Deep learning experiment

CIFAR-10 CNN: A Reproducible Baseline

A compact PyTorch image-classification experiment with a tracked configuration, learning history, class-level diagnostics, and an honest reference result.

Role: Project author
Classification: Reproducible experiment

Inspect repository

One number covering ten very different ones: 64.26% overall, from 33.5% on cat to 82.0% on automobile.

01 / 12One imageA CIFAR-10 test image, at the size the network actually receives it.

Dataset: CIFAR-1010 classes, 32 x 32 RGB
This image: cattest index 0, shown at native size
Test set: 10,0001,000 per class

A CIFAR-10 test image of class cat, 32 by 32 pixels — Real CIFAR-10 test images at their native 32 x 32 RGB. Every input to this model is 3,072 numbers, and no more.

A CIFAR-10 test image of class dog, 32 by 32 pixels — Real CIFAR-10 test images at their native 32 x 32 RGB. Every input to this model is 3,072 numbers, and no more.

Shapes, measured from the module rather than described

Input3 × 32 × 32input
Conv block 132 × 16 × 1610,272 params
Conv block 264 × 8 × 855,680 params
Conv block 3128 × 4 × 4221,952 params
Classifier10527,114 params

Custom CNN (3 conv blocks + FC), 815,018 trainable parameters. Trained on 15,000 of the 50,000 training images for 12 epochs with Adam and OneCycleLR.

Per-class accuracy against the 64.26% aggregate

airplane73%
automobile82%
bird35.3%
cat33.5%
deer46.9%
dog60.5%
frog81.6%
horse72.8%
ship76.3%
truck80.7%

The aggregate is the mean of a 48.5-point spread, from 33.5% on cat to 82% on automobile. The four vehicle classes average 78%; the six animal classes average 55.1%.

Confusion matrix, 10,000 test images. Rows are the true class, columns the predicted class.

Counts of test images by true class and predicted class. The largest off-diagonal value is 291, where cat was predicted as dog.
True class	airp	auto	bird	cat	deer	dog	frog	hors	ship	truc
airplane	730	38	57	5	9	6	19	19	88	29
automobile	14	820	1	2	2	3	10	10	11	127
bird	111	9	353	67	113	123	133	53	16	22
cat	20	13	58	335	37	291	145	39	26	36
deer	40	8	70	37	469	45	170	147	10	4
dog	13	5	49	139	50	605	38	76	10	15
frog	9	0	22	59	54	17	816	7	10	6
horse	9	3	18	26	55	114	17	728	1	29
ship	109	59	11	9	0	7	10	3	763	29
truck	31	104	2	6	1	5	13	17	14	807

The largest single cell off the diagonal is 291: cat predicted as dog. The model finds 33.5% of cats, and sends nearly as many of them to dog as it gets right. 82% of all 3,574 mistakes stay inside the vehicle group or inside the animal group — the boundary the model learned best is not one of the ten it was asked for.

Test accuracy: 64.26%
Best validation: 62.47%
Final train accuracy: 57.7%
Epochs: 12

A bounded educational baseline15,000 of the 50,000 training images and twelve epochs on a laptop GPU. It is not a state-of-the-art result and the repository does not present it as one.
No checkpoint, so no activationsThe run's weights are deliberately not versioned. Every feature map in this world is drawn from the architecture's real shapes; none is a recorded activation, and no prediction is shown for any individual image.
The run had not convergedValidation loss was still falling at the final epoch and training accuracy was still below validation accuracy. Twelve epochs was a budget, not a stopping criterion.
The aggregate is the least useful number64.26% is the mean of a 48.5-point spread, from 33.5% on cat to 82.0% on automobile. Quoting it alone describes almost nothing about the model's behaviour.
The matrix was read from a figureOnly the confusion figure is tracked, not its underlying array. It was transcribed and then checked against the per-class accuracies, the trace and every precision in the classification report before being used.
One run, one seedA single reference run. Nothing here establishes variance across seeds or across repeated training.

01Problem

Why this work exists

Small vision experiments are easy to overstate when only the best headline number survives and the run configuration is lost.

02Contribution

What I can claim

Implemented the training and evaluation path, tracked one bounded reference run, and retained per-class performance and plots rather than claiming an unrecorded full-dataset result.

Method

Detail traded for meaning

A convolutional stack answers a classification question by discarding almost everything, in a specific order.

291 cat images were predicted as dog.

Represent
The input is a grid of pixels - all detail, no interpretation.
Abstract
Each layer keeps fewer, larger features. Resolution falls as meaning rises, which is the trade the architecture exists to make.
Decide
What survives is a short list of scores over classes.
What this does not show
This is a compact baseline on a small, well-studied dataset. Its value is a reproducible reference point, not a competitive result.

03System

Workflow and decisions

01CIFAR-10 subset
02CNN training
03Validation selection
04Test evaluation
05Class diagnostics

PyTorch
torchvision
NumPy
Matplotlib
scikit-learn

04Evidence

What is actually versioned

Tracked test accuracy64.26%

15,000 training samples, 12 epochs, and 815,018 parameters in the versioned reference run.

Diagnostic range33.5–82.0%

Per-class accuracy exposes large variation hidden by the aggregate score.

05Quality controls

How the work is checked

Run configuration and metrics are stored together
Learning curves, confusion matrix, and class-level results are tracked
The portfolio uses the recorded 64.26% result rather than an aspirational 85% claim

06Limitations

Where the evidence stops

A bounded educational baseline, not a state-of-the-art result
The reference run uses a 15,000-image training subset
No trained checkpoint is versioned

What this changed in my practice

Reproducibility means keeping the unglamorous context—the subset, epochs, configuration, and weak classes—next to the score.

Source and documentation Discuss relevant work