Release Notes

Optimium (ver.1.2.0)

Optimium

Performance Enhancements on Target Devices

• Leveraged NEON intrinsics to accelerate critical layers, resulting in notable speedups on ARM64 platforms such as Raspberry Pi, Jetson, and mobile SoCs.
• Introduced various optimizations across both ARM64 and x64 backends which contribute to reduced latency in key operations.

Nadya

Dramatically Improved Compilation Speed

• Integrated a series of compiler optimizations and structural changes that lead to much faster build times.

Optimium (ver.1.0.5)

Optimium

“Supported layers” speed improvement

• Additional speed improvement for SOFTMAX layer

Speed improvement through graph optimization

• Additional speed improvement for consecutive PAD and 'CONV_2D /MAX_POOL_2D /AVERAGE_POOL_2D' layer through graph optimization

Improvement for tuning stability

• Fixed a bug that caused a value mismatch when tuning CONCATENATION layer

Optimium (ver.1.0.4)

Optimium

“Supported layers” speed improvement

• Additional speed improvement for CONV_2D operations (CONV_2D and DEPTHWISE_CONV_2D)
• Additional speed improvement for Reduction operations (REDUCE_PROD, REDUCE_MAX, REDUCE_PROD, ARG_MAX, ARG_MIN, MEAN, SUM)

Improvement for model processing stability

• Additional improvement for PyTorch models that use third-party libraries
• See more details here

Optimium (ver.1.0.3)

Optimium

“Supported layers” update

• Fixed a constraint that was only supported when BATCHNORM_2D is positioned immediately after CONV_2D
• See more details here

Optimium (ver.1.0.2)

Optimium

“Supported layers” update

• Additional support for TFLite layers (ARG_MAX, LOG and more)
• Additional support for PyTorch layers (ARG_MAX, LOG and more)
• See more details here

Support PyTorch Model Composition

• Provide an interface for importing third-party modules used by users for model definition
• See more details here

Optimium Runtime

Bug Fix

• Fixed a bug that caused data corruption may occur when the sum of layer inputs and outputs exceeded 6
• Fixed a bug that caused segfault may occur if the program terminates while the model is running
• Fixed a bug that caused hang may occur if the program exits without executing after loading the model

Optimium (ver.1.0.1)

Optimium

“Supported layers” update

• Additional support for TFLite layers (HardSigmoid, Swish and more)
• Additional support for PyTorch layers (ADD, MUL and more)
• See more details here

Optimium Runtime

“Kotlin” update

• Fixed a bug where Kotlin bindings did not recognize extensions when used on Linux.
• Fixed a bug in Kotlin bindings that caused JVM Segmentation Fault when using “InferRequest.setCallback”