Name | Modified | Size | Downloads / Week |
---|---|---|---|
Parent folder | |||
README.md | 2022-04-15 | 1.5 kB | |
v1.3.0 source code.tar.gz | 2022-04-15 | 77.0 MB | |
v1.3.0 source code.zip | 2022-04-15 | 79.7 MB | |
Totals: 3 Items | 156.7 MB | 0 |
Added
- Support on-device training for MLP, CNN(lenet, resnet50, mobilnetv1), transformer/bert(text to speech)
- Support change model input and output names in X2bolt
- Support more graph optimizations : Transpose+Convolution, Swish, Quantization, Power+Scale
- Support dynamic output related operators : Shape, ConstantOfShape, GenerateProposals, NonZero, NonMaxSuppression, Reshape, etc
- Support more operators : GridSample, CumSum, OneHot, Round, Floor, Ceil
- Support more networks on CPU : yolov2, yolov3, yolov4, yolov5, faster-rcnn, mask-rcnn, retinanet, dfsmn, frill, conformer, unet, etc
- Support Armv8 int8 to accelerate NLP network
- Improve inference performance on avx2 CPU
- Support netron to visualize bolt model
- Support not to bind CPU core
- Add C API MemoryCheck to check bolt memory leak
Changed
- X2bolt add -I and -O options to change model input and output names.
- X2bolt add -t option to convert model for on-device training.
- C API CreateModel and AllocAllResultHandle return value is set to NULL when unsuccessful.
- install.sh add --neon option to close arm neon acceleration on old platform.
- some operator parameter defination
Fixed
- Fix GPU depth2space and deconv bug
- Fix GPU preprocess tool on armv8 platform bug
- Fix x86 Sigmoid precision
- Fix C API CloneResultHandle bug
- Fix mobilnetv1 int8 inference
- Fix Java API build bug on Windows
- Fix ONNX converter deconv, pooling parameter bug
Removed
- Equal operator is replaced with Check.