Table 3.

Computational performance (median time for N = 25 batches of 32 examples in seconds over N = 5 repetitions) and mean peak memory consumption (one batch of 32 examples in MiB, N = 6 repetitions) of the compared frameworks for the classification and segmentation benchmarks.

Task	deepee (ours)	Opacus	Pyvacy
Classification	38.82 s [38.67 to 39.08]	16.39 s [16.29 to 16.69]	73.11 s [72.41 to 75.40]
	6366 MiB [6201 to 6448]	7014 MiB [6816 to 7213]	2044 MiB [1992 to 2102]
Segmentation	70.89 s [70.41 to 71.01]	78.47 s [78.08 to 79.86]	97.89 s [97.26 to 99.16]
	9770 MiB [9508 to 9829]	9909 MiB [9812 to 10112]	2085 MiB [1890 to 2205]
Segmentation (Transposed Conv.)	47.27 s [45.12 to 51.15]	–	64.68 s [62.76 to 66.32]
	12014 MiB [11598 to 12249]	–	1537 MiB [1399 to 1620]

Ranges in angled brackets. The Segmentation (Transposed Conv.) row showcases framework performance in a U-Net architecture using transposed convolutions. Opacus is incompatible with this layer type.