Edge Impulse Command line converter

Feature requests
#1

Does Edge Impulse have a command line converter? I am very envious of the Quantization to INT8 but with Float32 inputs and outputs. How I have previousely been doing quantiziation is a command like this from a tensorflow keras saved_model.pb file with folders assets and vairables using tflite_convert

installed with

pip install tf-nightly  tensorflowjs

one big command line

tflite_convert --saved_model_dir=./ --inference_type=tf.int8 --inference_output_type=tf.int8 --mean_value=128 --std_value=127 --output_file=./model_tf_INT8.tflite
xxd -i model_tf_INT8.tflite model_tf_INT8.h

Is there any way to do command line conversions on Edge Impulse (pehaps using edge-impulse-cli) and if not is this something edge impulse could allow from the main webpage.

#2

So you can grab the quantized tflite files (both with int8 and float32 inputs/outputs) from the dashboard:

image
image771×280 16 KB

Then you just need to run xxd over it. Or you can export as C++ Library, disable EON, and then find the xxd’ed file in tflite-model.

#3

Sorry @janjongboom, I was not clear about my question. I want to use an external model, not a model created using Edge Impulse. I can create tons of simple tensorflow keras .pb format pre-made models. Just wondering if their is an easy way to use the Edge Impulse tools to quantize my models?

#4

Not at the moment unfortunately. We need to have access to raw data to properly quantize the models, and are not set up to import pretrained models.

1 Like
#5

That makes sense. Thought I would ask.