I’m using X-Cube-AI Toolbox 9.1 to realize Vision Transformer deployment but got stuck. I exported from torch using ONNX opset=15. I then used onnx2tf to convert ViT to .tflite format, which is obviously a valid model as shown in Netron.
1.When I’m trying to analyze it in .tflite, there popped up an error says “BATCH_MATMUL not supported.”
2. Without conversion, analyzing .onnx model popped up something like “tuple has no method get_shape” and failed.
3. I further converted it to .keras and .h5, but .keras says: Unable to load the file, and then the whole CUBE IDE got stuck and expects a restart. Meanwhile .h5 only says something like the Length is four.
All of the trials above deploying ViT model failed with only a single line of error note during model analysis. So I have the following questions:
A) I tried to unroll the matmul in FOR loops, but unfortunately LOOP operator is also not supported.
B) I think one important thing is that the scaled dot product attention module seems to rely heavily on matmul or batch_matmul. So I wonder if the matmul operators, such as @ operator in torch, torch.bmm operator, or torch.einsum, could be supported? Or is there an interface to realize matmul customized operator in X-Cube-AI?
