Skip to content

1.1.md

Latest commit

 

History

History
56 lines (50 loc) · 2.24 KB

1.1.md

File metadata and controls

56 lines (50 loc) · 2.24 KB

ML Suite v1.1 Release Notes

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

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 (MAX)
    • Deconvolution
    • Concat
    • Eltwise (SUM)
    • BatchNorm
    • Scale
    • Slice
    • Layers supported in CPU:
    • InnerProduct
    • Softmax

Known Issues

  • AWS Overlays - Only overlay_2 and overlay_3 are available, and they are clock rate reduced. The overlays will be updated shortly, and will offer better performance post update.
  • Networks expecting inputs of 128x128x3 will not function. There is a bug with image preprocessing in xfdnn. To be fixed shortly.
  • quantize.py: See note below:
The quantizer expects a particular naming convention to be used in the deploy prototxt for caffe flows
Layer names must match the name of the "top" blob.
This is a common convention followed by most models
In the future we will add code to alleviate this requirement
  • test_classify and batch_classify scripts for VGG16 result in segmentation faults.
  • test_classify and batch_classify scripts for networks with average pool layers, e.g., MobileNet and Inception v3, do not produce desired accuracies, to be fixed in a future release.