2021-11-08
31
IDS NXT: Technical manual IDS NXT rome
·
Squeeze-and-Excitation layer
·
ReLU activation
·
ReLU6 activation
·
Swish activation
·
Sigmoid activation
·
Batch normalization
Filter kernel/pooling parameters
·
Kernel/pooling window: any rectangle up to 15 x 15 pixels
·
Kernel depth: any depth
·
Kernel/pooling stride: up to 15 pixels
Inference time
See
11.1.2 Benchmark
Reference models from tensorflow.keras.applications
Architecture
Input format
output format
Single image inference time
[ms]
MobileNet V1 α=1.0
(224, 224, 3) > (1000)
66
MobileNet V1 α=0.75
(224, 224, 3) > (1000)
50
MobileNet V1 α=0.5
(224, 224, 3) > (1000)
34
MobileNet V1 α=0.5
(128, 128, 3) > (1000)
16
MobileNet V1 α=0.25
(224, 224, 3) > (1000)
28
MobileNet V1 α=0.25
(128, 128, 3) > (1000)
12
MobileNet V2 α=1.4
(224, 224, 3) > (1000)
102
MobileNet V2 α=1.0
(224, 224, 3) > (1000)
71
MobileNet V2 α=0.5
(224, 224, 3) > (1000)
47
MobileNet V3 large alpha 1.0
(224, 224, 3) > (1000)
73
MobileNet V3 large alpha 0.75
(224, 224, 3) > (1000)
65
MobileNet V3 small alpha 1.0
(224, 224, 3) > (1000)
33
ResNet50
(224, 224, 3) > (1000)
297
Xception
(224, 224, 3) > (1000)
596
MobileNet_V1_SSD
(300, 300, 3) > (scores: (3323, 81),
boxes: (3323, 4), anchors: (3323, 4))
132
