Code

QKeras updates#30

Code

Merged

vloncar merged 18 commits intovloncar:csim_integrationvloncar/hls4ml:csim_integrationfrom thesps:csim_integrationthesps/hls4ml:csim_integrationCopy head branch name to clipboardMay 27, 2020

+450-150

Commits

Commits on May 4, 2020

• Increase precision of weight printing

  thesps

  committed
  6d27701

  Copy full SHA for 6d27701

Commits on May 5, 2020

• Merge branch 'precision_fix' of https://github.com/thesps/hls4ml into csim_integration

  thesps

  committed
  83d33f8

  Copy full SHA for 83d33f8

• Get class name differently for profiling. Import profiling in model/__init__ for easier import elsewhere

  thesps

  committed
  8a3c915

  Copy full SHA for 8a3c915

• Make optimizer passes configurable. API is with list in config file. Splitting hls_model.py into hls_model.py and hls_layers.py was necessary to remove circular import dependency from optimizers im…

  Show description for d7024df
  thesps

  committed
  d7024df

  Copy full SHA for d7024df

• QKeras use of integer bits seems to differ a bit from ap_fixed. e.g. quantized_bits(4,0).max() is 1.0, whereas it would be 0.5 with ap_fixed. So, add 1 bit to the integer for ap_fixed types

  thesps

  committed
  69e3563

  Copy full SHA for 69e3563

• Add 1 bit elsewhere for QKeras, since they don't count the sign bit

  thesps

  committed
  a0597a2

  Copy full SHA for a0597a2

Commits on May 12, 2020

• QKeras/low precision improvements. Add Optimizer pass to set rounding modes (required qkeras converter split for circular import), and hooks for rounding mode to IntegerPrecisionType, FixedPrecisio…

  Show description for be63c1a
  thesps

  committed
  be63c1a

  Copy full SHA for be63c1a

• Merge vloncar/csim_integration

  thesps

  committed
  0f53ad7

  Copy full SHA for 0f53ad7

• Cleanup after merge

  thesps

  committed
  3d8a228

  Copy full SHA for 3d8a228

• New softmax implementation

  thesps

  committed
  2bfec0b

  Copy full SHA for 2bfec0b

• Remove some prints

  thesps

  committed
  79c59ec

  Copy full SHA for 79c59ec

Commits on May 13, 2020

• Extra support for QKeras compatibility. In particular for 'alpha' in quantizers

  thesps

  committed
  4840d73

  Copy full SHA for 4840d73

• Support alpha='auto_po2' in quantizers. Adds QKerasFactorizeAlpha Optimizer pass to factorize out alpha scale and insert new 'ApplyAlpha' (BatchNormalization) layer to apply it back. Attach data_un…

  Show description for 2a44923
  thesps

  committed
  2a44923

  Copy full SHA for 2a44923

Commits on May 18, 2020

• Fix rounding/saturation mode in HLSTypes. Add FuseConsecutiveBatchNorm optimizer. Fix other passes to prevent multiple matches.

  thesps

  committed
  937bbf9

  Copy full SHA for 937bbf9

• Add new Softmax to all converters. Remove rounding, saturation mode from Integer type. Fix match in rounding, saturation optimizer pass, add layer name to match API.

  thesps

  committed
  4c75f03

  Copy full SHA for 4c75f03

Commits on May 19, 2020

• Use BinaryQuantizer for QKeras binary kernel quantizers

  thesps

  committed
  b1d3b54

  Copy full SHA for b1d3b54

Commits on May 27, 2020

• Remove 'alpha' from non-QKeras quantizers. Treat binary_tanh activation properly. Propagate new Softmax to keras_to_hls.

  thesps

  committed
  5173511

  Copy full SHA for 5173511

• Make quantizer default to None for WeightVariable. Remove rogue include

  thesps

  committed
  0377d34

  Copy full SHA for 0377d34