Changelog

Tracking Changes

1.1.8

• Fixed GraphFormat error when creating models.

• Fixed build errors when building for WebGL.

• NatML now requires Unity 2022.3+.

1.1.7

• Fixed JsonException when creating some models (#58).

• Fixed crash when building on Unity Cloud Build with macOS worker (#56).

1.1.6

• Fixed DllNotFoundException when building on Linux (#54).

• Fixed Illegal byte sequence encountered in the input error when building on macOS and Windows (#55).

• Refactored PredictionSession class to EndpointPrediction.

• Refactored NatML.PredictionSessions field to EndpointPredictions.

• Removed MLAudioFeature.FromStreamingAssets method. Use MLUnityExtensions.StreamingAssetsToAbsolutePath to get a path and create an audio feature.

• Deprecated PredictorSession class.

• Deprecated NatML.PredictorSessions field.

1.1.5

• Added support for Safari 16.4, bringing NatML to 88% of browsers.

• Added MLArrayFeature.CopyTo method overload that accepts a Texture2D.

• Added MLEdgeModel.Create method overload that accepts a NatMLClient instance in place of an accessKey.

• Added MLCloudModel.Create method overload that accepts a NatMLClient instance in place of an accessKey.

• Added Feature.dictValue field for working with dictionary features.

• Refactored Dtype.List enumeration member to Dtype.List.

• Refactored Dtype.Dictionary enumeration member to Dtype.Dict.

• Removed MLCloudModel.Predict method overload that accepts a Dictionary<string, MLFeature>.

• Removed MLImageFeature.RegionOfInterest method.

1.1.4

• Added NatML.API namespace for accessing the full NatML web API from .NET.

• Fixed MLEdgeModel.Create method blocking for a long time on Android and Windows (#49).
• Fixed MLEdgeModel.Create method throwing error when called from non-Unity threads.
• Updated default edge prediction configuration to use the CPU and neural processor (and not the GPU) on Android (#49).
• Refactored MLDepthFeature.ViewportToWorldPoint method to Unproject.
• Removed MLEdgeModel.AudioFormat struct. Use NatML.API.Types.AudioFormat class instead.

• Removed MLEdgeModel.Normalization struct. Use NatML.API.Types.Normalization class instead.

• Removed MLImageFeature.AspectMode enumeration. Use NatML.API.Types.AspectMode enumeration instead.

• Removed MLAudioType.FromAudioClip utility method.

• Removed MLAudioType.FromVideoClip utility method.

• Removed MLVideoType.FromVideoClip utility method.

1.1.3

• Added support for making edge predictions on the GPU on Android with MLEdgeModel.ComputeTarget.GPU.
• Fixed MLEdgeModel.Predict sporadically crashing on Android.

1.1.2

• Added support for CoreML models that consume pixel buffers (CVPixelBufferRef) instead of multi-arrays.

• Fixed MLEdgeModel.Create ignoring configuration argument when model is fetched from NatML Hub.
• Fixed MLAsyncPredictor accumulating massive amounts of memory on iOS and macOS.
• Refactored MLTextFeature class to MLStringFeature.
• Refactored MLTextType class to MLStringType.
• Refactored MLModelDataEmbed attribute to MLEdgeModel.Embed.

1.1.1

• Added model embedding support for encrypted models.

1.1.0

• Added MLCloudModel class for making predictions with predictor endpoints.
• Added MLEdgeModel.Create methods for creating an edge model from a predictor tag or an MLModelData instance.
• Added MLEdgeModel.ComputeTarget enumeration for specifying the compute target for model predictions.
• Added MLEdgeModel.Configuration data class for configuring the edge model creation process.
• Added MLEdgeModel.labels property for inspecting classification labels required by the model.
• Added MLEdgeModel.aspectMode property for inspecting the aspect mode required in image data consumed by the model.
• Added MLEdgeModel.audioFormat property for inspecting the audio format required in audio data consumed by the model.
• Added support for model encryption, protecting your valuable intellectual property.

• Added support for pinning MLArrayFeature instances with fixed statements.
• Added support for pinning MLImageFeature instances with fixed statements.

• Added MLArrayFeature constructor from an MLCloudFeature for making predictions with predictor endpoints.
• Added MLImageFeature constructor from an MLCloudFeature for making predictions with predictor endpoints.

• Added MLTextFeature constructor from an MLCloudFeature for making predictions with predictor endpoints.

• Added MLArrayFeature.CopyTo(MLArrayFeature) method for copying image data from one image feature to another.
• Added MLImageFeature.CopyTo(MLImageFeature) method for copying image data from one image feature to another.

• Added implicit conversion from float to MLFeature.
• Added implicit conversion from int to MLFeature.
• Added implicit conversion from bool to MLFeature.
• Added implicit conversion from int[] to MLFeature.
• Updated MLImageFeature class to be sealed thus preventing inheritance.
• Updated float[] implicit conversion to MLArrayFeature to have a 1D shape instead of a null shape.
• Refactored MLModelData.ComputeTarget enumeration to MLEdgeModel.ComputeTarget.
• Refactored MLModelData.Normalization enumeration to MLEdgeModel.Normalization.
• Refactored MLModelData.AudioFormat enumeration to MLEdgeModel.AudioFormat.
• Deprecated MLImageFeature.RegionOfInterest method. Use MLImageFeature.CopyTo method instead.

• Deprecated MLModelData.ComputeTarget.CPUOnly enumeration member. Use ComputeTarget.CPU instead.
• Removed MLEdgeModel constructor. Use MLEdgeModel.Create method instead.
• Removed MLModelData.Deserialize method. Use MLEdgeModel.Create method instead.
• Removed MLModelData.FromHub method. Use MLEdgeModel.Create method instead.
• Removed MLModelData.FromFile method. Use MLEdgeModel.Create method instead.
• Removed MLModelData.tag property.

• Removed MLModelData.computeTarget property. Use MLEdgeModel.Configuration.computeTarget property instead.
• Removed MLModelData.computeDevice property. Use MLEdgeModel.Configuration.computeDevice property instead.
• Removed MLModelData.labels property. Use MLEdgeModel.labels property instead.
• Removed MLModelData.normalization property. Use MLEdgeModel.normalization property instead.
• Removed MLModelData.audioFormat property. Use MLEdgeModel.audioFormat property instead.
• Removed MLModelData.ComputeTarget enumeration.

• Removed MLArrayFeature.CopyTo(T[]) method overload.

• Removed MLArrayFeature.CopyTo(NativeArray<T>) method overload.

• Removed MLArrayFeature.CopyTo(T*) method overload.

• Removed MLImageFeature.CopyTo(T[]) method overload.

• Removed MLImageFeature.CopyTo(NativeArray<T>) method overload.

• Removed MLImageFeature.CopyTo(T*) method overload.

• NatML now requires iOS 14+.

1.0.19

• Added MLEdgeModel public constructor and moved the class to the top-level NatML namespace.
• Fixed app storage size increasing every time that MLModelData.Deserialize was invoked on iOS.

• Deprecated MLModelData.Deserialize method. Use MLEdgeModel constructor instead.

1.0.18

• NatML now defaults to accelerating predictions on Windows by using the GPU. This results in much better performance and lower CPU usage for many models.

• Added MLModelData.computeTarget property for specifying the compute target used for prediction.

• Added MLModelData.computeDevice property for specifying the compute device used for prediction.

• Added MLImageFeature.CopyTo overloads that accepted pixel buffers in managed and unmanaged memory.
• Fixed MLImageFeature producing incorrect prediction results on WebGL.
• Fixed rare crash when calling MLModelData.Deserialize on low-end Android devices.

1.0.17

• Upgraded to Hub 1.0.12.

1.0.16

• Refactored MLDepthFeature.TransformPoint method to ViewportToWorldPoint.
• Removed MLImageFeature.CopyTo overloads that accepted pixel buffers. Use Texture2D overload instead.

• Removed MLImageFeature.ToTexture method. Use MLImageFeature.CopyTo method instead.

1.0.15

• Added initial support for WebGL! NatML can now be used in the browser.

• Improved MLImageFeature texture constructor to avoid copying pixel buffers when possible.

• Improved performance of MLImageFeature.NonMaxSuppression for large number of candidate boxes.

• Changed MLEdgeFeature.dataType property type from System.Type to NatML.DataType.

1.0.13

• Added support for rotated ROI in MLImageFeature.RegionOfInterest method.
• Added MLModelData.FromFile method to load ML model data from model files.

• Added MLImageFeature.TransformPoint method for transforming detection points from feature space to image space.
• Added MLDepthFeature.TransformPoint method for projecting 2D points into 3D space using depth.
• Added MLEdgeFeature.dataType property for inspecting the data type of Edge features.

• Added MLAsyncPredictor class for making predictions on a background thread.
• Added MLPredictorExtensions.ToAsync extension method for converting predictor to an async predictor.
• Improved prediction performance on Android devices with dedicated neural processing units.

• Changed MLEdgeFeature class to readonly struct to prevent GC pressure.

• Fixed MLImageFeature not respecting AspectMode.AspectFit when making predictions.
• Fixed sporadic NullReferenceException when MLModelData.FromHub is called on some Android devices.

• Updated MLDepthFeature.Sample method to accept a Vector2 point instead of individual coordinates.
• Removed IMLCloudFeature interface as it is no longer supported by the NatML Hub API.

• Removed MLCloudFeature class as it is no longer supported by the NatML Hub API.

• Removed MLCloudModel class as it is no longer supported by the NatML Hub API.

• Removed IMLAsyncPredictor interface.

• Removed MLFeature.CloudType utility method as it is no longer supported by the NatML Hub API.

• Removed MLArrayFeature constructor that accepted an MLCloudFeature.

• Removed MLAudioFeature constructor that accepted an MLCloudFeature.

• Removed MLImageFeature constructor that accepted an MLCloudFeature.

• Removed MLImageFeature constructor that accepted an encoded image byte[].

• Removed MLImageFeature.Contiguous method.

• Removed MLTextFeature constructor that accepted an MLCloudFeature.

• Removed MLEdgeFeature.ReleaseFeature method as it has long been deprecated.

1.0.12

• Upgraded to Hub 1.0.8.

1.0.11

• Added MLModelDataEmbed attribute for embedding model data at build time, making models immediately available in builds without downloads.

• Added MLImageFeature.NonMaxSuppression method for performing non-maximum suppression on detection proposals.
• Added MLImageFeature.TransformRect method for transforming detection rectangles from feature space to image space.
• Added custom icon for identifying ML model files imported by NatML.

• Migrated MLPredictorExtensions.ToAsync extension method and MLAsyncPredictor class to NatMLX.
• Removed MLPredictorExtensions.RectifyAspect method. Use MLImageFeature.TransformRect method instead.

• Removed MLPredictorExtensions.NonMaxSuppression method. Use MLImageFeature.NonMaxSuppression method instead.

• Refactored top-level namespace from NatSuite.ML to NatML for parity with our other API's.

1.0.10

• Improved prediction performance on Windows systems with dedicated GPU's.

• Added MLArrayFeature constructors that accept NativeArray<T> native arrays to minimize memory copies.
• Added setter accessors to MLArrayFeature indexers allowing for writing values to feature data.
• Added MLAudioFeature.Contiguous method for decoding encoded audio feature into memory.

• Added support for creating greyscale image MLEdgeFeature features from MLImageFeature features.
• Added MLImageFeature.RegionOfInterest method for extracting an ROI from an image.
• Added MLImageFeature.Contiguous method for decoding encoded image feature into memory.

• Added MLDepthFeature.width convenience property for getting width of depth features.
• Added MLDepthFeature.height convenience property for getting height of depth features.
• Added MLImageType.interleaved property for checking whether image feature is interleaved or planar.
• Added IMLCloudFeature interface to create cloud ML features for making cloud predictions.

• Added "Clear Predictor Cache" menu item for clearing predictor cache in the editor.

• Added NatMLHub.Subscribe method for making subscription requests to the NatML API.

• Fixed memory leak when using certain vision predictors like Robust Video Matting.

• Refactored MLHubModel class to MLCloudModel.

• Refactored MLHubFeature class to MLCloudFeature.

• Refactored HubDataType class to DataType.

• Deprecated IMLHubFeature interface.

• Removed MLAudioFeature.ReadToEnd method. Use MLAudioFeature.Contiguous method instead.

1.0.9

• Added exclusive support for running CoreML graphs on iOS and macOS.

• Added exclusive support for running TensorFlow Lite graphs on Android.

• Added support for working with CoreML .mlmodel files in Unity projects.

• Added support for working with TensorFlow Lite .tflite files in Unity projects.

• Added support for Apple Silicon on macOS.

• Added MLVideoFeature class for making ML predictions on video files.
• Added MLVideoType feature type for inspecting video features.
• Added MLDepthFeature abstract class for working with predictors that use depth data.
• Added support for audio feature resampling in MLAudioFeature with sampleRate and channelCount fields.
• Added MLEdgeFeature type for added type safety when authoring edge predictors.

• Added MLImageFeature constructor which accepts a NativeArray<byte> pixel buffer.
• Added MLImageFeature.width convenience property for getting width of image feature.
• Added MLImageFeature.height convenience property for getting height of image feature.
• Added MLImageFeature.CopyTo methods for copying pixel data from image feature.
• Added MLAudioFeature constructor that accepts a NativeArray<float> to minimize memory copies.
• Added MLAudioFeature path constructor for reading audio features from audio and video files.

• Added MLAudioFeature.CopyTo methods for copying audio data from audio feature.

• Added MLAudioFeature.FromStreamingAssets method for creating an audio feature from an audio file in the StreamingAssets folder.

• Added MLAudioFeature.ToAudioClip method for converting audio feature to an AudioClip.

• Added MLAudioFeature.ReadToEnd method to read audio data into memory for audio features backed by an audio file.

• Added MLAudioType.FromFile method for inspecting the audio type of a video or audio file.
• Added MLAudioType.FromAudioClip method for inspecting the audio type of an audio clip.
• Added MLAudioType.FromVideoClip method for inspecting the audio type of a video clip.
• Added MLAudioType.FromStreamingAssets method for inspecting the audio type of a video file in the StreamingAssets folder.
• Added MLImageType constructor that accepts image channels.
• Fixed MLImageFeature type incorrectly reporting 3 channels instead of 4.
• Fixed MLImageFeature default normalization standard deviation having 0 for alpha channel.
• Removed IMLModel interface as it has long been deprecated.

• Removed IMLFeature interface as it has long been deprecated.

1.0.8

• Fixed Cannot deserialize graph exception when deserializing cached predictors.

• Fixed MLModelData being cached for DRAFT predictors.

1.0.7

• Added MLArrayFeature constructor that accepts an MLHubFeature for working with Hub predictors.
• Added MLAudioFeature constructor that accepts an MLHubFeature for working with Hub predictors.
• Added MLImageFeature constructor that accepts an MLHubFeature for working with Hub predictors.
• Added MLTextFeature constructor that accepts an MLHubFeature for working with Hub predictors.
• Fixed DirectoryNotFoundException when loading cached MLModelData on iOS.

• Removed prediction analytics reporting to NatML Hub, relieving network bandwidth pressure.

• Removed cache flag in MLModelData.FromHub method.

1.0.6

• Introduced Hub Predictors, which make predictions using server-side processing on NatML Hub.
• Added MLHubModel class for authoring predictors that make cloud-based predictions using NatML Hub.

• Added MLEdgeModel class for authoring predictors that make edge (on-device) predictions.
• Added IMLHubFeature interface for creating server-side features when making predictions with NatML Hub.

• Added IMLEdgeFeature interface for creating native features when making edge (on-device) predictions.

• Added MLTextType feature type for inspecting MLTextFeature instances.
• Added MLModelData.tag property to identify the predictor tag from NatML Hub.
• Added MLModel.metadata dictionary for inspecting model metadata.
• Added MLArrayFeature.Squeeze to remove singleton dimensions from an array feature.

• Added MLArrayFeature.Flatten to flatten an array feature into one-dimensional array feature.

• Added MLArrayFeature.ToArray to convert an array feature into a flattened primitive array.
• Added MLImageFeature.ToTexture to convert an image feature into a Texture2D.
• Added MLImageType.FromType static method for converting arbitrary feature types to image types.
• Added implicit conversion from MLFeatureType to bool indicating if the type is non-null.
• Added implicit conversion from MLTextFeature to string.
• Fixed MLImageType image resolution constructor assuming planar format instead of interleaved format.
• Moved IMLPredictor interface to the top-level NatSuite.ML namespace.
• Moved IMLAsyncPredictor interface to the top-level NatSuite.ML namespace.

• Deprecated IMLModel interface. Cast model to MLEdgeModel class instead.
• Deprecated IMLFeature interface. Cast feature to IMLEdgeFeature interface instead.

• Deprecated MLPredictorExtensions.GetImageSize static method. Use MLImageType.FromType instead.
• Removed MLModelData.FromFile method. Use NatML Hub instead.
• Removed MLModelData.FromStreamingAssets method. Use NatML Hub instead.
• Removed MLPredictorExtensions.SerializeAudio method.

• Removed MLPredictorExtensions.SerializeImage method.

• Removed MLModel dictionary indexers. Use MLModel.metadata property instead.

1.0.5

• Changed MLImageFeature.mean and std types to Vector4 to support normalization for alpha channel.
• Fixed bitcode not being generated for iOS NatML.framework.

• Removed metadata accessors from IMLModel interface. Cast to MLModel instead.

1.0.4

• Added IMLAsyncPredictor interface for making server-side ML predictions with NatML Hub.

• Added MLArrayFeature<T> constructor which accepts a native array feature for easy interop.
• Added multi-indexing support to MLArrayFeature<T> for post-processing native array features.
• Added MLArrayType.elementCount property to get the total number of elements for an array type.
• Added MLArrayFeature<T>.shape property which returns the feature type's shape for convenience.
• Added MLArrayFeature<T>.elementCount property which returns the feature type's element count for convenience.
• Added MLArrayFeature<T>.CopyTo method to copy feature data into an array.
• Added MLArrayFeature<T>.Permute method to create a shallow array view with permuted dimensions.

• Added MLArrayFeature<T>.View method to create a shallow array view with a different shape.

• Added MLPredictorExtensions.NonMaxSuppression method for working with detection models.
• Added MLPredictorExtensions.GetImageSize method for making predictions with image features.

• Added MLPredictorExtensions.SerializeAudio method for making Hub predictions with audio features.

• Added MLPredictorExtensions.SerializeImage method for making Hub predictions with image features.

• Fixed MLAsyncPredictor predictions never completing if backing predictor encountered exception.

1.0.3

• Greatly improved performance and memory pressure when performing multi-indexing with MLArrayFeature<T>.
• Added MLPredictorExtensions.RectifyAspect extension method for correcting detection rects from aspect-scaled images.
• Fixed crash when making predictions with recurrent models on previous state features.

• Fixed crash when getting native array feature shape for MLArrayFeature<T>.
• Fixed memory leak when making predictions with image features on iOS and macOS.

1.0.2

• Added MLModelData.audioFormat property for working with audio and speech ML models.

• Added MLTextFeature for working with natural language processing models.
• Exposed mean and std arrays in MLModelData.Normalization struct for models that require arbitrary normalization.

• Removed generic MLClassificationPredictor and MLDenseClassificationPredictor predictors.

• Removed ability to specify class labels for local .onnx file in project. Use NatML Hub instead.

1.0.0

• First release.