I have fine tuned a VIT_B32 models via keras, from GitHub - faustomorales/vit-keras: Keras implementation of ViT (Vision Transformer), which does use gelu activation function.
I tried parsing TFLITE or ONNX variant but i was getting the following error around wrong shapes:
elif len(output_shapes) == len(input_format) == 4:
TypeError: object of type 'NoneType' has no len()
(hailo_virtualenv) hailo@dev:/local/workspace/12_f
upon searching posts here and looking into ONNX model via netron i saw there were a lot of unk__ dimensions (for batch) so using onnx i adjusted it to 1.
tryig parsing again via command:
hailo parser onnx --hw-arch hailo8l fixed_mdl_fold1.onnx
i got past the previous error and got a new error as per below log
raise MisspellNodeError(f"Unable to find {err_str}: {wrong_names}, please verify and try again.")
i tried with python API and fed the input/output names and input sizes and got more info on unsupported layers (reduced log due to size limit of the post:
UnsupportedShuffleLayerError in op vision_transformer/vit-b32/Transformer/encoderblock_11/MlpBlock_3/Transformer/encoderblock_11/Dense_1/Tensordot: Failed to determine type of layer to create in node vision_transformer/vit-b32/Transformer/encoderblock_11/MlpBlock_3/Transformer/encoderblock_11/Dense_1/Tensordot
UnsupportedShuffleLayerError in op vision_transformer/vit-b32/Transformer/encoderblock_11/MultiHeadDotProductAttention_1/out/Tensordot: Failed to determine type of layer to create in node vision_transformer/vit-b32/Transformer/encoderblock_11/MultiHeadDotProductAttention_1/out/Tensordot
UnsupportedSquareLayerError in op vision_transformer/dense_1/Gelu/Pow: Pow operator vision_transformer/dense_1/Gelu/Pow can only be supported as square (got power of 3.0)
UnsupportedSquareLayerError in op vision_transformer/dense_2/Gelu/Pow: Pow operator vision_transformer/dense_2/Gelu/Pow can only be supported as square (got power of 3.0)
UnsupportedSoftmaxLayerError in op vision_transformer/dense_3/Softmax: Unsupported softmax
Please try to parse the model again, using these start node names: vision_transformer/dense/Gelu/Tanh, vision_transformer/dense/Gelu/mul, vision_transformer/dense/Gelu/add
Please try to parse the model again, using these end node names: vision_transformer/dense_1/Gelu/mul, vision_transformer/dense_1/BiasAdd:0
is there any tips on how to convert this model or there are too many non-supported layers here?