Training Curve Explorer

Validation accuracy against wall-clock time and against epochs for every recorded run, on CPU. Pick a dataset and a pretrained backbone, then move the target to see how long each approach needed to first reach it. Further down: how the picture changes with the number of training images per class, and a cost comparison of all backbones. Everything is computed in your browser from the recorded curves; there is no server.

Dataset Pretrained backbone

Target validation accuracy 60%

Accuracy vs wall-clock time

Accuracy vs epochs

Thin lines are individual seeds; the bold line is the mean across seeds. The dashed line is the target. Markers on the bold lines show where the mean curve first reaches it.

Time to target at 60%

Approach	Runs reaching target	Mean time (s)	Mean epochs	Time saving vs scratch	Epoch saving vs scratch

Data efficiency: how many training images per class?

MobileNetV2 variants trained on 10, 25, 50, 100 or all training images per class (validation and test sets stay fixed). Each point is the mean of the recorded seeds, small dots are single seeds; hollow markers mean some runs never reached the target.

Final validation accuracy Time to target (uses the slider above)

Images per class	Approach	Epoch budget	Final val accuracy	Runs reaching target	Mean time to target (s)	Mean epochs to target

Backbone comparison on this dataset

All backbones use the same pipeline and seeds, trained on all training images. Time to target follows the slider above.

Backbone	Approach	Trainable / total parameters	Epoch 1 (s)	Steady epoch (s)	Final val accuracy	Runs reaching target	Mean time to target (s)

How it works: feature extraction vs fine-tuning

All three approaches classify the same small images (flowers or satellite tiles) and share the same input scaling, augmentation, classification head, optimiser family and batch size. They differ only in which weights start pretrained and which weights are updated.

From scratch. A small CNN learns edges, textures and shapes from only the training images.

Feature extraction. A pretrained backbone turns each image into features. Only the small classifier on top is trained, so far fewer weights change.

Fine-tuning. The same backbone, but its last blocks are also updated so high-level features can adapt to the new images. This costs extra compute per epoch.

Time-to-target is the cumulative wall-clock training time (including the per-epoch validation pass) at the end of the first epoch whose validation accuracy is at least the target. A run that never reaches the target is charged its full budget, so its time is a lower bound. The saving is one minus the mean time of an approach divided by the mean time of training from scratch.

Independent experiment on public data (TF Flowers, CC BY 2.0 photos; EuroSAT RGB, MIT licence); the recorded curves come from the runs committed in this repository. This page does not reproduce, and is not, the training-time result reported on the author's CV.