Latest (TF 2.2) version TF lite converter converts MatMul Op to FULLY_CONNECTED + TRANSPOSE.
But, old version drops it as CUSTOM(MatMul).
We may need to add a Pass to do this if our import reads this Op.
From #1675
@seanshpark
We may need to add a Pass to do this if our import reads this Op.
Do you mean CUSTOM(MatMul) to FULLY_CONNECTED conversion?
I also want to point, that latest TF is not able to convert all variants of MatMul operator.
For example, we still can get custom op for tf.linalg.matmul(x, y, transpose_a=True, transpose_b=False)
Do you mean CUSTOM(MatMul) to FULLY_CONNECTED conversion?
Yes..
I also want to point, that latest TF is not able to convert all variants of MatMul operator.
Yes, not all but it is good if we can convert any possible cases.
@seanshpark is this task still available? If you don't mind, I'd like to take a crack at it
is this task still available? If you don't mind, I'd like to take a crack at it
Yes! it is available :) Please go ahead!
@kvochko, finished?
Yes, it's all done!
Thank you!!!