Deploying quantized Tensorflow lite model using zynq-library on Zynq-7000

Hi!

Have you got an Android OS (AOSP) on your Zedboard?  

Arm Compute Library (ACL) is a collection of low-level machine learning functions optimized for Arm architectures, but it's not designed for direct use by TensorFlow / TF Lite.  For that you're likely to want ArmNN, rather than writing a lot of code yourself.

ArmNN uses ACL, and to deploy your model on that, there's a great tutorial here: https://developer.arm.com/solutions/machine-learning-on-arm/developer-material/how-to-guides/deploying-a-quantized-tensorflow-lite-mobilenet-v1-model-using-the-arm-nn-sdk/single-page 

To build ArmNN on Android, there are instructions: https://github.com/ARM-software/armnn/blob/branches/armnn_21_02/BuildGuideAndroidNDK.md 

Hope that helps!

Ben

Hi!

Have you got an Android OS (AOSP) on your Zedboard?  

Arm Compute Library (ACL) is a collection of low-level machine learning functions optimized for Arm architectures, but it's not designed for direct use by TensorFlow / TF Lite.  For that you're likely to want ArmNN, rather than writing a lot of code yourself.

ArmNN uses ACL, and to deploy your model on that, there's a great tutorial here: https://developer.arm.com/solutions/machine-learning-on-arm/developer-material/how-to-guides/deploying-a-quantized-tensorflow-lite-mobilenet-v1-model-using-the-arm-nn-sdk/single-page 

To build ArmNN on Android, there are instructions: https://github.com/ARM-software/armnn/blob/branches/armnn_21_02/BuildGuideAndroidNDK.md 

Hope that helps!

Ben

HI!

Actually, due to the limited memory on Zedboard I was not able to build ARM NN and I choose the cross compilation of ARM NN on it. Now I want to implement a quantized model but I could'nt find any way for that 

If you're just wanting to use ACL you'd need to write the model parsing code yourself.  ACL provides all the functions once you're running, but ArmNN's designed to do all the setup.

Hi,

I tried to use ArmNN as you said and when testing ArmNN quantized Mobilenet in the step of building the mobilenetv1_quant_tflite program I get this error

arm-linux-gnueabihf-g++: error: -E or -x required when input is from standard input
arm-linux-gnueabihf-g++: error: -E or -x required when input is from standard input
Makefile:14: recipe for target 'mobilenetv1_quant_tflite' failed
make: *** [mobilenetv1_quant_tflite] Error 1

Any suggestions please??

That looks like a bug in your command line.  That's a standard gcc / clang error, likely an incorrect flag, or misplaced '-'

This is the makefile. Actually I checked and I couldn't find the incorrect flag

ARMNN_ROOT = /home/lsa/armnn-pi/armnn
ARMNN_BUILD = /home/lsa/armnn-pi/armnn/build
BOOST_ROOT = /home/lsa/armnn-pi/boost
CXX= arm-linux-gnueabihf-g++
CPPFLAGS=-DARMNN_TF_LITE_PARSER -I$(ARMNN_ROOT)/include -I$(ARMNN_ROOT)/src/backends -I$(ARMNN_ROOT)/src/armnnUtils -I$(ARMNN_ROOT)/tests -I$(BOOST_ROOT)/include
CFLAGS=-Wall -O3 -std=c++14 -fPIE

LDFLAGS=-pie -L$(ARMNN_BUILD) -L$(ARMNN_BUILD)/tests -L$(BOOST_ROOT)/lib
LDLIBS=-larmnn -larmnnTfLiteParser -lboost_system -lboost_filesystem -lboost_program_options

all: mobilenetv1_quant_tflite

mobilenetv1_quant_tflite: mobilenetv1_quant_tflite.cpp inference_test_image.cpp utils.cpp
$(CXX) $(CPPFLAGS) $(CFLAGS) $^ -o $@ $(LDFLAGS) $(LDLIBS)

clean:
-rm -f mobilenetv1_quant_tflite

It's likely to be the command line rather than the Makefile.  What have you got there?

I just wrote make

Ah, so it's not that then! :)

I also can't see what's weird in the makefile, but others are more expert at make than me, so someone else might pick up something that's missing or whatever.

Ok thank you for your Help !