AlexNet feature-extracting backbone (no fully-connected head).
Builds an AlexNet with the Krizhevsky 2014 single-stream
OWT topology: five convolutional blocks (Conv1: 3→64 with
stride-4; Conv2: 64→192 with ;
Conv3: 192→384, Conv4: 384→256, Conv5: 256→256, all
), and overlapping max-pools
after blocks 1, 2, 5. Approximately 2.5 M parameters in the
convolutional trunk alone (≈61.1 M for the full classifier
variant).
Model Size
Parameters
pretrainedbool= False**overridesobject= {}AlexNetConfig to
customise individual fields (in_channels=1 for grayscale,
dropout=0.0 to disable the dropout layers of the
classifier head).Returns
AlexNetBackbone with the AlexNet configuration applied (or with
overrides merged on top of it).
Notes
Krizhevsky et al., NIPS 2012 reports a 15.3% ImageNet-1k top-5 validation error — more than ten percentage points below the runner-up — and is widely credited with re-igniting deep-learning research after the 2006-2011 lull. Channel widths follow Krizhevsky 2014 ("One weird trick for parallelizing convolutional neural networks"), the single-stream re-derivation of the original two-GPU model, since that is the topology every published reference checkpoint targets. No paper-cited "small / large" variants exist (H11) — AlexNet is a single architecture.
Examples
>>> import lucid
>>> from lucid.models.vision.alexnet import alexnet
>>> model = alexnet()
>>> x = lucid.randn(1, 3, 224, 224)
>>> out = model(x)
>>> out.last_hidden_state.shape # (B, 256, 6, 6)
(1, 256, 6, 6)