onnc-bench 1.4.0

ONNC-bench is a Python wrapper of ONNC

ONNC-bench

ONNC-bench is a Python wrapper of ONNC

Installation

Using pip

pip install onnc-bench

Python API Example

Here is an example to show how to use ONNC python API

# Setup your ONNC API key
api_key = "Your API KEY"

# Instantiate a workspace for deploying model for device `M487`
workspace = launch(api_key, 'NUMAKER_IOT_M487')

# Quantize model to improve performance and reduce memory footprint.
# Here we need quantization dataset, using validation dataset
# is surfficent.
workspace.quantize(x_test)

# Compile the model and get the compilation results
report = workspace.compile(model, "input_1", "dense_1")["report"]

# Save the compiled model
workspace.save('./output')

# Release disk space in cloud
workspace.close()

print(report)
"""
{'ram': 2490, 'rom': 101970}

The report shows we need:
    2,490 bytes of SRAM
  101,970 bytes of ROM
to run this model on a CortexM device.
"""

CLI tools

onnc-bench comes with cli tools to help you deploy model faster. Follow below commands to scaffolding a bench.

1. Create and enter your bench

onnc-create mybench
cd mybench

2. Setup API key

onnc-login --key "Your-API-Key-Here"

3. Create an infer myinfer1 base on template vww

./create-project -t vww -o myinfer1

4. Compile the pretrained model

./build-project -t myinfer1 -d NUMAKER_IOT_M487

5. Deoply the compiled model

./deploy-project -t myinfer1 -o ./output

More examples can be found in examples, currently we provide below examples:

1. Keras MNIST: Contains a MNIST example in Keras from training to development.
2. Simple Example: Compile a serialized model, and download loadable with demo code in c++.