lazy-toad/alexnet-imagenette

AlexNet — Imagenette (320px)

A from-scratch AlexNet implementation trained on the Imagenette 320px dataset (10 classes, ~9.5k training images).

Trained as a learning project to understand classic CNN architectures end-to-end — data loading, training loop, mixed precision, evaluation, and deployment.

Results

Metric	Value
Test accuracy	79.09%
Best validation accuracy	79.53%
Training epochs	30
Optimizer	SGD (lr=0.01, momentum=0.9, weight_decay=5e-4)
LR schedule	StepLR (step_size=15, gamma=0.1)
Hardware	Kaggle T4 GPU
Mixed precision	FP16 (AMP)
Training time	~25-30 min

Model architecture

Standard AlexNet, modified for 320×320 input:

Layer	Type	In → Out	Kernel / Stride / Padding
1	Conv2d	3 → 64	11×11 / 4 / 2
	ReLU + MaxPool2d		3×3 / 2
2	Conv2d	64 → 192	5×5 / 1 / 2
	ReLU + MaxPool2d		3×3 / 2
3	Conv2d	192 → 384	3×3 / 1 / 1
	ReLU
4	Conv2d	384 → 256	3×3 / 1 / 1
	ReLU
5	Conv2d	256 → 256	3×3 / 1 / 1
	ReLU + MaxPool2d		3×3 / 2
	Flatten	256×5×5 → 6400
6	Linear + Dropout(0.5) + ReLU	6400 → 4096
7	Linear + Dropout(0.5) + ReLU	4096 → 4096
8	Linear	4096 → 10

Classes

tench, english springer, cassette player, chain saw, church, french horn, garbage truck, gas pump, golf ball, parachute

Usage

Load weights

from huggingface_hub import hf_hub_download
import torch

weights_path = hf_hub_download(
    repo_id="lazy-toad/alexnet-imagenette",
    filename="alexnet_imagenette.pt",
)
state_dict = torch.load(weights_path, map_location="cpu")