Table 3 Training and performance... | Emerald Publishing

Table 3

Training and performance hyperparameters for benchmark CNN models

Parameter	ResNet-18	Deep Baseline CNN
Input size	2 × 500 × 500	2 × 500 × 500
Parameter quantity	12.5 m	11m
Batch size	64	64
Learning rate	1 × 10^–3	1 × 10^–3
Optimizer	Adam	Adam
Weight decay	1 × 10^–5	1 × 10^–5
Data augmentation	Random flips, crops	Random flips, crops
Epochs	100	110
Training time (wall-clock)	6 h	6.5 h
GPU memory peak	8 GB	8 GB
Inference latency per sample	45 ms	50 ms
List of channels in conv layers	[ 64, 128, 256, 512 ]	[ 64, 128, 256, 512 ]
Sizes of FC hidden layers	[ 512, 1,024, 256 ]	[ 512, 1,024, 256 ]
Dropout rate	0.3	0.3

Parameter	ResNet-18	Deep Baseline CNN
Input size	2 × 500 × 500	2 × 500 × 500
Parameter quantity	12.5 m	11m
Batch size	64	64
Learning rate	1 × 10^–3	1 × 10^–3
Optimizer	Adam	Adam
Weight decay	1 × 10^–5	1 × 10^–5
Data augmentation	Random flips, crops	Random flips, crops
Epochs	100	110
Training time (wall-clock)	6 h	6.5 h
GPU memory peak	8 GB	8 GB
Inference latency per sample	45 ms	50 ms
List of channels in conv layers	[ 64, 128, 256, 512 ]	[ 64, 128, 256, 512 ]
Sizes of FC hidden layers	[ 512, 1,024, 256 ]	[ 512, 1,024, 256 ]
Dropout rate	0.3	0.3