Release Note 1.9.0
ONE Compiler
Compiler supports more operations
NonMaxSuppressionV4, NonMaxSuppressionV5, PadV2, Unique
Changes
Quantization enhancements: channel wise UINT8 quantization(Conv2D, DepwiseConv, TransposeConv, FullyConnected)
Experimental requantization from INT8 to UINT8
Adding more operator value tests
tf2tfliteV2 supports conversion from Keras model, saved model
Refactoring for better maintenance long Class codes using visitor patterns
Introducing optimization pass that fuses batch normalization with Transposed Convolution.
ONE Runtime
Runtime backend operation support
CPU backend: RANK
CPU backend qasymm uint8: LOG_SOFTMAX
ACL-CL backend: LEAKY_RELU, RESIZE_NEAREST_NEIGHBOR
Optimization
Copy Elimination between compatible backends
Operation Implementation
Operations with same parameters are unified
Change
CPU backend qasymm uint8 performance enhancement: arithmetic operations