Comparison of the computational cost and the number of weights for the uncompressed, structurally-compressed, and deep-compressed models. The computational cost is quantified by FLOPS, i.e., the number of floating-point operations required in each layer to classify one image. The compression ratio is defined as the number of FLOPS (or weights) required for the uncompressed model divided by those required for the compressed model. Note that for the structurally-compression model, the compression ratio based on the number of FLOPS is equal to the compression ratio based on the number weights because all filters are removed during this form of compression. Structural compression is also not defined for the fully connected layers; therefore no number is reported for these layers.

<p>Comparison of the computational cost and the number of weights for the uncompressed, structurally-compressed, and deep-compressed models. The computational cost is quantified by FLOPS, i.e., the number of floating-point operations required in each layer to classify one image. The compressio...

Full description