Skip to content

Latest commit

 

History

History
122 lines (100 loc) · 4.83 KB

1.x.md

File metadata and controls

122 lines (100 loc) · 4.83 KB
Raw

ML Suite Release Notes

Framework Support and Layers

  • Caffe: 1.0.0
  • Tensorflow: 1.8
  • Darknet (Specifically for YOLO)
  • Supported Layers
    • Convolution
    • ReLU - supported following Convolution / Eltwise Layers
    • Pooling
    • Deconvolution (Only compiler support, runtime is pending)
    • Concat
    • Eltwise (sum)
    • BatchNorm
    • Scale
    • Slice
    • Layers supported in CPU:
      • InnerProduct
      • Softmax
      • Region/NMS

Release 1.3

New Features

  • Updated to support 2018.2 SDx and XRT
  • New Platforms Supported:
    • VCU1525, Alveo U200, Alveo U250
  • Added Support for xDNNv3 - Delivers higher throughput at lower latency
  • Auto-detect Platform/DSA
  • Both v2 and v3 libxfdnn.so runtime libraries automatically built with make
  • Auto-detect xclbin version and correctly load corresponding libxfdnn.so runtime library for XDNN_v2 or XDNN_v3
  • Auto-detect batch size based on # PEs (batch size = # PEs, except for xDNNv2 8-bit where it will be double)
  • XDNN_v3 max throughput end-to-end demo for Googlenet_v1 and Resnet50
    • Customers can use this to enjoy command-line performance that matches FPGA kernel time
  • pytest suite to exercise all documented examples/apps
  • REST server updated to use new XFDNN API; handles concurrent requests with exec_async
  • Perpetual Demo now supports xDNNv3 and upto 8 FPGA cards
  • Uptdated xfDNN APIs to allow higher throughput streaming on xDNNv3, and improve ease of use
  • xfDNN Compiler enhanced to report better memory utilization and estimate throughput/latency of network

Bug fixes

  • batch_classify example renamed to streaming_classify
  • Improvements to Tensorflow Compiler/Quantizer
  • Improved accuracy for Quantized Resnet101 models

Known Issues

  • examples/classification/test_classify_async_multinet.py - incorrectly reloads network schedule in some circumstances
  • XDNNv3 fpga binary (overlay_4.xclbin) is not included for the AWS platform. Stay tuned.
  • xfDNN runtime API documentation is out of date. Please use the examples & notebooks for reference

Release 1.2

New Features

  • Added support for 2018.2 XRT, and Alveo-U200
  • New Jupyter Notebook available:
    • Image Classification with TensorFlow
  • Benchmark utility script added (GoogLeNet, ResNet50)
  • Enhanced Documentation with API markdown guides
  • Added Power 9 Support

Bug fixes

  • Improved layer merging optimizations
  • xdnn.computeFC optimized for 2x speedup w/ vectorized numpy implementation
  • xdnn.computeSoftmax optimized for 10x speedup w/ vectorized numpy implementation
    • In the future we will move to C++ implementation
  • Fixed bug where image dimensions mod 32 would cause invalid preprocessing, and zero detections

Known Issues

  • examples/classification/test_classify_async_multinet.py - incorrectly reloads network schedule in some circumstances

Release 1.1

New Features

  • Added Jupyter Notebook Support
  • New Jupyter Notebooks available:
    • Image Classification with Caffe
    • Using the xfDNN Compiler w/ a Caffe Model
    • Using the xfDNN Quantizer w/ a Caffe Model
    • Object Detection w/ YOLOv2 + Darknet to Caffe conversion
  • Image Classification Googlenet Demo for VCU1525
  • Enhanced Documentation with ML Suite Overview, Overlay Selector Guide and FAQ
  • Introducing Nimbix Support
  • Updated SDx DSA support to xilinx_vcu1525_dynamic_5_1 for VCU1525 and Nimbix

Bug fixes

  • AWS overlay names have been updated, removing 'aws' prefix
  • General Enhancements and Bug fixes for xfDNN Compiler/Quantizer
  • Batch Nom layers implementation corrected from Darknet
  • Default file permission fixed
  • Root dir issues involving "ml-suite" resolved
  • xdnn.execute api no longer needs images/batch argument
  • Quantizer updated to allow for custom file names for output files
  • xdnn_io updated to return 'none' if there are not FC layers in network

Known Issues

  • Batch Normalization layers not supported by the quantizer
  • Local Response Normalization layers not supported
  • Hardware solution for average pool causes some accuracy loss, to be fixed in a future release
  • Standard ReLU is the only supported non-linearity (Leaky ReLU Networks must be modified/retrained

Release 1.0

This release is the first push of Xilinx's ML Suite to Github.
Releasing to github will enable rapid release cycles, a smaller release footprint, and open source contribution.

New Features

  • FPGA Accelerated Image Classification, support for many networks
  • FPGA Accelerated Object Detection, YOLOv2 support
  • Python API for deploying inference to FPGA
  • Precompiled xfDNN Library
  • Updated Support for xDNNv2

Known Issues

  • Batch Normalization layers not supported by the quantizer
  • Local Response Normalization layers not supported
  • Hardware solution for average pool causes some accuracy loss, to be fixed in a future release
  • Standard ReLU is the only supported non-linearity (Leaky ReLU Networks must be modified/retrained)