Error loading a frozen graph ( float incompatible with float_ref )

[mhaghighat]

I froze the 20170131-234652 model using the freeze_graph.py, but I cannot load it in C++.

I first read the binaryproto successfully as:

tensorflow::GraphDef graph_def;
Status load_graph_status =  ReadBinaryProto(tensorflow::Env::Default(), graph_file_name, &graph_def);

But, it gives an error while creating the graph to be used for the session:

std::unique_ptr<tensorflow::Session> session(tensorflow::NewSession(tensorflow::SessionOptions()));
tensorflow::Status sessionCreateStatus = session->Create(graphDef);

The error is:

Invalid argument: Input 0 of node InceptionResnetV1/Block8/Branch_1/Conv2d_0c_3x1/BatchNorm/cond/AssignMovingAvg_1/Switch was passed float from InceptionResnetV1/Block8/Branch_1/Conv2d_0c_3x1/BatchNorm/moving_variance:0 incompatible with expected float_ref.

Any ideas how to solve this problem?

Thanks in advance :)

[mhaghighat]

BTW, this error also happens when I load the .pb model in Python:

ValueError: graph_def is invalid at node 'InceptionResnetV1/Conv2d_1a_3x3/BatchNorm/cond/AssignMovingAvg/Switch': Input tensor 'InceptionResnetV1/Conv2d_1a_3x3/BatchNorm/moving_mean:0' Cannot convert a tensor of type float32 to an input of type float32_ref.

[Lunrot]

Refer to the following url. https://www.bountysource.com/issues/36614355-unable-to-import-frozen-graph-with-batchnorm

[mhaghighat]

Hi @Lunrot, I've already tried the solution mentioned in this link, but it does not work in our case.

[Lunrot]

@mhaghighat Does the same error occur in python? I have not tested it(freeze & load) in C ++, but there is no error in python.

[mhaghighat]

Yes, it gives this error in Python:

ValueError: graph_def is invalid at node 'InceptionResnetV1/Conv2d_1a_3x3/BatchNorm/cond/AssignMovingAvg/Switch': Input tensor 'InceptionResnetV1/Conv2d_1a_3x3/BatchNorm/moving_mean:0' Cannot convert a tensor of type float32 to an input of type float32_ref.

[Lunrot]

my code

        saver = tf.train.import_meta_graph(os.path.join(os.path.expanduser(args.model_dir), 
            'model-' + os.path.basename(os.path.normpath(args.model_dir)) + '.meta'), clear_devices=True)
        tf.get_default_session().run(tf.global_variables_initializer())
        tf.get_default_session().run(tf.local_variables_initializer())
        saver.restore(sess, tf.train.latest_checkpoint(os.path.expanduser(args.model_dir)))

        output_node_names = 'embeddings'

        # for fixing the bug of batch norm
        gd = sess.graph.as_graph_def()
        for node in gd.node:            
            if node.op == 'RefSwitch':
                node.op = 'Switch'
                for index in xrange(len(node.input)):
                    if 'moving_' in node.input[index]:
                        node.input[index] = node.input[index] + '/read'
            elif node.op == 'AssignSub':
                node.op = 'Sub'
                if 'use_locking' in node.attr: del node.attr['use_locking']
            elif node.op == 'AssignAdd':
                node.op = 'Add'
                if 'use_locking' in node.attr: del node.attr['use_locking']

        converted_graph_def = graph_util.convert_variables_to_constants(sess, gd, output_node_names.split(","))
        tf.train.write_graph(converted_graph_def, args.output_dir, args.output_filename, as_text=False)

[mhaghighat]

@Lunrot:

I had done the same, but the resulted protobuf was not loading successfully. The only difference between my code and yours was that I was feeding sess.graph.as_graph_def() directly to the second argument of the convert_variables_to_constants as in:

output_graph_def = graph_util.convert_variables_to_constants(
                sess, sess.graph.as_graph_def(), output_node_names.split(","))

However, if I change it to the way that you've done, creating gd = sess.graph.as_graph_def(), and then:

output_graph_def = graph_util.convert_variables_to_constants(
                sess, gd, output_node_names.split(","))

there is no error anymore.

I know it sounds absurd, but this is the case!!! The problem is solved but I'm still confused, why???

[Lunrot]

Since gd (=sess.graph.as_graph_def()) has changed in the bug finxing, gd shoud be used instead of sess.graph.as_graph_def().

[mhaghighat]

I had the sess.graph.as_graph_def() in the bug fixing loop as:

for node in sess.graph.as_graph_def().node:  
    [perform all the fixing]

but the resulted protobuf still had the issue.

Maybe, it cannot alter the nodes in the original sess.graph.as_graph_def(); so a copy of it (i.e., gd) needs to be created on which we can perform the bug fixing. Can it be right?!

[ugtony]

@Lunrot, thanks for your code. I made the same mistake as mhaghighat did(also thank @mhaghighat for finding out the difference).

I noted that tf.gfile.GFile is replaced by tf.train.write_graph. What's the difference between the two functions? Can they be used for save/load interchangeably?

I found some of the tensors' shape information are eliminated from the frozen model. For example, the input shape was (?, 160, 160, 3) in the original model but became in the frozen model. It made me unable to use tensor.get_shape() to check the input shape.

For curiosity, I fed the network with some inputs with different shapes: (90, 160, 160, 3), (90, 220, 280, 3), (90, 160, 160, 1) to see if the network can work with any input shape. The last one failed. It means that shapes doesn't imply all shapes are allowed. So now, I have to hard code a proper input size in my program, which is not so convenient. Do you have any idea why the shape information was gone after the freezing operation?

[Lunrot]

@mhaghighat I think so. @ugtony (90, 160, 160, 1) is probably a black and white image. If you change that image to (90,160,160,3), I think it will work correctly.

[ugtony]

Hi @Lunrot, I can understand why (90, 160, 160, 1) doesn't work and why (90, 220, 280, 3) does work for this network. But without shape information, it's not easy to use the frozen model for those who haven't seen inception_resnet_v1.py before.

[Lunrot]

@ugtony The CNN model is just a classifier. It is common to perform image preprocessing (image resizing, cropping, etc.) on CNN input values.

[tengshaofeng]

@mhaghighat hi, can you show me your args? python freeze_graph.py (args). and can you show me the code how you load the frozen model?

[tengshaofeng]

@mhaghighat @Lunrot @ugtony @rtkaleta @scotthong hi, all after Converting model.ckpt to model.pb, how to load the model with the model.pb?

[ugtony]

@tengshaofeng A great tutorial is here.

[mhaghighat]

@tengshaofeng: I just submitted a pull request with the updated freeze_graph.py. This is how I call the function:

python freeze_graph.py ~/models/facenet/20170131-234652 ~/models/facenet/20170131-234652/facenet.pb

For loading the protobuf graph in Python, you can use:

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())

    with tf.Graph().as_default() as graph:
        tf.import_graph_def(
            graph_def, 
            input_map=None, 
            return_elements=None, 
            op_dict=None, 
            producer_op_list=None
        )

    return graph

graph = load_graph('./facenet/facenet.pb')

For loading it in C++, you can use:

tensorflow::GraphDef graphDef;
tensorflow::ReadBinaryProto(tensorflow::Env::Default(), "facenet.pb", &graphDef);

std::unique_ptr<tensorflow::Session> session = tensorflow::NewSession(tensorflow::SessionOptions());
tensorflow::Status sessionCreateStatus = session->Create(graphDef);

[davidsandberg]

This has been fixed in #172.

[tengshaofeng]

@mhaghighat @ugtony thanks so much.

[NicoCoallier]

In my case , I had this error because I was saving the totally of my variable into constant. When I selected only the correct operations in the ouput_node_names, the loading was a success . EX: output_node_names = "Loss/predictions"

[cvJie]

I try to use Facenet by tensorflow C++ API (VS2015),it can load graph,but it doesn't work. Not found: FeedInputs: unable to find feed output phase_train code like this: [tensorflow::Tensor input_tensor(DT_FLOAT, TensorShape({ 2 , iHeight, iWidth, depth })); auto input_tensor_mapped = input_tensor.tensor<float, 4>(); tensorflow::Tensor phase(DT_BOOL, TensorShape()); phase.scalar()() = FALSE; input_tensor.shape(); std::vector<std::pair<std::string, tensorflow::Tensor>> inputs = { { "input", input_tensor }, //it's OK {" phase_train",phase } // it' bad
};

could you give me a hand?thanks

[mhaghighat]

@cvJie: This works for me:

tensorflow::Tensor phaseTrain(tensorflow::DT_BOOL, tensorflow::TensorShape());
phaseTrain.scalar<bool>()() = false;
std::vector<std::pair<std::string, tensorflow::Tensor>> inputs = { 
{ "input", faceTensor } ,
{ "phase_train", phaseTrain } 
};

[arvidzt]

with tf.Session() as sess:
    saved_model_dir = "saved_model_dir_signature"
    meta_graph_def = tf.saved_model.loader.load(sess, [tf.saved_model.tag_constants.SERVING], '')
    for node in sess.graph_def.node:
      if node.op == 'RefEnter':
        node.op = 'Enter'
        for index in range(len(node.input)):
          if 'moving_' in node.input[index]:
            node.input[index] = node.input[index] + '/read'
      if node.op == 'RefSwitch':
        node.op = 'Switch'
        for index in range(len(node.input)):
          if 'moving_' in node.input[index]:
            node.input[index] = node.input[index] + '/read'
      elif node.op == 'AssignSub':
        node.op = 'Sub'
        if 'use_locking' in node.attr: del node.attr['use_locking']
      elif node.op == 'AssignAdd':
        node.op = 'Add'
        if 'use_locking' in node.attr: del node.attr['use_locking']

How can i modify the graph_def in session? if i do it in this way, the model saved by sess didn't change from RefSwitch to Switch. Can someone tell me how to modify the graph_def in sess? thanks.

[Priyashbhugra]

Hello

I am facing this issue

raise ValueError(str(e)) ValueError: Input 0 of node import/global_step/Assign was passed int32 from import/global_step:0 incompatible with expected int32_ref.

in line below code while loading frozen.pb file

    with tf.Graph().as_default() as graph:
        **tf.import_graph_def(
            graph_def,
            input_map=None,
            return_elements=None,
            op_dict=None,
            producer_op_list=None
        )**

here is my full code: model_dir = '/home/priyash/avod/avod/data/outputs/pyramid_cars_with_aug_example/checkpoint_freeze/pyramid_cars_with_aug_example-00120000.pb' log_dir = '/home/priyash/avod/avod/data/outputs/pyramid_cars_with_aug_example/checkpoint_freeze/logs/'

with tf.Session() as sess: model_filename = model_dir with gfile.FastGFile(model_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_map=None,
            return_elements=None,
            op_dict=None,
            producer_op_list=None
        )
    print(graph)
    #
    # g_in = tf.import_graph_def(graph_def)
    # print(g_in)

train_writer = tf.summary.FileWriter(log_dir) train_writer.add_graph(sess.graph)

[glennford49]

FusedBatchNorm/Switch:1 incompatible with expected half error, im trying to convert it to fp16 precision

[xiexie123]

FusedBatchNorm/Switch:1 incompatible with expected half error, im trying to convert it to fp16 precision

hi I have the save error. do you fix it?