Migrating to This Release

This topic contains information about changes that may affect existing VisionPro Deep Learning projects when migrating projects from previous releases to the current release.

3.0 → 3.1

Re-optimization Required When Using TensorRT

When using TensorRT with VisionPro Deep Learning 3.1, please note that even if you have already optimized TensorRT with VisionPro Deep Learning3.0, you must optimize it again for VisionPro Deep Learning3.1.

2.1 (2.1.1) → 3.0

Removal of "Extra Large" from Green Classify High Detail Network Models

Due to the Green Classify High Detail architecture update in VisionPro Deep Learning 3.0, the Extra Large option in Network Model parameters for a Green Classify High Detail is removed from VisionPro Deep Learning. If you want to use a Green Classify High Detail in VisionPro Deep Learning 3.0, saved with the Extra Large network size in the previous VisionPro Deep Learning versions, import the tool and change its Network Model parameter from a blank to another one (Small, Normal, or Large).

Removal of Red Analyze High Detail Quick

Due to internal reasons, Red Analyze High Detail Quick is not supported in VisionPro Deep Learning 3.2.1. When you open a Red Analyze High Detail Quick tool trained in an earlier VisionPro Deep Learning version on VisionPro Deep Learning 3.2.1:

This tool will be opened as a Red Analyze High Detail tool.
The trained network of this tool will be lost.
The views and their labels of this tool will be preserved.
The Tool Parameter values will be reset to the default values of a Red Analyze High Detail tool.

Also, importing a Red Analyze High Detail Quick tool trained in an earlier VisionPro Deep Learning version into a VisionPro Deep Learning 3.2.1 workspace is disabled.

High Detail Backwards Compatibility

When you open a Red Analyze High Detail or Green Classify High Detail tool trained in an earlier VisionPro Deep Learning version on VisionPro Deep Learning 3.2.1:

The trained network of this tool will be lost.
The views and their labels of this tool will be preserved.
The Tool Parameter values will be preserved.

When you import a Red Analyze High Detail tool (.vrta) or a Green Classify High Detail tool (.vgta) trained in an earlier VisionPro Deep Learning version into a VisionPro Deep Learning 3.2.1 workspace:

The trained network of this tool will be lost.
The views and their labels of this tool will be preserved.
The Tool Parameter values will be preserved.

Runtime Workspace Backwards Compatibility

The runtime workspace created in the earlier versions of VisionPro Deep Learning 3.0 that contains a Red Analyze High Detail, Red Analyze High Detail Quick, Green Classify High Detail, or Green Classify High Detail Quick can't be imported in VisionPro Deep Learning 3.0. In this case, you need to create a new runtime workspace in VisionPro Deep Learning 3.0.

SuaKIT Segmentation and Classification Projects Compatibility

As it becomes unable to import the High Detail tools created and trained in the earlier versions of VisionPro Deep Learning 3.0, it is also not available to import the trained models (neural networks) of existing segmentation or classification projects from SuaKIT into VisionPro Deep Learning 3.0. See Import a SuaKIT Classification Project and Import a SuaKIT Segmentation Project for more details.

2.0 → 2.1 (2.1.1)

Termination of CIE1 Deployment Support

The CIE1 deployment target is officially no longer supported by VisionPro Deep Learning 2.1 as VisionPro Deep Learning does not support smart cameras. The support for CIE1 deployment target has unofficially been terminated since VisionPro Deep Learning 2.0, but the option was visible in VisionPro Deep Learning GUI in 2.0.

1.1 → 2.0

Termination of Designer Support

The integration with Cognex Designer is no longer supported from VisionPro Deep Learning 3.2.1. For this particular update, the users who are required to use Designer should use VisionPro Deep Learning 1.1.

1.0.1 → 1.1

Change in Validation Loss Calculation

The base of calculating the Validation Loss in High Detail modes (Red Analyze and Green Classify) is changed from 128 iterations to 0.125 epochs. According to this change, the Patience parameter, which was calculated on 128 iterations (N times of 128 iterations) basis, is no longer used in High Detail modes and a new parameter named Patience Epoch is used instead. This new parameter is calculated on an epoch basis (N epochs). For the Green Classify High Detail tools already trained, you have to reconfigure the value of Patience Epoch manually due to this base change.

Loss Tracking in Log Messages

The records of the validation loss in High Detail modes (Green Classify and Red Analyze High Detail) are no longer available in Log Messages as its function of providing the validation loss history is replaced with the Loss Inspector.

Region Filter for Confusion Matrix

For all Red Analyze tools, the region filter is now also applied to the result of the confusion matrix. This application of the region filter is either performed by using filtered() expression.

1.0 → 1.0.1

The workspaces created in VisionPro Deep Learning 1.0.0 are fully compatible with VisionPro Deep Learning 1.0.1

VisionPro ViDi 4.1 → VisionPro Deep Learning 1.0

Architecture of Existing Green Classify Tool

When you open workspaces created from previous version (ViDi Suite) and if there is a Green Classify Tool, you would see as a Green Classify Tool in Focused mode. Existing architecture of Green Classify Tool is called Focused mode and a new architecture of Green Classify Tool is called High Detail mode.

Keeping Existing names of APIs

Since VisionPro Deep Learning inherits ViDi, it retains the name ViDi in many parts of the APIs

Internal version of APIs

Product version is 1.0.0 but, 5.0.0 is used for the version in APIs and binaries.

No longer supported or Deprecated APIs

GPUMode.MultipleDevicesPerTool is no longer supported. For more information, see API Deprecated Features .

Changes to Designer functionality

A project built with the 4.1 (or previous) version of the Designer Deep Learning Plugin is not compatible with the Cognex Deep Learning Plugin. To force the correct plugin to be used, use the Project Plugin Configurator in Designer’s Utilities.

You cannot train in Designer, you can import Cognex Deep Learning Studio Runtime Workspace(.vrws file) created from Cognex Deep Learning Studio 1.0 and process the images.

The Runtime Block includes only one output pin. Individual Tools in the Tool Chain are not visible in the UI. And there are no preconfigured result pins so you need script for every single application.

Deep Learning Runtime Block does not support an External Fixture.

3.4.1 → 4.1

Changes to the Feature Filter Syntax

In the 4.1.0 release, the syntax for the Feature Filter Processing tool parameter was modified. If you were using width and height in your expression, they must now be qualified with view, such as view.width and view.height.

For example, in version 3.2 and earlier, syntax used to limit the found features to those where the center of the feature lies within the ROI would have been the following:

x < width and x > 0 and y > 0 and y < height

In 4.1.0 and later, it must now be expressed as the following:

x < view.width and x > 0 and y > 0 and y < view.height

Note:

This syntax can also more easily be expressed by using the in operator. For example, the following syntax could be used to simplify the expression:

x in [0,view.width] and y in [0,view.height]

3.4 → 3.4.1

Changes to the Viewing of Overlay Graphics

In the 3.4.1 releases, the keyboard shortcuts used to show and hide overlay graphics in the Image Display Area have been changed. The new keystrokes are the following:

The Alt key plus the Right Arrow key or the Alt key plus the Left Arrow key

These are used to cycle through the available combinations of display graphics (labels, markings, labels and markings, or no overlay graphics).

In previous releases, the space bar could be used to show or hide all overlay graphics, but that is no longer supported.

3.3 → 3.4

Changes to Tool Training Workflow

In the 3.4.0 release, there is a new method for specifying the images/views that will be used to train a tool. If you will be re-training a tool that was trained in a previous release, you will need to configure the tool to use the new training set paradigm, via the Select Training Set dialog. Please refer to the Train Neural Network section of each tool.

C Library Overlay Function Deprecation

VisionPro Deep Learning provides two mechanisms, the C API and the .NET API, to display result graphics. However, the following C library overlay graphics functions are being deprecated in this release, and will be removed in the following release:

vidi_training_tool_get_overlay ();
vidi_runtime_sample_get_overlay ();
vidi_runtime_get_overlay ();
vidi_training_sample_get_overlay();

To access the overlay graphics in the future, you will need to use either CognexDesigner, the .NET API or In-Sight VisionPro Deep Learning.

3.2 → 3.3

Importing workspace archives created with version 2.1 and older is not supported. If you are migrating from VisionPro Deep Learning Suite version 2.1 or older, you must first import and then export your workspaces using version 3.0, 3.1, 3.2, or 3.2.1 before you can import them into VisionPro Deep Learning 3.3.0.
VisionPro Deep Learning 3.3.0 requires the use of an NVIDIA driver version greater than or equal to 411.31.

To obtain the updated NVIDIA drivers, visit: https://www.nvidia.com/download/find.aspx

For more information on VisionPro Deep Learning driver compatibility, visit: https://support.cognex.com/en/downloads/detail/deep-learning/3763/1033
The Client/Server functionality in VisionPro Deep Learning 3.3.0 has been updated to use the HTTP communication protocol, and no longer supports the TCP protocol. To use the new HTTP protocol in your client/server application, you must upgrade both the client and server to VisionPro Deep Learning 3.3.0. There is no cross-version client/server compatibility (for example clients running 3.2.1 can only connect to servers running 3.2.1, and clients running 3.3.0 can only connect to servers running 3.3.0).

Note: HTTP communications generally provide better throughput for run-time image processing. However, for high-latency network topologies, HTTP communications may increase interface latency during training and development.

3.1 → 3.2

C Library Function Deprecation

A number of C library functions were deprecated, as follows:

Runtime functions

Note: The deprecated functions are located in a new vidi_runtime_legacy.h header file.

Previous Version	New Version
vidi_runtime_get_overlay	Not available
vidi_runtime_process_sample	vidi_runtime_sample_process
vidi_runtime_get_image	vidi_runtime_sample_get_image

Training functions

Note: The deprecated functions are located in a new vidi_training_legacy.h header file.

Previous Version	New Version
vidi_training_stream_process_sample	vidi_training_sample_process
vidi_training_blue_add_feature	vidi_training_blue_set_feature
vidi_training_blue_add_model	vidi_training_blue_create_model
vidi_training_tool_set_mask	vidi_training_roi_set_mask
vidi_training_tool_get_mask	vidi_training_roi_get_mask
vidi_training_ini_sample	vidi_training_add_sample
vidi_training_process_sample	vidi_training_sample_process

NuGet Package Deprecation

The use of NuGet packages have been deprecated, and are no longer included in the VisionPro Deep Learning installer. If you have runtime and/or training applications that reference the NuGet packages from an earlier VisionPro Deep Learning release, you will need to remove those references, and replace them with references to the .NET .dll libraries, instead.

To remove NuGet packages in Visual Studio, select Tools -> NuGet Package Manager -> Manage NuGet Packages for Solution...

To add the .NET .dlls, in your project, right-click and select Add -> Reference.

VisionPro ViDi 3.0 → 3.1

For the Blue Locate Tool, a change was made in the behavior of X and Y position clamping when using model matching. In version 2.1, these constraints were interpreted to apply to the position of the origin of the model coordinates. In version 3.0, they are interpreted as constraints on the position of the centroid of the nodes in the model.