tensorflow pb to tflite 精度下降详解

之前希望在手机端使用深度模型做OCR，于是尝试在手机端部署tensorflow模型，用于图像分类。

思路主要是想使用tflite部署到安卓端，但是在使用tflite的时候发现模型的精度大幅度下降，已经不能支持业务需求了，最后就把OCR模型调用写在服务端了，但是精度下降的原因目前也没有找到，现在这里记录一下。

工作思路：

1.训练图像分类模型；2.模型固化成pb；3.由pb转成tflite文件；

但是使用python 的tf interpreter 调用tflite文件就已经出现精度下降的问题，android端部署也是一样。

1.网络结构

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf
slim = tf.contrib.slim

def ttnet(images, num_classes=10, is_training=False,
  dropout_keep_prob=0.5,
  prediction_fn=slim.softmax,
  scope='TtNet'):
end_points = {}

with tf.variable_scope(scope, 'TtNet', [images, num_classes]):
net = slim.conv2d(images, 32, [3, 3], scope='conv1')
# net = slim.conv2d(images, 64, [3, 3], scope='conv1_2')
net = slim.max_pool2d(net, [2, 2], 2, scope='pool1')
net = slim.batch_norm(net, activation_fn=tf.nn.relu, scope='bn1')
# net = slim.conv2d(net, 128, [3, 3], scope='conv2_1')
net = slim.conv2d(net, 64, [3, 3], scope='conv2')
net = slim.max_pool2d(net, [2, 2], 2, scope='pool2')
net = slim.conv2d(net, 128, [3, 3], scope='conv3')
net = slim.max_pool2d(net, [2, 2], 2, scope='pool3')
net = slim.conv2d(net, 256, [3, 3], scope='conv4')
net = slim.max_pool2d(net, [2, 2], 2, scope='pool4')
net = slim.batch_norm(net, activation_fn=tf.nn.relu, scope='bn2')
# net = slim.conv2d(net, 512, [3, 3], scope='conv5')
# net = slim.max_pool2d(net, [2, 2], 2, scope='pool5')
net = slim.flatten(net)
end_points['Flatten'] = net

# net = slim.fully_connected(net, 1024, scope='fc3')
net = slim.dropout(net, dropout_keep_prob, is_training=is_training,
     scope='dropout3')
logits = slim.fully_connected(net, num_classes, activation_fn=None,
        scope='fc4')
end_points['Logits'] = logits
end_points['Predictions'] = prediction_fn(logits, scope='Predictions')

return logits, end_points
ttnet.default_image_size = 28

def ttnet_arg_scope(weight_decay=0.0):
with slim.arg_scope(
 [slim.conv2d, slim.fully_connected],
 weights_regularizer=slim.l2_regularizer(weight_decay),
 weights_initializer=tf.truncated_normal_initializer(stddev=0.1),
 activation_fn=tf.nn.relu) as sc:
return sc

基于slim，由于是一个比较简单的分类问题，网络结构也很简单，几个卷积加池化。

测试效果是很棒的。真实样本测试集能达到99％+的准确率。

2.模型固化，生成pb文件

#coding:utf-8

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
from nets import nets_factory
import cv2
import os
import numpy as np
from datasets import dataset_factory
from preprocessing import preprocessing_factory
from tensorflow.python.platform import gfile
slim = tf.contrib.slim
#todo
#support arbitray image size and num_class

tf.app.flags.DEFINE_string(
'checkpoint_path', '/tmp/tfmodel/',
'The directory where the model was written to or an absolute path to a '
'checkpoint file.')

tf.app.flags.DEFINE_string(
'model_name', 'inception_v3', 'The name of the architecture to evaluate.')
tf.app.flags.DEFINE_string(
'preprocessing_name', None, 'The name of the preprocessing to use. If left '
'as `None`, then the model_name flag is used.')
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_integer(
'eval_image_size', None, 'Eval image size')
tf.app.flags.DEFINE_integer(
'eval_image_height', None, 'Eval image height')
tf.app.flags.DEFINE_integer(
'eval_image_width', None, 'Eval image width')
tf.app.flags.DEFINE_string(
'export_path', './ttnet_1.0_37_32.pb', 'the export path of the pd file')
FLAGS = tf.app.flags.FLAGS
NUM_CLASSES = 37

def main(_):
network_fn = nets_factory.get_network_fn(
 FLAGS.model_name,
 num_classes=NUM_CLASSES,
 is_training=False)
# pre_image = tf.placeholder(tf.float32, [None, None, 3], name='input_data')
# preprocessing_name = FLAGS.preprocessing_name or FLAGS.model_name
# image_preprocessing_fn = preprocessing_factory.get_preprocessing(
#  preprocessing_name,
#  is_training=False)
# image = image_preprocessing_fn(pre_image, FLAGS.eval_image_height, FLAGS.eval_image_width)
# images2 = tf.expand_dims(image, 0)
images2 = tf.placeholder(tf.float32, (None,32, 32, 3),name='input_data')
logits, endpoints = network_fn(images2)
with tf.Session() as sess:
output = tf.identity(endpoints['Predictions'],name="output_data")
with gfile.GFile(FLAGS.export_path, 'wb') as f:
 f.write(sess.graph_def.SerializeToString())

if __name__ == '__main__':
tf.app.run()

3.生成tflite文件

import tensorflow as tf

graph_def_file = "/datastore1/Colonist_Lord/Colonist_Lord/workspace/models/model_files/passport_model_with_tflite/ocr_frozen.pb"
input_arrays = ["input_data"]
output_arrays = ["output_data"]

converter = tf.lite.TFLiteConverter.from_frozen_graph(
graph_def_file, input_arrays, output_arrays)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

使用pb文件进行测试，效果正常；使用tflite文件进行测试，精度下降严重。下面附上pb与tflite测试代码。

pb测试代码

with tf.gfile.GFile(graph_filename, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())

with tf.Graph().as_default() as graph:
tf.import_graph_def(graph_def)
input_node = graph.get_tensor_by_name('import/input_data:0')
output_node = graph.get_tensor_by_name('import/output_data:0')
with tf.Session() as sess:
 for image_file in image_files:
  abs_path = os.path.join(image_folder, image_file)
  img = cv2.imread(abs_path).astype(np.float32)
  img = cv2.resize(img, (int(input_node.shape[1]), int(input_node.shape[2])))
  output_data = sess.run(output_node, feed_dict={input_node: [img]})
  index = np.argmax(output_data)
  label = dict_laebl[index]
  dst_floder = os.path.join(result_folder, label)
  if not os.path.exists(dst_floder):
   os.mkdir(dst_floder)
  cv2.imwrite(os.path.join(dst_floder, image_file), img)
  count += 1

tflite测试代码

model_path = "converted_model.tflite" #"/datastore1/Colonist_Lord/Colonist_Lord/data/passport_char/ocr.tflite"
interpreter = tf.contrib.lite.Interpreter(model_path=model_path)
interpreter.allocate_tensors()

# Get input and output tensors.
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
for image_file in image_files:
abs_path = os.path.join(image_folder,image_file)
img = cv2.imread(abs_path).astype(np.float32)
img = cv2.resize(img, tuple(input_details[0]['shape'][1:3]))
# input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
interpreter.set_tensor(input_details[0]['index'], [img])

interpreter.invoke()
output_data = interpreter.get_tensor(output_details[0]['index'])
index = np.argmax(output_data)
label = dict_laebl[index]
dst_floder = os.path.join(result_folder,label)
if not os.path.exists(dst_floder):
 os.mkdir(dst_floder)
cv2.imwrite(os.path.join(dst_floder,image_file),img)
count+=1

最后也算是绕过这个问题解决了业务需求，后面有空的话，还是会花时间研究一下这个问题。

如果有哪个大佬知道原因，希望不吝赐教。

补充知识：.pb 转tflite代码，使用量化，减小体积，converter.post_training_quantize = True

import tensorflow as tf

path = "/home/python/Downloads/a.pb" # pb文件位置和文件名
inputs = ["input_images"] # 模型文件的输入节点名称
classes = ['feature_fusion/Conv_7/Sigmoid','feature_fusion/concat_3'] # 模型文件的输出节点名称
# converter = tf.contrib.lite.TocoConverter.from_frozen_graph(path, inputs, classes, input_shapes={'input_images':[1, 320, 320, 3]})
converter = tf.lite.TFLiteConverter.from_frozen_graph(path, inputs, classes,
             input_shapes={'input_images': [1, 320, 320, 3]})
converter.post_training_quantize = True
tflite_model = converter.convert()
open("/home/python/Downloads/aNew.tflite", "wb").write(tflite_model)

来源：https://blog.csdn.net/qian1122221/article/details/84590248