IDS NXT rome - System Specification; Cnn; Deep Ocean Core

2021-11-08 30
IDS NXT: Technical manual IDS NXT rome
11 System specification
CPU
1.3 GHz Dual Core ARM Cortex-A53
FPGA
Xilinx Zynq UltraScale+ (ZU3CG)
RAM
2 GB DDR4
Flash
8 GB (with the system)
11.1 CNN
Operating system
Linux
Image preprocessing
· Demosaicing
· AOI
· Resolution scaling
· CNN preprocessing modes
o zeroto255
o zerotoone
· CNN preprocessing modes from
keras.applications.imagenet_utils.preprocess_input
o tf
o caffe
o torch
Data postprocessing on CPU
· Softmax operation for classification
· Box regression for object detection
· User programmable processing on Dual Core ARM Cortex-A53
11.1.1 deep ocean core
Input data formats
· Up to 512 x 512 pixels
· Tensor of shape (x, y, 3) for color
· Tensor of shape (x, y, 1) for mono
Output data formats
· Float
· Shape formating can be defined in post-processing (depending on
model and task)
Internal number format
· 16-bit fixed point
· Integer and fractional bits individually optimized for each layer
Maximum number of layers in
model
· 180 layers
Maximum model size
Any size fitting the RAM size (maximum 245 MB (1,6 MP mono),
minimum 160 MB (6 MP color)
Supported operations/layers
· 2D convolution layer
· Depthwise separable convolution layer
· Average pooling layer
· Max. pooling layer
· Dense layer
· Add layer
· Concatenate layer