Releases: zetane/viewer
v1.7.4
Viewer Engine
- Added support for
ONNX 1.10.2 - Added support for
ONNX Runtime 1.10.0 - Added support for
Keras/TensorFlow 2.9.1 - Improved progress notifications when loading Keras models
- Fixed crash cause by nested Keras models.
- Reduced Tensor viewer memory usage
- Dropped support for Ubuntu 16.04 LTS. See the up-to-date Minimum Requirements.
- Deprecated support for macOS 10.14 Mojave
API
- Added the Zetane API context manager to automate view updates and cleanup, resulting in less verbose code.
- Added support for
Python 3.9 - Dropped support for
Python 3.6 - Fixed
protobufdependency versioning
1.7.0
1.6.2
- Added output blocks for models to prevent navigation to the end of the model graph
- Added a Top-K output view for tensors that match certain shapes, e.g. (1, N). Classification models now have a more human understandable output.
- Update to onnxruntime 1.8.1 to support latest ONNX opset.
- Improve autodetection of input shapes to allow more inputs to pass inference without shape errors.
- Fixes for RAM overuse
- Fixes for Mesh API
1.5.0
1.4.0
Release 1.4.0 adds many new features, including support for native Keras models!
- Load .h5 models and properly shaped inputs for Keras models to view internal tensors and tensor metrics.
- Load ztn snapshots with sample inputs
- A number of fixes and performance improvements, including fixing localization issues.
[PRO] Load custom inputs for ZTN curated models and Keras models.
1.3.0
- When ONNX models are loaded, an inference pass with sample data is run by default. That means all tensors / feature maps / weights / biases should be viewable immediately after input load. Please let us know if there are models that don't succeed at this initial pass so we can fix them!
(PRO) User input nodes are now attached to the model architecture diagram. When using Zetane Viewer Pro ($15/month) you can load custom inputs and send them through the model. Currently supported formats are .npy, .npz, .pb, and the majority of image formats (jpg, png, tiff, hdr, pic).
(PRO) When user inputs are misshapen, the engine will display an error about the model's shape expectation. Note that this feature is also usable by free users without the error popup, the input node will load the user input and show dimensions before attempting to run inference with the model.
(PRO) Any errors during model inference will also appear in the UI. An example is the shape error above. Individual graph operations may fail at any point during the inference pass-- the engine will attempt to populate the graph outputs up until the point of the error, a stack trace of the model run.
As always, we welcome feedback, bug reports, and any suggestions you might have.
1.2.0
- Shape mismatch errors for running model inference are shown in the UI, describing the expected input and the given input. (PRO)
- Changed default UI interaction with a mouse wheel to zoom by default, right click to drag the UI.
- Panels now scroll or move on hover, not just after being selected.
- Tensor viewer displays the original shape from file or API, without reordering the dimensions to fit the view panel.
- User notification for version upgrade now appears in the UI.
- Mac / Linux now run in API mode by default.
- Added a new ZTN snapshot for XAI features.
- User inputs now show above the Model Explorer panel's input node.
- A number of bug fixes and performance improvements
