Error when saving model using model.save

[le-tan-phuc]

Description

Hi, I'm new to Tensorflow.net. I'm just playing around with the "Toy version of ResNet in Keras" example on the main page and got an error at the model.save("./toy_resnet_model");

C# threw an exception: System.InvalidOperationException: 'Collection was modified; enumeration operation may not execute.'

I tried to debug and trace the problem and it seems like the exception was thrown at some point within this function:

(MetaGraphDef, Graph, TrackableSaver, AssetInfo, IList<Trackable>, IDictionary<Trackable, IEnumerable<TrackableReference>>) tuple = _build_meta_graph(obj, signatures, options, metaGraphDef);

which is part of

(saved_nodes, node_paths) = SavedModelUtils.save_and_return_nodes(model, filepath, signatures, options);

which is part of

KerasSavedModelUtils.save_model(this, filepath, overwrite, include_optimizer, signatures, options, save_traces);

Package installed:

• TensorFlow.NET 0.150.0
• TensorFlow.Keras 0.15.0
• SciSharp.TensorFlow.Redist 2.16.0

Any help would be appreciated! Thank you.

[AdrienDeverin]

I encountered the same problem months ago. #1017
My conclusion was that some layers (notably the Cropping layer in my case) weren't managed properly. Could you provide the model that caused the problem?

[le-tan-phuc]

Hi @AdrienDeverin, I just put the exact example from the TensorFlow.NET GitHub page to try out and face this problem. I attached the code again here for easy reference, and the error occurs at the last line, which is the model.save. I have also tried changing the save_format from "tf" to "h5" and it runs without error, but nothing was saved:

using static Tensorflow.Binding;
using static Tensorflow.KerasApi;
using Tensorflow;
using Tensorflow.NumPy;

var layers = keras.layers;
// input layer
var inputs = keras.Input(shape: (32, 32, 3), name: "img");
// convolutional layer
var x = layers.Conv2D(32, 3, activation: "relu").Apply(inputs);
x = layers.Conv2D(64, 3, activation: "relu").Apply(x);
var block_1_output = layers.MaxPooling2D(3).Apply(x);
x = layers.Conv2D(64, 3, activation: "relu", padding: "same").Apply(block_1_output);
x = layers.Conv2D(64, 3, activation: "relu", padding: "same").Apply(x);
var block_2_output = layers.Add().Apply(new Tensors(x, block_1_output));
x = layers.Conv2D(64, 3, activation: "relu", padding: "same").Apply(block_2_output);
x = layers.Conv2D(64, 3, activation: "relu", padding: "same").Apply(x);
var block_3_output = layers.Add().Apply(new Tensors(x, block_2_output));
x = layers.Conv2D(64, 3, activation: "relu").Apply(block_3_output);
x = layers.GlobalAveragePooling2D().Apply(x);
x = layers.Dense(256, activation: "relu").Apply(x);
x = layers.Dropout(0.5f).Apply(x);
// output layer
var outputs = layers.Dense(10).Apply(x);
// build keras model
var model = keras.Model(inputs, outputs, name: "toy_resnet");
model.summary();
// compile keras model in tensorflow static graph
model.compile(optimizer: keras.optimizers.RMSprop(1e-3f),
    loss: keras.losses.SparseCategoricalCrossentropy(from_logits: true),
    metrics: new[] { "acc" });
// prepare dataset
var ((x_train, y_train), (x_test, y_test)) = keras.datasets.cifar10.load_data();
// normalize the input
x_train = x_train / 255.0f;
// training
model.fit(x_train[new Slice(0, 2000)], y_train[new Slice(0, 2000)],
            batch_size: 64,
            epochs: 10,
            validation_split: 0.2f);
// save the model
model.save("./toy_resnet_model");

[AdrienDeverin]

Have you try with a direct path folder (folder need to be created before) ? Example : @"C:\GitHub\TestTF\ResnetModel" Normally you get .pb file in it after

[le-tan-phuc]

Thanks for your prompt response. I've tried to put a full path folder, but the problem remains.
This is what is shown in the output log:

Exception thrown: 'System.InvalidOperationException' in mscorlib.dll
An unhandled exception of type 'System.InvalidOperationException' occurred in mscorlib.dll
Collection was modified; enumeration operation may not execute.

[AdrienDeverin]

It's really strange. I tried it myself an no problem appear. Everything go well.

Try with this config (normally it doesn't matter, your seems good) :

• Dependancy : SciSharp.TensorFlow.Redist (2.16.0)
• directly download the Tensorflow.Net-master repo and add project Tensorflow.Keras and Tensorflow.Biding (Tensorflow.Core.sln) to your solution. Add them as reference of your test program.

My import :

using Tensorflow;               
using Tensorflow.NumPy;   
using Tensorflow.Keras;
using Tensorflow.Keras.Layers;
using Tensorflow.Keras.Saving;
using Tensorflow.Keras.Engine; 
using Tensorflow.Keras.Losses;
using Tensorflow.Keras.Utils;
using Tensorflow.Keras.ArgsDefinition;
using Tensorflow.Keras.ArgsDefinition.Reshaping;
using Tensorflow.Operations.Activation;
using Tensorflow.Operations.Initializers;
using Tensorflow.Common.Types;
using static Tensorflow.KerasApi;
using static Tensorflow.Binding; 
using static Tensorflow.ops;
using static Tensorflow.ApiDef.Types;

[AsakusaRinne]

Hi, I have tried your code but I failed to reproduce it. I ran it and everything seemed to go well.

The only difference between our code is that I changed epoch to 1 and batch to 4 to make it faster to complete the training. I guess that doesn't matter.

P.S. I was using the CPU redist package.

[le-tan-phuc]

Hi @AdrienDeverin and @AsakusaRinne, thank you both for your help. Let me explain what exactly happened: I have 2 computers, A and B. I started trying the example code on computer A, where I created a new project with .NET Framework 4.8. Then I faced the problem described at the beginning. I tried different stuff as suggested by @AdrienDeverin but it didn't solve the problem. Later on, I created a new project based on .NET 8.0 on computer A, and it worked like a charm. I tried to replicate the solution on computer B, also with .NET 8 and all the same Nuget packages installed. It now gives me a different error on the model.save:

System.NotImplementedException: ''

Have you experienced this before? This is pretty confusing for me.

Ps: this is the project properties for your reference. I'm using VS2022 V17.9.6

[AsakusaRinne]

I didn't manage to reproduce it on my PC. Could you please clone the repo and add project reference to it, so that a detailed trace back will be shown?

[AdrienDeverin]

Me too, I tried to reproduce your bug, but it's working correctly on my computer... :/
(To add more, I was testing it in .NET 6.0)

Another idea to understand where the problem lies: since you have done what I said earlier, you could use the debug mode and see in the code step by step where you go (and compare with computer A)...

[le-tan-phuc]

Hi all, thank you for your support. I've tried again with a fresh project on computer B based on .NET 8.0 in a local folder and it works perfectly now. I guess the previous problem on System.NotImplementedException: maybe because the project folder was placed in Onedrive of computer A, got synced to computer B and somehow it ends up with that error when run on computer B because some files are missing.
I also tried with .NET 6.0 as @AdrienDeverin , and it worked as well now.
Nevertheless, a new project based on .NET Framework 4.5 still has the original problem, maybe it's not supported.

[AsakusaRinne]

Nevertheless, a new project based on .NET Framework 4.5 still has the original problem, maybe it's not supported.

It's expected to support .NET framework 4.5. Could you please run with the tf.net repo and put the detailed traceback here if you'd like to dig on it?

[le-tan-phuc]

I was figuring out a solution to make .NET Framework-based app work, but not sure if this is a bug or anything. Let me detail the process so that someone facing the same issue knows how to get through it.
The original problem was:

1. I created a new c# winform app using .NET Framework 4.8, and set it to work with x64 only.
2. Installed nuget packages: TensorFlow.NET, SciSharp.TensorFlow.Redist, and TensorFlow.Keras. To successfully install TensorFlow.Keras, I needed to install the PureHDF separately first (by checking the Include prerelease).
3. Copy the example from the main SciSharp github page
4. Copy the tensorflow.dll from SciSharp.TensorFlow.Redist.2.16.0 package into the debug folder (to clear the backend not found exception)
5. Got the System.InvalidOperationException: 'Collection was modified; enumeration operation may not execute.' at the model.save

How did I get it work:

1. Download the entire TensorFlow.Net repo -> create a new .NET Framework 4.8 project within the TensorFlow.NET solution.
2. Installed nuget packages: TensorFlow.NET, SciSharp.TensorFlow.Redist, and TensorFlow.Keras. This is to get all the dependencies to be installed. After that, uninstall TensorFlow.NET and TensorFlow.Keras.
3. Copy the tensorflow.dll into the debug folder
4. In my project, add a reference to the Tensorflow.Binding and Tensorflow.Keras from the repo.
5. The application works smoothly now without error.

When I checked the output debug folder, I noticed the size difference in the Tensorflow.Binding.dll and Tensorflow.Keras.dll between the original and the solution. Copying these two files from the updated solution folder to the previous project folder solve the error too. Thus, I guess there should be some differences in the Tensorflow.Binding.dll and Tensorflow.Keras.dll between the release NuGet packages and the repo. Do you have any idea on this @AsakusaRinne?

[Oceania2018]

It might be the NuGut package is not up to date.

[abrguyt]

I was figuring out a solution to make .NET Framework-based app work, but not sure if this is a bug or anything. Let me detail the process so that someone facing the same issue knows how to get through it. The original problem was:

1. I created a new c# winform app using .NET Framework 4.8, and set it to work with x64 only.

2. Installed nuget packages: TensorFlow.NET, SciSharp.TensorFlow.Redist, and TensorFlow.Keras. To successfully install TensorFlow.Keras, I needed to install the PureHDF separately first (by checking the Include prerelease).

3. Copy the example from the main SciSharp github page

4. Copy the tensorflow.dll from SciSharp.TensorFlow.Redist.2.16.0 package into the debug folder (to clear the backend not found exception)

5. Got the System.InvalidOperationException: 'Collection was modified; enumeration operation may not execute.' at the `model.save`

How did I get it work:

1. Download the entire TensorFlow.Net repo -> create a new **.NET Framework 4.8** project within the TensorFlow.NET solution.

2. Installed nuget packages: TensorFlow.NET, SciSharp.TensorFlow.Redist, and TensorFlow.Keras. This is to get all the dependencies to be installed. After that, uninstall TensorFlow.NET and TensorFlow.Keras.

3. Copy the tensorflow.dll into the debug folder

4. In my project, add a reference to the Tensorflow.Binding and Tensorflow.Keras from the repo.

5. The application works smoothly now without error.

When I checked the output debug folder, I noticed the size difference in the Tensorflow.Binding.dll and Tensorflow.Keras.dll between the original and the solution. Copying these two files from the updated solution folder to the previous project folder solve the error too. Thus, I guess there should be some differences in the Tensorflow.Binding.dll and Tensorflow.Keras.dll between the release NuGet packages and the repo. Do you have any idea on this @AsakusaRinne?

Indeed doing a custom build with the official Repo solves the model.save() exception. After tinkering around I found two more structural alternative solutions:

Option 1) Just add only the Tensorflow.Keras project from the Repo (after download/unpacking) to your own solution, and then reference that added (external) project from your own project in your own solution. That way all necessary dependencies (from both Tensorflow.Keras and Tensorflow.NET) will be included in your debug/release folders.

alternative:

Option 2) Create your own local NuGet packages from a fresh build:

• change the PackageId of the Tensorflow.Keras and Tensorflow.Binding projects (in their respective project file) to a custom value - e.g. just add a postfix like '.custom' (to avoid Package Id clashes with the official NuGet repo)
• package both the Tensorflow.Keras and Tensorflow.Binding projects (right click 'package' in VS 2022)
• store both resulting .nupkg package files (stored in the /packages folder of the project) in your local custom NuGet folder
• add this local custom NuGet folder as a NuGet source in VS 2022
• install Tensorflow.Keras.custom (or whatever postfix you chose) from your local NuGet source folder into your project
• (make sure your project has a reference to tensorflow.dll from SciSharp.TensorFlow.Redist)

Option 1) is easier but requires that you add the Repo Tensorflow.Keras project to every solution/project. Option 2) is cleaner as it's easily re-usable for new solutions/projects - it requires minimal installation (only the local Tensorflow.Keras NuGet package + tensorflow.dll ref) for new development.

Thanks so much for putting me on the right trail for a solution. Of course it would be most wonderful if SciSharp could just resubmit a new working build as a package to NuGet - all these workarounds would be unnecessary. The Repo code itself seems to be fine.