Open MichaelMcCulloch opened 1 year ago
It appears that the NNDelegate code has crashed, and my suspicion is that it is using the xnn pack delegate by default. To use the NNDelegate, certain flags may need to be enabled during the building of the tflite library.
If you wish to run a model other than the English one, you will need to load the "filters_vocab_multilingual.bin" file from this GitHub repository: https://github.com/usefulsensors/openai-whisper/blob/main/models/filters_vocab_multilingual.bin. Please try running the model with this file loaded.
to just confirm if nnapi delegate is working or not with tflite model. please use the below tflite model and run the test again https://github.com/tensorflow/tflite-micro/blob/main/third_party/xtensa/examples/pytorch_to_tflite/mobilenet_v2_quantized_1x3x224x224.tflite
to just confirm if nnapi delegate is working or not with tflite model. please use the below tflite model and run the test again https://github.com/tensorflow/tflite-micro/blob/main/third_party/xtensa/examples/pytorch_to_tflite/mobilenet_v2_quantized_1x3x224x224.tflite
This tflite model appears to work.
It appears that the NNDelegate code has crashed, and my suspicion is that it is using the xnn pack delegate by default. To use the NNDelegate, certain flags may need to be enabled during the building of the tflite library.
If you wish to run a model other than the English one, you will need to load the "filters_vocab_multilingual.bin" file from this GitHub repository: https://github.com/usefulsensors/openai-whisper/blob/main/models/filters_vocab_multilingual.bin. Please try running the model with this file loaded.
Bookmarked, thank you.
to just confirm if nnapi delegate is working or not with tflite model. please use the below tflite model and run the test again https://github.com/tensorflow/tflite-micro/blob/main/third_party/xtensa/examples/pytorch_to_tflite/mobilenet_v2_quantized_1x3x224x224.tflite
This tflite model appears to work. Looks like Whisper model is not compatible for NNAPI delegate. If I learn more about it I will post here..
Are you able to tell me what factors contribute to it being compatible?
The whisper hybrid tflite model may not be compatible due to its mixed precision format, where weights are in int8 and activations are in float32. It's possible that generating a full float model could resolve this issue. Additionally, other causes of incompatibility could be related to unsupported ops, meaning that certain operations included in the model are not supported by the NNDelegate's available ops
Logcat when the interpreter is loaded thus: