Models
Models are deep learning specific pipeline elements
.
Nimble provides first class support for three different hardware targets: cpu
, igpu
and gpu
and second level support for fpga
and asic
.
Each model lists the hardware targets they support, please contact a Megh Representative if you are interested in either fpga
or asic
devices.
Models can be configured to operate in one of two modes: asynchronous or synchronous.
Models operating in asynchronous mode are generally used for high throughput workloads whereas synchronous mode offers lower latency for those workloads that are sensitive to inference time.
Depending on the type of mode they may also include additional parameters such as score_threshold
and iou_threshold
for filtering the inference results.
A model can be described in a pipeline in two forms: short-form or long-from.
When describing the parameter lists f
refers to a floating point value, s
refers to a string, e
refers to an enum, l
refers to a list, and i
refers to an integer.
Short-form:
<e-hw_mode>:<s-model_name>:<e-mode>:<f-score_threshold>:<f-iou_threshold>
e-hw_mode
(optional defaults tocpu
): valid hardware mode.s-model_name
: valid model names as listed in Megh's Model Zoo.e-mode
(optional defaults toa
):s
latency mode, ora
throughput mode.f-score_threshold
(optional): The score thresholds.f-iou_threshold
(optional): The iou thresholds.
Long-form:
name: "model-name" # `s-model_name` (required)
hw: "cpu" # `e-hw_mode` (optional, default: "cpu")
mode: "a" # `e-mode` (optional, default: "a")
score_threshold: 0.4 # `f-score_threshold` (optional, default: 0.4)
iou_threshold: 0.3 # `f-iou_threshold` (optional, default: 0.3)
A good starting point for person detection are the Megh Models. They have been trained on our internal datasets and are optimised for surveillance camera and eye level deployments in street, factory, and supermarket scenarios. Each model has different SKUs and can be tailored to your accuracy and hardware requirements.