将 keras/tensorflow h5/json 转换为 tensorflow pb 时遇到问题

Question

我使用 keras（张量流后端）训练了一个网络，并将模型保存为 json，权重保存为 h5。我现在尝试将其转换为单个张量流 pb 文件，它抱怨输出节点的名称。系统信息：Tensorflow 2.3.0 Keras 2.4.3 Cuda 10.1 Cudnn 7转换脚本非常简单：import jsonfrom tensorflow import kerasfrom keras import backend as Kimport tensorflow as tfjson_file = "my-trained-model.json"h5_file = "my-trained-model.h5"Output_Path = "./trained_models/"Frozen_pb_File = "my-trained-model.pb"def freeze_session(session, keep_var_names=None, output_names=None, clear_devices=True):       from tensorflow.python.framework.graph_util import convert_variables_to_constants    graph = session.graph    with graph.as_default():        freeze_var_names = list(set(v.op.name for v in tf.compat.v1.global_variables()).difference(keep_var_names or []))        output_names = output_names or []        output_names += [v.op.name for v in tf.compat.v1.global_variables()]        # Graph -> GraphDef ProtoBuf        input_graph_def = graph.as_graph_def()        if clear_devices:            for node in input_graph_def.node:                node.device = ""        frozen_graph = convert_variables_to_constants(session, input_graph_def,                                                      output_names, freeze_var_names)        return frozen_graphwith open(json_file, 'r') as json_file:    model = keras.models.model_from_json(json_file.read())model.load_weights(h5_file)model.summary()# get output node namesOutputNames = [out.op.name for out in model.outputs]print("\nOutput Names:\n", OutputNames)  # this prints "concatenate/concat" as the only output node name# freeze the modelfrozen_graph = freeze_session(tf.compat.v1.keras.backend.get_session(), output_names=OutputNames)# save the output files# this is the .pb file (a binary file)tf.io.write_graph(frozen_graph, Output_Path, Frozen_pb_File, as_text=False)当我运行这个时，AssertionError: concatenate/concat is not in graph因此，由于某种原因，它正在读取“concatenate/concat”的输出节点名称。下面给出模型总结，可以看到输出节点是“concatenate”；但是，即使我将输出节点名称硬编码为“连接”，我也会收到类似的断言错误：AssertionError: concatenate is not in graph

Answer 1

看起来这都是由于尝试冻结 TensorFlow 2.3 模型而引起的。显然，Tensorflow 2.0+ 已弃用“冻结”概念，转而采用“保存模型”概念。一旦发现这一点，我就能够立即将 h5/json 保存到已保存的模型 pb 中。

我仍然不确定这种格式是否针对推理进行了优化，所以我将对此进行一些跟进，但由于我的问题是关于我看到的错误，我想我会发布导致问题的原因。

作为参考，这是我的 python 脚本，用于将 keras h5/json 文件转换为 Tensorflow 保存的模型格式。

import os

from keras.models import model_from_json

import tensorflow as tf

import genericpath

from genericpath import *

def splitext(p):

    p = os.fspath(p)

    if isinstance(p, bytes):

        sep = b'/'

        extsep = b'.'

    else:

        sep = '/'

        extsep = '.'

    return genericpath._splitext(p, sep, None, extsep)

def load_model(path,custom_objects={},verbose=0):

    from keras.models import model_from_json

    path = splitext(path)[0]

    with open('%s.json' % path,'r') as json_file:

        model_json = json_file.read()

    model = model_from_json(model_json, custom_objects=custom_objects)

    model.load_weights('%s.h5' % path)

    # if verbose: print 'Loaded from %s' % path

    return model

json_file = "model.json"  # the h5 file should be "model.h5"

model = load_model(json_file) # load the json/h5 pair

model.save('my_saved_model') # this is a directory name to store the saved model