Name: Model Compression Lab
Author: MySimulator

Compression

Pruning (% weights removed): 0% Quantization: 32-bit float

Pruning removes the smallest-magnitude weights from a trained network, on the theory they contribute least. Quantization reduces the numeric precision used to store each remaining weight, trading a little accuracy for a much smaller file and faster inference.

Result

Model size—

Size reduction—

Test accuracy—

Weight distribution & decision boundary