NVIDIA Nsight Deep Learning Designer changes in version 2022.2:

• We added support to launch the PyTorch exporter from a virtual environment (Conda or virtualenv).
• We improved the overall performance of Channel Inspector by separating the visualization of per-layer weights from per-layer features and changing the way how we visualize NxCx1x1 weights.
• We added an experimental feature that allows users to directly import existing PyTorch models into NVIDIA Nsight Deep Learning Designer without starting from scratch.
• We switched to using cubic curves to represent layer links to reduce path-finding overhead.
• We added support to visualize inference results of classification networks in the Analysis Mode.
• We added support for custom padding (in addition to the current same and valid options) to the relevant layers.
• We fixed numerous bugs.

We have also decided to remove some uncommonly used operators as we modernize the inference library and unify our model exporters. The following layers are now deprecated and will be removed in a future product release:

• Local response normalization (LRN) layer: Only region values of within are being removed. Normalization across channels is still fully supported.
• Mono-four-stack layer: Replace with a custom layer.
• Mono-to-RGB layer: Replace with a custom layer.
• Network layer: Import the subnetwork as a template instead.
• Output layer: Use of output layers for tensor slicing (the width, height, channels, and offset parameters) is deprecated. Use an explicit slice layer if these operations are required.

The following activations are now deprecated:

• Leaky sigmoid activation: Replace with a custom layer.
• Leaky tanh activation: Replace with a custom layer.
• ReLU activation with alpha values other than zero is being removed. Normal ReLU is still fully supported. The element-wise max layer can replace these activations.

NvNeural changes in version 2022.2:

• Changed the signature of nvneural::XmlNetworkBuilder::createLayerObject to receive the original serialized type used to select the layer object being instantiated. Custom classes deriving from XmlNetworkBuilder must be updated if they override this function.

NVIDIA Nsight Deep Learning Designer changes in version 2022.1:

• Added support to save all tensors in the analysis mode.
• Added support for using nested templates to construct hierarchical network graphs.
• We significantly improved the performance of the type-checking process in the Editor.
• Fixed a bug that prevented PyTorch exports on Linux from succeeding.
• Removed clamping behavior from the Affine layer. It no longer restricts the values of its scale and offset parameters. The options_on parameter has been deprecated; users wishing to hide interactive controls for this layer during analysis should set the new include_ui parameter to false.
• Fixed a bug that blocked FP16 inference when fusing 7x7 convolutions with batch normalizations.

NvNeural changes in version 2022.1:

• Added a new analysis layer: Signal Injector.
• Added a new Input (Constant) layer which supports direct embedding of scalar constants.
• Optimized the performance of the BatchNorm layer.
• Optimized the performance of the Upscale layer.
• Added support for downscaling and fixed-size scaling to the Upscale layer.
• The NvRTC wrapper in nvneural::ICudaRuntimeCompiler has been replaced with a stub when type-checking networks from the GUI. Plugins that rely on the ability to execute generated kernel code during initialization or nvneural::ILayer::reshape should call NvRTC directly, but for performance reasons we do not recommend this approach.
• The forward() function in the exported PyTorch class now takes in keyword-only arguments. User should explicitly name the input paramters while calling the model/function.
• The INetwork::inferenceSubgraph method now applies queued reshape operations. Queued reshapes are not cleared upon failure and will continue to block inference and inferenceSubgraph calls until they succeed.

NVIDIA Nsight Deep Learning Designer changes in version 2021.2:

• The Channel Inspector can display summary and per-channel statistics about a layer's output tensor: mean, minimum/maximum, standard deviation, and sparsity (percentage of tensor elements close to zero).
• Output tensor shapes are now visible during editing.
• ConverenceNG can now save network outputs as .npy files.
• Users can now expand or collapse the parameters list of a layer glyph in the editor view.
• In Channel Inspector, users can now toggle a checkbox to perform auto scale and shift of the channels.
• Users can now save a template as a file that can be imported into another model.

We have added more data to network profiling reports:

• Per-layer device memory footprints
• Whole-network device memory footprint
• Percentage view for inference timings
• Layers' distance from the nearest input, for sorting by network depth
• Template-level inference timings

NvNeural changes in version 2021.2:

• Plugin initialization has been refactored to reduce its reliance on translation-unit-scoped static initialization. The ExportPlugin framework now expects user plugin code to provide an implementation of the function void nvneural::plugin::InitializePluginTypes(). This function should call static ClassRegistry methods to make its export types visible to the client application.
• The SkipConcatenation optimization has been rewritten. Custom concatenation layers should implement nvneural::IConcatenationLayer2 to participate in this optimization.
• We added two new analysis layers: Saliency Generator and Saliency Mix. The Saliency Generator layer converts their input tensors into a single-channel tensor with the same H and W as their inputs. The saliency mix layer simply overlays saliency information (output of a Saliency Generator) onto another input tensor.