pb转tflite

tensorflow官方模型转换教程 https://tensorflow.google.cn/lite/convert

tensorflow1.8版本好像一些功能没有，我们还是通过bazel的方式转换

 bazel build tensorflow/lite/toco:toco
 

bazel-bin/tensorflow/lite/toco/toco --input_file='../kerasyolo3/m/transfer_multi_4cl_1207.pb' --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --output_file='../kerasyolo3/m/transfer_multi_4cl_1207.tflite' --inference_type=FLOAT --input_type=FLOAT --input_arrays=input_1 --output_arrays=output_1,output_2,output_3 --input_shapes=1,416,416,3

--post_training_quantize用于模型压缩

 

tensorflow1.13.1 通过下面的代码也可以模型压缩，但是不知道压缩了什么，为什么反而慢这么多。

converter=tf.lite.TFLiteConverter.from_keras_model_file('transfer_multi_4cl_1207.h5',input_shapes={"input_1" : [1, 416, 416, 3]})
converter.post_training_quantize=True
tflite_quantized_model=converter.convert()
open('quantized_model.tflite', 'wb').write(tflite_quantized_model)

---

tf-nightly 1.14  版本下借助于tf.lite.Interpreter，可以在电脑段运行tflite。采用tflite.invoke。但是tensorflow lite貌似只能在cpu下跑，不管是不是装的gpu版本。

我的测试结果，h5模型在gpu下运行4.8张/秒，在cpu下运行大概1.2张/秒；pb模型不采用8位存储 gpu下4.8张/秒，8位存储gpu下3.6张/秒；转化为tflite，不采用post_training_quantize，只能在cpu下，也是1.2张/秒，但是采用post_training_quantize后，只能在cpu下运行 13秒/张（我要崩溃了，还特意装的最新每日版tensorflow，结果反而慢了。）

 

#coding=utf-8
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
import cv2
import numpy as np
import time

import tensorflow as tf

def preprocess_input(image, net_h, net_w):
    new_h, new_w, _ = image.shape

    # determine the new size of the image
    if (float(net_w) / new_w) < (float(net_h) / new_h):
        new_h = (new_h * net_w) // new_w
        new_w = net_w
    else:
        new_w = (new_w * net_h) // new_h
        new_h = net_h

    # resize the image to the new size
    resized = cv2.resize(image[:, :, ::-1] / 255., (new_w, new_h))

    # embed the image into the standard letter box
    new_image = np.ones((net_h, net_w, 3)) * 0.5
    new_image[(net_h - new_h) // 2:(net_h + new_h) // 2, (net_w - new_w) // 2:(net_w + new_w) // 2, :] = resized
    new_image = np.expand_dims(new_image, 0)

    return new_image

if __name__=='__main__':

    test_image_dir = ''
    # model_path = "./model/quantize_frozen_graph.tflite"
    model_path = "transfer_multi_4cl_1207.tflite"

    # Load TFLite model and allocate tensors.
    interpreter = tf.lite.Interpreter(model_path=model_path)
    interpreter.allocate_tensors()


    # Get input and output tensors.
    input_details = interpreter.get_input_details()
    print(str(input_details))
    output_details = interpreter.get_output_details()
    print(str(output_details))

    # x = interpreter.set_tensor(input_details[0]['index'])
    x = interpreter.tensor(interpreter.get_input_details()[0]["index"])
    y = interpreter.tensor(interpreter.get_output_details()[0]["index"])
    # y1 = interpreter.set_tensor(output_details[0]['index'])
    # y2 = interpreter.set_tensor(output_details[1]['index'])
    # y3 = interpreter.set_tensor(output_details[2]['index'])
    # 载入图片和调整
    image = cv2.imread('003000.jpg')
    batch_input = np.zeros((1, 416, 416, 3))
    batch_input[0] = preprocess_input(image, 416, 416)
    batch_input = batch_input.astype('float32')



    with tf.Session() as sess:
        #x().fill(batch_input)
        #interpreter.invoke()
        #print("inference %s" % y())

        #y_out = sess.run(y, feed_dict={x: batch_input})
        interpreter.set_tensor(input_details[0]['index'], batch_input)
        interpreter.invoke()
        output_data = [interpreter.get_tensor(output_details[0]['index']),interpreter.get_tensor(output_details[1]['index']),interpreter.get_tensor(output_details[2]['index'])]

        print(output_data)