← All models

MODEL EXPLORER / VISION MILESTONES

The convolutional network that changed ImageNet

How do five convolutions turn an image into a classification vector?

Curated by TensorVizAlexNet · 4-step tour
Static architecture
How do five convolutions turn an image into a classification vector?
Opening the interactive graph…
Architecture snapshot · No Python requiredDownload graph ↓

RECORDED NUMERICAL EXAMPLE

Build features while reducing image resolution.

Five convolutions lead to a 16×6×6 representation and three fully connected layers. ReLU, overlapping pooling, LRN, grouped computation and dropout preserve the main architecture choices in a smaller model.

These are thumbnails of at most 8×8 cells, average-pooled from a recorded feature channel of a synthetic stripe image. The first two paths apply cross-channel LRN before pooling; the fifth pools directly. The displayed maps are untrained features and do not demonstrate ImageNet accuracy.

Before normalization / pooling · thumbnail

Min 0 · Max 0.006857

Recorded cell values
Row / col01234567
00.0068570.0068570.004285000.0042850.0068570.006857
10.0068570.0068570.004285000.0042850.0068570.006857
20.0068570.0068570.004285000.0042850.0068570.006857
30.0068570.0068570.004285000.0042850.0068570.006857
40.0068570.0068570.004285000.0042850.0068570.006857
50.0068570.0068570.004285000.0042850.0068570.006857
60.0068570.0068570.004285000.0042850.0068570.006857
70.0068570.0068570.004285000.0042850.0068570.006857
After pooling · thumbnail

Min 0 · Max 0.004077

Recorded cell values
Row / col01234567
00.0040770.0040770.002446000.0024460.0040770.004077
10.0040770.0040770.002446000.0024460.0040770.004077
20.0040770.0040770.002446000.0024460.0040770.004077
30.0040770.0040770.002446000.0024460.0040770.004077
40.0040770.0040770.002446000.0024460.0040770.004077
50.0040770.0040770.002446000.0024460.0040770.004077
60.0040770.0040770.002446000.0024460.0040770.004077
70.0040770.0040770.002446000.0024460.0040770.004077

Each heatmap uses its own scale. Mint is positive, rust is negative, and the lightest color is zero. Table values are rounded to four significant digits.

Feature shape · channels / height / width
8
27
27

Bars share one scale within this example. Values are rounded to four significant digits.

Spatial width before pool
55
Spatial width after pool
27
Values after pool · all channels
5832
Verified by the local recipe
  • The 227-pixel convention produces 55-pixel first-convolution maps and ten output logits
  • LRN matches the original cross-channel sum formula, accounting for PyTorch's alpha convention
  • Overlapping 3×3 stride-2 pooling matches an independent sliding-window maximum
  • Grouped convolution isolates its channel groups; dropout scales survivors in training and is identity in evaluation

Download the source and run.py to reproduce these checks. Choosing an example here replays recorded values.

ABOUT THIS EXAMPLE

Inspect an AlexNet reference with reduced widths, including grouped convolutions, local response normalization, overlapping pooling and dropout.

Architecture · 2012

AlexNet

Five convolutions and three fully connected layers with widths 8/16/24/24/16 and 64/64/10. Uses a 227-pixel input convention, grouped conv2/4/5, LRN after the first two ReLUs, 3×3 stride-2 pools and 50% dropout.

Source, capture & limitations +

Untrained reduced-width reference. It does not recreate the original two-GPU execution, ImageNet training, augmentation or accuracy. The 227-pixel input follows the common implementation convention that yields a 55-pixel first map. CPU capture runs in evaluation mode, so dropout is inactive.

Original TensorViz teaching example. PyTorch provides the underlying operators.

No separate redistribution license has been declared for these project examples.

Content revision
fa1d0156d71ecbf0
Source SHA-256
8e1a9aaec3de8fc9305b61411c0209c32aacbc6812aa2647307e55a6e4e05f38
Captured
2026-09-17 · Python 3.13.13 / Torch 2.7.1

Reduced AlexNet CPU forward, independent LRN/pooling calculations, grouped-convolution isolation and dropout-mode checks. No trained ImageNet evaluation.

Read the model manifest ↗