Tensorflow的slim框架可以写出像keras一样简单的代码来实现网络结构（虽然现在keras也已经集成在tf.contrib中了），而且models/slim提供了类似之前说过的object detection接口类似的image classification接口，可以很方便的进行fine-tuning利用自己的数据集训练自己所需的模型。

官方文档提供了比较详细的从数据准备，预训练模型的model zoo，fine-tuning，freeze model等一系列流程的步骤，但是缺少了inference的文档，不过tf所有模型的加载方式是通用的，所以调用方法和调用其他pb模型是一样的。

根据TF开发人员是说法Tensorflow对于模型读写的保存和调用的步骤一般如下：Build your graph --> write your graph --> import from written graph --> run compute etc。

以下我们使用slim提供的网络inception-resnet-v2作为例子：

1. export inference graph

import tensorflow as tf

import nets.inception_resnet_v2 as net

slim = tf.contrib.slim

# checkpoint path

checkpoint_path = "/your/path/to/inception_resnet_v2.ckpt" # ckpt file obtained during model training or fine-tuning

# set up and load session

sess = tf.Session()

arg_scope = net.inception_resnet_v2_arg_scope()

# initialize tensor suitable for model input

input_tensor = tf.placeholder(tf.float32, [None, 299, 299, 3])

with slim.arg_scope(arg_scope):

    logits, end_points = net.inception_resnet_v2(inputs=input_tensor)

# set up model saver

saver = tf.train.Saver()

saver.restore(sess, checkpoint_path)

with tf.gfile.GFile('/your/path/to/model_graph.pb', 'w') as f:   # save model to given pb file

    f.write(sess.graph_def.SerializeToString())

f.close()

2. freeze model

这里用tf提供的tensorflow/python/tools下的freeze_graph工具：

$ bazel build tensorflow/python/tools:freeze_graph

$ bazel-bin/tensorflow/python/tools/freeze_graph \

    --input_graph=/your/path/to/model_graph.pb  \  # obtained above

    --input_checkpoint=/your/path/to/inception_resnet_v2.ckpt \

    --input_binary=true

    --output_graph=/your/path/to/frozen_graph.pb \

    --output_node_names=InceptionResnetV2/Logits/Predictions   # output node name defined in inception resnet v2 net

(Optional) visualize frozen graph

LOG_DIR = ‘/tmp/graphdeflogdir’

model_filename = '/your/path/to/frozen_graph.pb'

with tf.Session() as sess:

    with tf.gfile.FastGFile(model_filename, 'rb') as f:

        graph_def = tf.GraphDef()

        graph_def.ParseFromString(f.read())

        graph = tf.import_graph_def(graph_def, name='')

    writer = tf.summary.FileWriter(LOG_DIR, graph_def)

writer.close()

然后用tensorborad --logdir=LOG_DIR选择graph就可以查看到frozen后的网络结构。

3. inference

import cv2

import numpy as np

def preprocess_inception(image_np, central_fraction=0.875):

    image_height, image_width, image_channel = image_np.shape

    if central_fraction:

        bbox_start_h = int(image_height * (1 - central_fraction) / 2)

        bbox_end_h = int(image_height - bbox_start_h)

        bbox_start_w = int(image_width * (1 - central_fraction) / 2)

        bbox_end_w = int(image_width - bbox_start_w)

        image_np = image_np[bbox_start_h:bbox_end_h, bbox_start_w:bbox_end_w]

    # normalize

    image_np = 2 * (image_np / 255.) - 1

    return image_np

image_np = cv2.imread("test.jpg")

# preprocess image as inception resnet v2 does

image_np = preprcess_inception(image_np)

# resize to model input image size

image_np = cv2.resize(image_np, (299, 299))

# expand dims to shape [None, 299, 299, 3]

image_np = np.expand_dims(image_np, 0)

# load model

with tf.gfile.GFile('/your/path/to/frozen_graph.pb')

    graph_def = tf.GraphDef()

    graph_def.ParseFromString(f.read())

    graph = tf.import_graph_def(graph_def, name='')

    with tf.Session(graph=graph) as sess:

        input tensor = sess.graph.get_tensor_by_name("input:0")   # get input tensor

        output_tensor = sess.graph.get_tensor_by_name("InceptionResnetV2/Logits/Predictions:0")  # get output tensor

        logits = sess.run(output_tensor, feed_dict={input_tensor: image_np})

        print "Prediciton label index:", np.argmax(logits[0], 1)

        print "Top 3 Prediciton label index:", np.argsort(logits[0], 3)

参考:

1. https://*.com/questions/42961243/using-pre-trained-inception-v4-model
2. https://gist.github.com/cchadowitz-pf/f1c3e781c125813f9976f6e69c06fec2
3. https://blog.metaflow.fr/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc
4. https://github.com/tensorflow/models/blob/master/slim/README.md
5. https://gist.github.com/tokestermw/795cc1fd6d0c9069b20204cbd133e36b