DECORE · Compressed vs. Full VGG-16

One reinforcement-learning agent per channel learns which to drop. All three models run entirely in your browser — no server, no GPU.

63% fewer parameters 39% fewer FLOPs +0.3% accuracy ⚡ on-device inference
MetricFull VGG-16DECORE-prunedDECORE + INT8
Parameters14.72 M5.43 M5.43 M
Model size58.9 MB21.8 MB5.5 MB
Top-1 accuracy93.48%93.76%93.68%
vs. full size2.7× smaller~11× smaller
Pick an example below
or upload your own image

Examples (CIFAR-10 test set)

Pick an image to compare the models.

Full VGG-16

14.7M params
Loading…

DECORE-pruned

5.4M params
Loading…

DECORE + INT8 final deliverable

5.5 MB · INT8
Loading…

💰 What would this save you?

Same accuracy, ~2× faster end-to-end, 63% smaller. A model's real-world worth is load time + inference time combined — load is what a cold-starting serverless replica or a first-time visitor pays, inference is what every request pays. Enter your current spend to estimate the impact on your bill.

Load — · Inference — · Total —

Load + inference times are measured live in your browser (from page load and from running an image above). Savings assume compute cost scales with combined load+inference time at constant request volume — a directional estimate, not a quote.

Total speedup (load + inference)
Est. monthly savings $0
Est. annual savings $0