-
Autograd add 14 new operators: Gemm, GlobalAveragePool, ConstantOfShape, Dropout, ReduceSum, ReduceMean, Slice, Ceil, Split, Gather, Tile, NonZero, Cast, OneHot, as well as their test cases.
-
Sonnx add 14 new operators for both backend and frontend: Gemm, GlobalAveragePool, ConstantOfShape, Dropout, ReduceSum, ReduceMean, Slice, Ceil, Split, Gather, Tile, NonZero, Cast, OneHot, as well as their test cases.
-
Soonx model zoo add CV and NLP models, now support Bert-squad, Arcface, FER+ Emotion, MobileNet, ResNet18, Tiny Yolov2, Vgg16, Mnist.
-
Support multidirectional broadcasting for some operators: Add, Sub, Mul, Div, Pow, PRelu, Gemm
-
Reconstruct soonx, 1, support two types of weight value (Initializer and Constant Node); 2, for some operatos(BatchNorm, Reshape, Clip, Slice, Gather, Tile, OneHot), support moving some inputs to its arributes; 3, define and implement type conversion map. 4, fix some bugs within test cases.
-
Conv and Pooling support four bins padding(such as (2,3,2,3)), also support SAME_UPPER, SAME_LOWER pad mode and shape checking.