H5 model trained by keras to tflite

1、 Use transcoding

Training use tf1.13.1 and keras2.0
transformation tflite The environment is tf1.13.1, Close test available

The code is as follows ：

import tensorflow.lite as lite
import argparse

def parse_args():

    parser = argparse.ArgumentParser(description='Keras to TensorFlow Lite converter')

    parser.add_argument('--input_keras',
                        required=True,
                        type=str,
                        help='The input Keras file model (.h5)')

    parser.add_argument('--output_tflite',
                        required=True,
                        type=str,
                        help='The output TensorFlow Lite file model (.tflite)')

    parser.add_argument('--post_quantize',
                        required=False,
                        type=bool,
                        help='Use post-quantization')

    args = parser.parse_args()
    return args

def convert(args):

    input_file = args.input_keras
    output_file = args.output_tflite

    # Converts the Keras model to TensorFlow Lite
    converter = lite.TocoConverter.from_keras_model_file(input_file)
    converter.post_training_quantize = True
    tflite_model = converter.convert()
    open(output_file, "wb").write(tflite_model)


def run():
    args = parse_args()
    convert(args)

if __name__ == "__main__":
    run()

2、 Use tflite_convert Tools

I haven't tried yet ： Use tflite_convert The command tool will keras h5 The file is converted to tflite Easy guide to documentation
This tutorial is divided into two steps , First the h5 transformation （ By code API transformation ） by pb Model , then pb Model transformation （ the tflite_convert Tool conversion ） by tflite

3、 Personal test use tflite_convert Tools

Model training uses tf1.13.1 and keras2.0
transformation tflite The environment is tf1.13.1

3.1 brief introduction

tflite_convert Is a command line tool , You can take a look at will pb Model transfer tflite The general process of ：TF1.x and TF2.x Version trained PB Model transfer Tflite. Found that we are right keras The model is converted to tflite You need to know h5 Information about the model ：

usage: tflite_convert --output_file OUTPUT_FILE	#  The output path 
                      --keras_model_file KERAS_MODEL_FILE	# keras Of h5 Model file path 
                      --output_format TFLITE	#  Output format , Designated as TFLITE（ Fix ）
                      --input_arrays INPUT_ARRAYS	#  Enter the node name 
                      --input_shapes INPUT_SHAPES	#  Enter shape 
                      --output_arrays OUTPUT_ARRAYS	#  Output node name 

Now let's talk about input_arrays 、input_shapes、output_arrays How to get

3.2 Model information acquisition

Give the code directly ：

from keras.models import load_model
net = load_model(r"model.hdf5", compile=False)

print("inputs info",net.inputs)	
print("inputs name",[node.op.name for node in net.inputs])
print("output name",[node.op.name for node in net.outputs])

Program output ：

#  It can be seen that shape, stay tflite_convert Designated as 1,64,64,1
inputs info [<tf.Tensor 'input_1:0' shape=(?, 64, 64, 1) dtype=float32>]

#  It can be seen that the input node name is input_1
inputs name ['input_1']

#  You can see that the name of the output node is predictions/Softmax
output name ['predictions/Softmax']

In fact, you can directly print(net.inputs) and print(net.outputs), Output is

:0 It takes into account the multi input and multi output models , But in fact, this model is single input and single output , So that is 'input_1' and 'predictions/Softmax'

3.3 transformation

Execute on the command line ：

tflite_convert --output_file output.tflite 	#  Specify the output path 
			   --keras_model_file newdata_mini_XCEPTION.123-0.70.hdf5 #  Specify the model to convert 
			   --input_arrays input_1 	#  Specify the input node name 
			   --input_shapes 1,64,64,1 	#  Input shape
			   --output_arrays predictions/Softmax 	#  Output node name 
			   --output_format TFLITE	# TFLITE Pattern 

If no error is reported , And it exists output.tflite file , The conversion should be successful
output.tflite It should be stored in the directory where you execute the command line , Or specify the absolute path when executing the command , Easy to find output.tflite file .