Table 2:

Impact of network architecture on tomoDRGN homogeneous network reconstruction.

Box size (px)	Architecture	# Trainable parameters	# Data points per particle	Training time per 1k particles (min)	VRAM per particle (GB)	Max resolution (1/px)	Epochs to max resolution	Wall clock to max resolution (min)
64	64 × 3	24,962	96,776	0.15	1.42	0.48	1	0.75
64	128 × 3	74,498	96,776	0.17	1.54	0.48	1	0.85
64	256 × 3	247,298	96,776	0.22	1.76	0.48	1	1.11
64	512 × 3	887,810	96,776	0.37	2.15	0.48	1	1.84
64	768 × 3	1,921,538	96,776	0.58	2.79	0.48	1	2.88
128	64 × 3	37,250	194,064	0.28	1.85	0.38	43	60.74
128	128 × 3	99,074	194,064	0.33	2.04	0.49	15	24.94
128	256 × 3	296,450	194,064	0.43	2.47	0.49	4	8.69
128	512 × 3	986,114	194,064	0.75	3.33	0.49	2	7.48
128	768 × 3	2,068,994	194,064	1.22	4.53	0.49	1	6.12
256	64 × 3	61,826	378,516	0.88	3.48	0.20	42	185.81
256	128 × 3	148,226	378,516	0.94	3.67	0.26	47	221.67
256	256 × 3	394,754	378,516	1.15	4.42	0.32	33	190.30
256	512 × 3	1,182,722	378,516	1.81	6.10	0.34	22	199.42
256	768 × 3	2,363,906	378,516	3.61	11.92	0.35	20	360.84

Summary statistics for tomoDRGN homogeneous network training using the simulated ribosome class E particles at various box and pixel sizes, sweeping the number of nodes per layer in the decoder network.