Deep learning based high-throughput phenotyping of chalkiness in rice exposed to high night temperature

Table 4 Classification networks: training time and model size

Model	Training time (s)	Number of parameters	Size (MB)	Acc. (%)
DenseNet-121	1522.88	6955906	28.4	95.61
DenseNet-161	2157.04	26,476,418	107.1	95.12
DenseNet-169	1306.20	12,487,810	50.9	94.63
ResNet-18	546.77	11,177,536	44.8	94.63
ResNet-34	719.41	21,285,696	85.3	94.15
ResNet-50	1011.85	23,512,128	94.4	94.88
ResNet-101	1668.41	42,504,256	170.6	95.12
ResNet-152	2172.97	58,147,904	233.4	94.88
SqueezeNet-1.0	533.15	736,450	3.0	95.12
SqueezeNet-1.1	481.53	723,522	2.9	94.39
VGG-11	2382.44	128,774,530	515.1	94.88
VGG-13	2641.00	128,959,042	515.9	94.39
VGG-16	2745.00	134,268,738	537.1	95.12
VGG-19	3079.89	139,578,434	558.4	94.15
EfficientNetB0	1198.53	4,052,126	33.0	95.13
EfficientNetB1	2243.48	6,577,794	53.4	95.13
EfficientNetB2	1882.26	7,771,380	62.9	93.67
EfficientNetB3	2696.21	10,786,602	87.1	95.13
EfficientNetB4	3476.74	17,677,402	142.3	95.38
EfficientNetB5	3584.68	28,517,618	229.1	93.67
EfficientNetB6	4946.95	40,964,746	328.3	94.16
Mask R-CNN	14863.00	42,504,256	255.9	N/A

The number following a network’s name denotes the number of layers in the network (as in DenseNet-121 or ResNet-101) or the version of the network (as in SqueezeNet-1.0 or EfficientNetB0). All models are trained on AWS p3.2xlarge instances. The training time it took to train each model for 200 epochs is reported in seconds (s). Model complexity is reported as the number of trainable parameters of the model, as well as the size of the model in MB. The accuracy of each model is also shown, and the best accuracy (Acc.) obtained for each type of model is highlighted in bold font

ISSN: 1746-4811