python - 如何避免在修改时将 "keras_learning_phase"添加到模型中？

Question

我正在制作一个小程序来将模型从 .h5/.hdf5 转换为 .pb。

(将其输入层从其原始形状更改为[None, None, c])

((c的值与​​其原始模型的输入 channel 相同))

这样做的原因是我可以使用任何形状来推断训练好的模型。

到目前为止，我可以用两种不同的代码完成整个过程:

code_1: 将 .h5/.hdf5 转换为 .pb

from keras.models import load_model
import tensorflow as tf
import os
import json
from keras import backend as K
from tensorflow.python.framework import graph_io
import Unet
from keras.layers import Input

def focal_loss(gamma=2., alpha=.25):
    def focal_loss_fixed(y_true, y_pred):
        pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
        pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
        return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1+ K.pow(10.0, -9)))-K.sum((1-alpha) * K.pow( pt_0, gamma) * K.log(1. - pt_0+ K.pow(10.0, -9)))
    return focal_loss_fixed

def IOU_calc():
    def IOU_calc_fixed(y_true, y_pred):
        smooth = 0.001
        y_true_f = K.flatten(y_true)
        y_pred_f = K.flatten(y_pred)

        intersection = K.sum(y_true_f * y_pred_f)

        return 2*(intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
    return IOU_calc_fixed

def IOU_calc_loss():
    return lambda x: -IOU_calc()

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.global_variables()).difference(keep_var_names or []))
        output_names = output_names or []
        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)
        # add for new test
        frozen_graph = tf.graph_util.remove_training_nodes(frozen_graph)
        return frozen_graph

def h5_to_pb(h5_model, output_dir, model_name, out_prefix = "output_"):
    out_nodes = []
    for i in range(len(h5_model.outputs)):
        out_nodes.append(out_prefix + str(i + 1))
        tf.identity(h5_model.output[i], out_prefix + str(i + 1))
    # add for ignore dropout at inference process(?)
    K.set_learning_phase(0)

    sess = K.get_session()
    from tensorflow.python.framework import graph_util,graph_io
    init_graph = sess.graph.as_graph_def()
    main_graph = graph_util.convert_variables_to_constants(sess, init_graph, out_nodes)
    graph_io.write_graph(main_graph, output_dir, name = model_name, as_text = False)

if __name__ == '__main__':

    config_path = r'unet_h5_to_pb_config.json'
    with open(config_path) as config_buffer:
        a = config_buffer.read()
        a = a.replace('\\', '\\\\')
        config = json.loads(a)
    """-------------------------Set Paths-----------------------------------"""
    h5_model_path = config["h5_md_path"]

    if "\\" in h5_model_path:
        h5_name = h5_model_path.split("\\")[-1]
        pb_model_name = h5_name.split(".")[0] + ".pb"

        if len(config["pb_save_path"]) == 0:
            output_path = h5_model_path.strip(h5_model_path.split("\\")[-1])
            if h5_model_path[0] == ".":
                output_path = "." + output_path
        else:
            output_path = config["pb_save_path"]
            if os.path.exists(output_path) is False:
                os.mkdir(output_path)
        out_md_path = os.path.join(output_path, pb_model_name)
    else:
        h5_name = h5_model_path.split("/")[-1]
        pb_model_name = h5_name.split(".")[0] + ".pb"

        if len(config["pb_save_path"]) == 0:
            output_path = h5_model_path.strip(h5_model_path.split("/")[-1])
            if h5_model_path[0] == ".":
                output_path = "." + output_path
        else:
            output_path = config["pb_save_path"]
            if os.path.exists(output_path) is False:
                os.mkdir(output_path)
        out_md_path = os.path.join(output_path, pb_model_name)

    """---------------------------load h5 model-----------------------------"""
    K.set_learning_phase(0)
    try:
        net_model = load_model(h5_model_path,
                               custom_objects={'IOU_calc': IOU_calc()})
        print("***Handling model with structure***")
    except ValueError as E:
        if "Cannot create group in read only mode" in str(E):
            print("***Handling model without structure***")
            input_shape = tuple(config["input_shape"])
            model_class = config["model_class"]
            net_model = Unet.small_unet(input_shape, model_class)
            net_model.load_weights(h5_model_path)
        else:
            print(E)

    print('***input is :', net_model.input.name)
    print('***output is:', net_model.output.name)
    """----------------------------save pb file-----------------------------"""
    if config["name_output_node"]:
        h5_to_pb(net_model, output_dir=output_path, model_name=pb_model_name)
    else:
        sess = K.get_session()
        frozen_graph = freeze_session(K.get_session(),
                                      output_names=[net_model.output.op.name])
        graph_io.write_graph(frozen_graph,
                             output_path, pb_model_name, as_text=False)
    print(f"===== FINISH converting model to pb file {out_md_path} =====")

code_2: 改变.pb输入层形状

# ref: http://digital-thinking.de/tensorflow-replace-tensors-of-a-savedmodel-or-frozengraph/

import tensorflow as tf
from keras.layers import Input


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

    for n in graph_def.node:
        if "input_" in n.name:
            input_name = n.name + ":0"
            input_channel = n.attr["shape"].shape.dim[3].size

    with tf.Graph().as_default() as graph:
        tf.import_graph_def(graph_def, name='')

    return graph, input_name, input_channel


# ref:https://stackoverflow.com/a/56324805/10427809
def delete_ops_from_graph(graph_def, output_model_filepath):

    # Delete nodes
    nodes = []
    input_count = 0
    for node in graph_def.node:
        if "input" in node.name and "dropout" not in node.name:
            if input_count == 0:
                nodes.append(node)
                input_count += 1
            else:
                print('Drop', node.name)
        else:
            nodes.append(node)

    mod_graph_def = tf.GraphDef()
    mod_graph_def.node.extend(nodes)

    # Delete references to deleted nodes
    for node in mod_graph_def.node:
        inp_names = []
        for inp in node.input:
            if 'Neg' in inp:
                pass
            else:
                inp_names.append(inp)

        del node.input[:]
        node.input.extend(inp_names)
    with open(output_model_filepath, 'wb') as f:
        f.write(mod_graph_def.SerializeToString())
    print(f"*** save new model at{output_model_filepath}")

if __name__ == "__main__":

    tf_md_path = r'./pb/my_model_weights_480.pb'

    out_md_path = tf_md_path.replace(".pb", "_dynamic.pb")
    
    sess = tf.Session()
    graph_model, input_name, C = load_graph(tf_md_path)
    graph_model_def = graph_model.as_graph_def()
    resize_val = (None, None, C)
    rp_input = Input(resize_val)
    tf.import_graph_def(graph_model_def, name='',
                        input_map={input_name: rp_input})
    
    graph = tf.get_default_graph()
    graph_def = graph.as_graph_def()
    print(graph_def.node)
    delete_ops_from_graph = delete_ops_from_graph(graph_def, out_md_path)
    print("=== Finish update model with dynamic input size ===")

但是当我尝试将代码组合在一起时(像这样):

from keras.models import load_model
import tensorflow as tf
import os
import json
from keras import backend as K
from tensorflow.python.framework import graph_io
import Unet
from keras.layers import Input

def focal_loss(gamma=2., alpha=.25):
    def focal_loss_fixed(y_true, y_pred):
        pt_1 = tf.where(tf.equal(y_true, 1), y_pred, tf.ones_like(y_pred))
        pt_0 = tf.where(tf.equal(y_true, 0), y_pred, tf.zeros_like(y_pred))
        return -K.sum(alpha * K.pow(1. - pt_1, gamma) * K.log(pt_1+ K.pow(10.0, -9)))-K.sum((1-alpha) * K.pow( pt_0, gamma) * K.log(1. - pt_0+ K.pow(10.0, -9)))
    return focal_loss_fixed

def IOU_calc():
    def IOU_calc_fixed(y_true, y_pred):
        smooth = 0.001
        y_true_f = K.flatten(y_true)
        y_pred_f = K.flatten(y_pred)

        intersection = K.sum(y_true_f * y_pred_f)

        return 2*(intersection + smooth) / (K.sum(y_true_f) + K.sum(y_pred_f) + smooth)
    return IOU_calc_fixed

def IOU_calc_loss():
    return lambda x: -IOU_calc()

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.global_variables()).difference(keep_var_names or []))
        output_names = output_names or []
        print(output_names)
        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)
        # add for new test
        frozen_graph = tf.graph_util.remove_training_nodes(frozen_graph)
        return frozen_graph

def h5_to_pb(h5_model, output_dir, model_name, out_prefix = "output_"):
    out_nodes = []
    for i in range(len(h5_model.outputs)):
        out_nodes.append(out_prefix + str(i + 1))
        tf.identity(h5_model.output[i], out_prefix + str(i + 1))
    # add for ignore dropout at inference process(?)
    K.set_learning_phase(0)

    sess = K.get_session()
    from tensorflow.python.framework import graph_util,graph_io
    init_graph = sess.graph.as_graph_def()
    main_graph = graph_util.convert_variables_to_constants(sess, init_graph, out_nodes)
    graph_io.write_graph(main_graph, output_dir, name = model_name, as_text = False)

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

    for n in graph_def.node:
        if "input_" in n.name:
            input_name = n.name + ":0"

    with tf.Graph().as_default() as graph:
        tf.import_graph_def(graph_def, name='')
                
    return graph, input_name


def delete_ops_from_graph(graph_def, output_model_filepath):

    # Delete nodes
    nodes = []
    input_count = 0
    for node in graph_def.node:
        if "input" in node.name and "dropout" not in node.name:
            if "keras" not in node.name:
                if input_count == 0:
                    nodes.append(node)
                    input_count += 1
                else:
                    print('Drop', node.name)
            else:
                nodes.append(node)
        else:
            nodes.append(node)

    mod_graph_def = tf.GraphDef()
    mod_graph_def.node.extend(nodes)

    # Delete references to deleted nodes
    for node in mod_graph_def.node:
        inp_names = []
        for inp in node.input:
            if 'Neg' in inp:
                pass
            else:
                inp_names.append(inp)

        del node.input[:]
        node.input.extend(inp_names)

    with open(output_model_filepath, 'wb') as f:
        f.write(mod_graph_def.SerializeToString())


if __name__ == '__main__':

    config_path = r'unet_h5_to_pb_config.json'
    with open(config_path) as config_buffer:
        a = config_buffer.read()
        a = a.replace('\\', '\\\\')
        config = json.loads(a)
    """-------------------------Set Paths-----------------------------------"""
    h5_model_path = config["h5_md_path"]

    if "\\" in h5_model_path:
        h5_name = h5_model_path.split("\\")[-1]
        pb_model_name = h5_name.split(".")[0] + ".pb"

        if len(config["pb_save_path"]) == 0:
            output_path = h5_model_path.strip(h5_model_path.split("\\")[-1])
            if h5_model_path[0] == ".":
                output_path = "." + output_path
        else:
            output_path = config["pb_save_path"]
            if os.path.exists(output_path) is False:
                os.mkdir(output_path)
        out_md_path = os.path.join(output_path, pb_model_name)
    else:
        h5_name = h5_model_path.split("/")[-1]
        pb_model_name = h5_name.split(".")[0] + ".pb"

        if len(config["pb_save_path"]) == 0:
            output_path = h5_model_path.strip(h5_model_path.split("/")[-1])
            if h5_model_path[0] == ".":
                output_path = "." + output_path
        else:
            output_path = config["pb_save_path"]
            if os.path.exists(output_path) is False:
                os.mkdir(output_path)
        out_md_path = os.path.join(output_path, pb_model_name)

    """---------------------------load h5 model-----------------------------"""
    K.set_learning_phase(0)
    try:
        net_model = load_model(h5_model_path,
                               custom_objects={'IOU_calc': IOU_calc()})
        print("***Handling model with structure***")
    except ValueError as E:
        if "Cannot create group in read only mode" in str(E):
            print("***Handling model without structure***")
            input_shape = tuple(config["input_shape"])
            model_class = config["model_class"]
            net_model = Unet.small_unet(input_shape, model_class)
            net_model.load_weights(h5_model_path)
        else:
            print(E)

    print('***input is :', net_model.input.name)
    print('***output is:', net_model.output.name)
    """----------------------------save pb file-----------------------------"""
    if config["name_output_node"]:
        h5_to_pb(net_model, output_dir=output_path, model_name=pb_model_name)
    else:
        sess = K.get_session()
        frozen_graph = freeze_session(K.get_session(),
                                      output_names=[net_model.output.op.name])
        graph_io.write_graph(frozen_graph,
                             output_path, pb_model_name, as_text=False)
    print(f"===== FINISH converting model to pb file {out_md_path} =====")

    """----------------------------revise input-----------------------------"""
    K.clear_session()
    if config["dynamic_input"]:
        tf_md_path = os.path.join(output_path, pb_model_name)
        out_pb_path = tf_md_path.replace(".pb", "_dynamic.pb")
        resize_val = (None, None, config["model_inp_ch"])
        
        graph_model, input_name = load_graph(tf_md_path)
        graph_model_def = graph_model.as_graph_def()

        rp_input = Input(resize_val)
        
        tf.import_graph_def(graph_model_def, name='',
                            input_map={input_name: rp_input})

        graph = tf.get_default_graph()
        graph_def = graph.as_graph_def()
        #print(graph_def.node)
        delete_ops_from_graph = delete_ops_from_graph(graph_def, out_pb_path)
        print("=== Finish update model with dynamic input size ===")

改变.pb输入层形状的部分会在输入前加上两层，

"keras_learning_phase/input:0" 和 "keras_learning_phase:0"，输入类型为 bool。

因为如果我分别运行code_1和code_2，就不会出现这种情况，

我觉得可能是keras后端的东西或者K.set_learning_phase(0)(?)引起的

在进入code_2的部分之前，我在code_1的部分之后添加了"K.clear_session()"，

但结果还是一样。

所以我想知道我还能做些什么来避免这种情况发生吗？

(提前感谢您的帮助或建议!)

Answer 1

对此有 2 个解决方案，但都必须在图本身内执行。

最简单的方法是在您从默认图表导入的图表中设置K.set_learning_phase(0)。

如果不行的话。你可以试试这个技巧。因为您正在转换模型的权重。您可以通过移除这些 keras_learning_phase 节点与其他节点的连接来从图中移除这些节点。可以关注this solution

但是，为了给您快速演示:

def remove_keras_learning(od_graph_def):
    for node in od_graph_def.node:
        inp_names = []
        for inp in node.input:
            if not 'keras_learning_phase' in inp:
               inp_names.append(inp)

        del node.input[:]
        node.input.extend(inp_names)
                
    return od_graph_def

结果如下: