daltskin/customvision-tensorflow-csharp

Microsoft Cognitive Services CustomVision.ai - TensorFlow CLI

CustomVision.ai TensorFlow CLI - Offline image classification using C#

Cross platform CLI to run a pre-trained model exported from CustomVision.ai in the Tensorflow format for image classification using the TensorFlowSharp library.

To learn more about Microsoft Cognitive Custom Vision Service, please see here: https://azure.microsoft.com/en-gb/services/cognitive-services/custom-vision-service/

Mean Values (RGB)

Depending on the CustomVision.ai model's Domain, you'll need to set the appropriate values within the code:

Project's Domain	Mean Values (RGB)
General (Compact)	(123, 117, 104)
Landmark (Compact)	(123, 117, 104)
Retail (Compact)	(0, 0, 0)

These are set in BGR eg:

// General (Compact) 
// Landmark (Compact)
var bgrValues = new TFTensor(new float[] { 104.0f, 117.0f, 123.0f });

// Retail (Compact)
var bgrValues = new TFTensor(0f);

CLI Arguments

Argument name	shortcut	example
TensorFlowModelFilePath	-m	Assets\model.pb
TensorFlowLabelsFilePath	-l	Assets\labels.txt
TestImageFilePath	-t	Assets\test.jpg

Usage

• Open the solution
• Restore Nuget Packages
• Run & enjoy

In case you see the following error:

Unhandled Exception: System.DllNotFoundException: Unable to load DLL 'libtensorflow': The specified module could not be found.

Copy the libtensorflow.dll file from the relevant %userprofile%.nuget\packages\tensorflowsharp runtimes folder of your OS into the same folder as the compiled executable.

Image classification

The sample TensorFlow model (exported from CustomVision.ai) is for mushroom classification

CustomVision-TensorFlow.exe -m Assets\model.pb -l Assets\labels.txt -t Assets\test.jpg

Output

Running the above from the command line will display the following:

I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX AVX2
C:\Github\CustomVision-TensorFlow-CSharp\Assets\test.jpg = Agaricus bisporus (92.7477061748505%)
Total time: 00:00:00.4987094

Resources

• Link to Custom Vision Service Documentation
• Link to Custom Vision with Android
• Link to TensorFlow documentation
• Link to TensorFlowSharp
• Link to TensorFlow example code using Python
• Link to TensorFlow example code using C++
• Link to TensorFlow example code using Java

customvision-tensorflow-csharp's Issues

Object detection

This is not an issue, just an idea. How I should handle the object detection from the customvision.ai model? I define this input :
runner.AddInput(graph["Placeholder"][0], tensor).Fetch(graph["model_outputs"][0]);
but I have no idea on how to handle it.

Not working anymore

Have you tried with a new model exported lately?
When using a model exported last year, it works just fine. I tried today with a new model and it doesn't work anymore.
Any ideas?

I'm getting an exception:

Exception thrown: 'TensorFlow.TFException' in TensorFlowSharp.dll
An exception of type 'TensorFlow.TFException' occurred in TensorFlowSharp.dll but was not handled in user code
Input to reshape is a tensor with 97556 values, but the requested shape requires a multiple of 90944
	 [[Node: block0_0_reshape0 = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/replica:0/task:0/device:CPU:0"](block0_0_concat, block0_0_reshape0/shape)]]

Great work

Hello :),

I found your great project, one the way of also loading a Python model for a University project
Maybe you are so kind that you could send it to me
Because at the moment I don't understand why you use the 'loss'
in the CustomVision.cs line 50

Thank you very much for your help in advance

Best Regards
Chris