AIoT应用创新大赛-基于TFML的迁移学习实践

conda create -n eiq python=3.8
conda activate eiq

python -m pip install -U pip
python -m pip install -U setuptools
python -m pip install tflite-model-maker
python -m pip install onnxmltools mmdnn tensorflow-datasets opencv-python PILLOW
python -m pip install numpy scipy matplotlib ipython jupyter pandas sympy nose imageio 
python -m pip install netron

wget http://download.tensorflow.org/example_images/flower_photos.tgz
tar -xvzf flower_photos.tgz

import os
import tensorflow as tf
from tflite_model_maker import image_classifier
from tflite_model_maker import ImageClassifierDataLoader
from tflite_model_maker.config import QuantizationConfig

#Specify image directory
image_path = os.path.join(os.getcwd(), 'flower_photos')

#Split up images into different training categories for training, validation, and testing. 
data = ImageClassifierDataLoader.from_folder(image_path)
train_data, rest_data = data.split(0.8)
validation_data, test_data = rest_data.split(0.5)

mobilenetv1_spec = image_classifier.ModelSpec(uri='https://tfhub.dev/google/imagenet/mobilenet_v1_025_128/feature_vector/4',input_image_shape=[128,128], name='mobilenet_v1')
model = image_classifier.create(train_data, model_spec=mobilenetv1_spec, validation_data=validation_data)
model.summary()

#Evaluate final model
print('Done training\n')
loss, accuracy = model.evaluate(test_data)

#Write out .tflite file
print('Write out model\n')

config = QuantizationConfig.for_int8(representative_data=test_data, inference_input_type=tf.int8, inference_output_type=tf.int8,
    supported_ops=tf.lite.OpsSet.TFLITE_BUILTINS_INT8) 

model.export(export_dir='.',tflite_filename='flower_model_int8.tflite',label_filename='flower_labels.txt',with_metadata=False,quantization_config=config)

xxd -i flower_model.tflite > flower_model.h

#include <cmsis_compiler.h>
#define MODEL_NAME "mobilenet_v1_0.25_128_flower"
#define MODEL_INPUT_MEAN 127.5f
#define MODEL_INPUT_STD 127.5f
const char flower_model_tflite[] __ALIGNED(16) = {

const char* labels[] = {
    "daisy",
    "dandelion",
    "roses",
    "sunflowers",
    "tulips"
};