All the test passed.

Till now the development of pb format model saving has been finished. However, it's still phased. The following features are not supported:

1. The save of checkpoints and assets.
2. The save of signatures (if it's used).
3. Some layers of keras has not been supported (mainly cropping, zero_padding and those related with RNN).

Besides, saving the model training by tf with self-defined model has not been verified. Currently it's recommended to use keras with the pb format model saving.

Then, the following records are for developers of tf.net:

1. A member dependencies was added to SavedObject, which is a class related with protobuf serialization. Some methods were not correspondingly changed such as GetHashCode and MergeFrom. Though everything looks good now, I'm not sure if adding this member would cause some problems in the future.
2. AutoTrackable uses reflection to automatically get the properties which are needed to be tracked. However, the getters of some properties contain extra process. Therefore the reflection may lead to unexpected behavior. It may be considered to Define an attribute to specify the properties to track.
3. In order to realize the feature, two "context managers" are added, SaveContext and SharedObjectSavingScope. However, not only the design may not be the best, but also it's not thread-safe. In the future the thread-safe of these two class should be supported.
4. Break changes were introduced into MySaveableObject. It supports the MySaveableObject contains either Tensor or Variable. The behavior of old APIs of this class was slightly changed.
5. Args and Config of keras were both defined to inherits IKerasConfig, which is somehow misleading. It should be changed if there's a better design in the future.

Hi @AsakusaRinne, thank you for the add of this essential feature.
I see that some layers definitions aren't done for the loading part (and maybe also for the saving part).

In my exemple I use a GlobalAveragePooling2D layer, I don't get any issues during the saving part but I threw an Exeception during the loading part.
The issue occure in "generic_utils.cs" during the "deserialize_keras_object(string class_name, JToken config)" function.
Among others things, argType is null if class_name == "GlobalAveragePooling2D", which seems normal since I didn't see it in the Tensorflow.Binding.Keras.ArgsDefinition.

Did you mind if I add myself the missing feature in the binding (supposing that I just need to take others ArgsDef files as exemple) ? Or else could you add this feature to the project ?

Thank you.

One more question : Is a load of .onnx files supported or did I need to build/use a converter .pb to .onnx and vis versa ?

I add one issue in my case.
After a load I cannot save my model anymore. The learning works correctly.
I got an issue during the return of the _trackable_children() function, saying that KerasAPI already exist in the Dictionnay.
I will look in details the origin of the bug.

@AdrienDeverin Hi, thank you for reporting us this missing of feature. The model loading is now in rapid development and may be unstable. I'll fix your problem once I finish the feature I'm now working on.

One more question : Is a load of .onnx files supported or did I need to build/use a converter .pb to .onnx and vis versa ?

The loading of onnx file is on our schedule but unfortunately is now currently supported, please convert the onnx file to pb first. :)

@AsakusaRinne Hi, thank to you.
I have already build a patch to this problem (code below).

Issue : When we load a model (from a folder with correct .pb file), learn, and save, we get an issue during the saving part

After some research it seem that your save function have a little bug when save or load the _name attribut of Tf.Variable. Basically, you don't get the layer name/type but only is type.
To explain with an exemple, you have this.SharedName = "bias/" and not "dense_1/bias/".

My assumption : After a load, it didn't cause a problem during the learning since you don't use it during it, but it cause a issue when saved a second time as graph variable because you get multiple time the same name when you have more than one layer ("bias/" and "kernel/").

This bug could be temporally corrected by changed the function map_resources() in BaseResourceVariable.cs by something like this :

public override (IDictionary<Trackable, Trackable>, IDictionary<Tensor, Tensor>) map_resources(SaveOptions save_options)
        {
            BaseResourceVariable new_variable;
            if (save_options.experimental_variable_policy.save_variable_devices())
            {
                tf.device(this.Device);
                Debug.Assert(this is ResourceVariable);
                new_variable = resource_variable_ops.copy_to_graph_uninitialized((ResourceVariable)this);
            }
            else
            {
                // Patch here ! I let you do something more optimize (or directly change it at the origin of the problem) //
                string basename = this.Name.Split('/')[0];
                string sharedname = this._name.Split('/')[0]; // _name is currently in protected mode. I set it to public temporally to acces
                if (basename != sharedname)
                {
                    this._name = basename + "/" + this._name;
                }
                ////////////////////////////////////////////////////////////////////////////////////////
                new_variable = resource_variable_ops.copy_to_graph_uninitialized((ResourceVariable)this);
            }
            Dictionary<Trackable, Trackable> obj_map = new();
            Dictionary<Tensor, Tensor> resource_map = new();
            obj_map[this] = new_variable;
            resource_map[this.handle] = new_variable.handle;
            return (obj_map, resource_map);
        }

An other bug come from your way to use a Linq function. In _trackable_children() function of Functionnal you use Concat() and ToDictionnary() which cause a problem when we have multiple time the same element (occure due to the _unconditional_checkpoint_dependencies elements). If we change it by Union() it will resolve the problem.

public override IDictionary<string, Trackable> _trackable_children(SaveType save_type = SaveType.CHECKPOINT, IDictionary<string, IDictionary<Trackable, ISerializedAttributes>>? cache = null)
        {
            return LayerCheckpointDependencies.ToDictionary(x => x.Key, x => x.Value.GetTrackable()).Union(base._trackable_children(save_type, cache) ).ToDictionary(x => x.Key, x => x.Value);
        }

Hope it will help you and others people who get this problem.
Thank you again for all your work, these features were necessary

And for the first issue with GlobalAveragePooling2D Layers, just adding a
GlobalAveragePooling2DArgs.cs file in Tensorflow.Biding.Keras.Pooling resolve it. So it was nothing finally :)

using Newtonsoft.Json;
namespace Tensorflow.Keras.ArgsDefinition
{
    public class GlobalAveragePooling2DArgs : Pooling2DArgs
    {
    }
}