ashtonmeuser / godot-tf-inference Goto Github PK

The addon is unlikely to work as is on Windows systems. There will likely be issues with linking once the addon is compiled.

• Update SConstruct file to add Windows support
• Ensure Godot project (including model and addon) is exportable on Windows
• Update documentation to reflect any special steps that are required

Use smart default signature names on model load. This would avoid a user having to manually set signature names. Dependent on #7.

Currently, input and output tensors must be one-dimensional.

• Accept N-dimensional input tensor
• Return N-dimensional output tensor

Currently, only TensorFlow SavedModel format is supported. Support for ONNX, an open format, may be preferable. Conversion between the two formats is possible.

Precursor to support for arbitrary number of input tensors. If model info can be extracted (see #7), we can check to ensure arguments supplied to infer() match the number of input vectors.

An array should be given as the only argument to the infer() method. Currently, a string could be passed causing a failure in cppflow. We should catch this and return a relevant error. See #3.

An info() method that extracts a model's signature names may be helpful.

Currently, the load_model() and infer() methods don't report any errors they encountered. The calling GDScript has no idea whether a model was successfully loaded or an inference was correctly executed.

• Error returned from load_mode()
• Error returned from infer()

The addon is unlikely to work as is on Linux systems. There will likely be issues with linking once the addon is compiled.

• Update SConstruct file to add Linux support
• Ensure Godot project (including model and addon) is exportable on Linux
• Update documentation to reflect any special steps that are required