.

• cnn
• computer-vision
• coursera
• deep-learning
• Written in EN

Daniel Du
03 Apr 2023 · 3 min read

1x1 Convolutions and Network in Network

Why a 1x1 filter is not a no-op, how it becomes a per-pixel fully connected layer across channels, and the bottleneck it makes possible.

• cnn
• computer-vision
• coursera
• deep-learning
• Written in EN

Daniel Du
31 Mar 2023 · 3 min read