Caffe_v1.1.6

1. Optimize the inference performance of first layer INT8 convolution
2. Support multi-instance running with weight sharing in inference
3. Add Windows support in training and inference under single node
4. Fix bugs in FC INT8/LARS/calibration tool

Caffe_v1.1.6

1. Optimize the inference performance of first layer INT8 convolution
2. Support multi-instance running with weight sharing in inference
3. Add Windows support in training and inference under single node
4. Fix bugs in FC INT8/LARS/SSD detection/calibration tool

Caffe_v1.1.5

1. Support memory optimization for inference
2. Enable INT8 InnerProduct and its calibration support
3. Release the full INT8 model of ResNet-50 v1.0
4. Fix in-place concat for INT8 inference with single batch size

Caffe_v1.1.4

1. Enabled single-node VNET training and inference
2. Enhanced full convolution calibration to support models with customized data layer
3. Enabled inference benchmarking scripts with multiple instances inference support
4. Supported INT8 accuracy test in docker image

Caffe_v1.1.3

1. Upgraded to MKLDNN v0.17
2. Supported INT8 convolution with signed input
3. Added more 3D layers support

Caffe_v1.1.2a

Features:

1. Support multi-node inference

Caffe_v1.1.2

• Features

1. 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

2. FP32 inference
   Performance improved on detectionOutput layer with ~3X
   Add MKL-DNN 3D convolution support

3. Multi-node training
   SSD-VGG16 multi-node training is supported

4. 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

5. Build improvement
   Fixed compiler warnings using GCC7+ version

6. Misc
   MKLML upgraded to mklml_lnx_2019.0.20180710
   MKL-DNN upgraded to v0.16+ (4e333787e0d66a1dca1218e99a891d493dbc8ef1)

• Known issues

1. INT8 inference accuracy drop for convolutions with output channel 16-individable
2. FP32 training cannot reach SOTA accuracy with Winograd convolution

Caffe_v1.1.1a

• Features

1. Update the batch size for benchmark scripts

• Bug fixings

1. Fix docker image build target cpu-ubuntu

Caffe_v1.1.1

• Features

1. INT8 inference
   Inference speed improved with upgraded MKL-DNN library.
   Accuracy improved with channel-wise scaling factor. Support added in calibration tool as well.
2. 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.
3. Layer performance optimization: dilated convolution and softmax
4. 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
5. New models
   Support inference of R-FCN object detection model.
   Added the Inception-V3 multi-node model that converges to SOTA.
6. 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

1. MLSL upgraded to 2018-Preview
2. MKL-DNN upgraded to 464c268e544bae26f9b85a2acb9122c766a4c396

Caffe_v1.1.0

• Features

1. Support INT8 inference. A calibration tool is provided to transform FP32 models to INT8 models
2. Support convolution and element-wise sum fusion, boosting inference performance (e.g. ResNet-50)
3. Support SSD training and inference with pure MKLDNN engine
4. Enhance MSRA weight filler with scale parameter
5. Support performance collection on single node in the same way as multi-node
6. Set CPU_ONLY as default in CMake configuration

• Bug fixings

1. Fix correctness issue on layers with various engines
2. Sync sampling bug fix 96175b from Wei Liu’s SSD branch
3. Fix multi-node crash issue running from pycaffe
4. Correct link library of MLSL for multi-node
5. Fix build issue of weight quantization

• Misc

1. Upgrade MKLML to 2018.0.1.20171227 and MKLDNN to v0.12
2. Update models for multi-node training
3. Enhance installation and benchmarking scripts