tensorflow-workspace/K210code

pb file to kmodel file

Make a directory named ncc. Download nncase tool and uncompress it to ncc.

pb file to tflite

Copy the pretrained model mobilenetv1_1.0.pb in pretrained directory to ncc/bin.

Enter ncc/bin directory.

toco --graph_def_file=mobilenetv1_1.0.pb --output_file=mobilenetv1_1.0.tflite --output_format=TFLITE --input_shape=1,224,224,3 --input_arrays=inputs --output_arrays=MobileNetV1/Bottleneck2/BatchNorm/Reshape_1 --inference_type=FLOAT

tflite to kmodel

Enter ncc directory and place a few pictures of your dataset into ncc/dataset directory.

./ncc compile ./bin/mobilenetv1_1.0.tflite ./bin/mobilenetv1_1.0.kmodel -i tflite -o kmodel --dataset ./dataset/

Note: Pictures in ncc/dataset are used for quantization. They should cover all classes of your dataset.

Prepare image for test

Convert an image, for example eagle.jpg, to a C file.

import numpy as np
import matplotlib.pyplot as plt

img = plt.imread('eagle.jpg')
img = np.transpose(img,[2,0,1]) # KPU requires NCHW format, 
								# while Tensorflow requires NHWC.
with open('image.c','w') as f:
    print('const unsigned char gImage_image[]={', file=f)
    print(', '.join([str(i) for i in img.flatten()]), file=f)
    print('};', file=f)

Test

Copy the K210code directory to kendryte-standalone-sdk/src. Build and download to KD233 to check the results.

Note: develop branch of kendryte-standalone-sdk is required.