Releases: intel/caffe
Caffe_v1.1.6
Caffe_v1.1.6
- Optimize the inference performance of first layer INT8 convolution
- Support multi-instance running with weight sharing in inference
- Add Windows support in training and inference under single node
- Fix bugs in FC INT8/LARS/SSD detection/calibration tool
Caffe_v1.1.5
- Support memory optimization for inference
- Enable INT8 InnerProduct and its calibration support
- Release the full INT8 model of ResNet-50 v1.0
- Fix in-place concat for INT8 inference with single batch size
Caffe_v1.1.4
- Enabled single-node VNET training and inference
- Enhanced full convolution calibration to support models with customized data layer
- Enabled inference benchmarking scripts with multiple instances inference support
- Supported INT8 accuracy test in docker image
Caffe_v1.1.3
- Upgraded to MKLDNN v0.17
- Supported INT8 convolution with signed input
- Added more 3D layers support
Caffe_v1.1.2a
Features:
- Support multi-node inference
Caffe_v1.1.2
- Features
-
INT8 inference
Inference speed improved with upgraded MKL-DNN library.
In-place concat for latency improvement with batch size 1. Scale unify for concat for better performance. Support added in calibration tool as well -
FP32 inference
Performance improved on detectionOutput layer with ~3X
Add MKL-DNN 3D convolution support -
Multi-node training
SSD-VGG16 multi-node training is supported -
New models
Support training of R-FCN object detection model
Support training of Yolo-V2 object detection model
Support inference of SSD-MobileNet object detection model
Added the SSD-VGG16 multi-node model that converges to SOTA -
Build improvement
Fixed compiler warnings using GCC7+ version -
Misc
MKLML upgraded to mklml_lnx_2019.0.20180710
MKL-DNN upgraded to v0.16+ (4e333787e0d66a1dca1218e99a891d493dbc8ef1)
- Known issues
- INT8 inference accuracy drop for convolutions with output channel 16-individable
- FP32 training cannot reach SOTA accuracy with Winograd convolution
Caffe_v1.1.1a
- Features
- Update the batch size for benchmark scripts
- Bug fixings
- Fix docker image build target cpu-ubuntu
Caffe_v1.1.1
- Features
- INT8 inference
Inference speed improved with upgraded MKL-DNN library.
Accuracy improved with channel-wise scaling factor. Support added in calibration tool as well. - Multi-node training
Better training scalability on 10Gbe with prioritized communication in gradient all-reduce.
Support Python binding for multi-node training in pycaffe.
Default build now includes multi-node training feature. - Layer performance optimization: dilated convolution and softmax
- Auxiliary scripts
Added a script to parse the training log and plot loss trends (tools/extra/caffe_log_parser.py and tools/extra/plot_loss_trends.py).
Added a script to identify the batch size for optimal throughput given a model (scripts/obtain_optimal_batch_size.py).
Improved benchmark scripts to support Inception-V3 and VGG-16 - New models
Support inference of R-FCN object detection model.
Added the Inception-V3 multi-node model that converges to SOTA. - Build improvement
Merged PR#167 "Extended cmake install package script for MKL"
Fixed all ICC/GCC compiler warnings and enabled warning as error.
Added build options to turn off each inference model optimization.
Do not try to download MKL-DNN when there is no network connection.
- Misc
- MLSL upgraded to 2018-Preview
- MKL-DNN upgraded to 464c268e544bae26f9b85a2acb9122c766a4c396
Caffe_v1.1.0
- Features
- Support INT8 inference. A calibration tool is provided to transform FP32 models to INT8 models
- Support convolution and element-wise sum fusion, boosting inference performance (e.g. ResNet-50)
- Support SSD training and inference with pure MKLDNN engine
- Enhance MSRA weight filler with scale parameter
- Support performance collection on single node in the same way as multi-node
- Set CPU_ONLY as default in CMake configuration
- Bug fixings
- Fix correctness issue on layers with various engines
- Sync sampling bug fix 96175b from Wei Liu’s SSD branch
- Fix multi-node crash issue running from pycaffe
- Correct link library of MLSL for multi-node
- Fix build issue of weight quantization
- Misc
- Upgrade MKLML to 2018.0.1.20171227 and MKLDNN to v0.12
- Update models for multi-node training
- Enhance installation and benchmarking scripts