Quartus Custom Matrix Multiplication & Quantization #523

bo3z · 2022-04-08T14:42:17Z

Custom multiplication for Quartus, depending on the type of weights and input data (binary, ternary, exponential etc.)
Added ApplyAlpha and BatchNormalizationQuantizedTanh for Quartus
QKeras tests parameterised for both Quartus and Vivado
Basic tests for Quartus softmax; failing on accuracy (2%)

… both Quartus and Vivado

vloncar

Looks good. We'll follow up with the listed TODOs (like streaming stuff placeholders and strategy for Quartus)

bo3z force-pushed the quartus-matmult-quantization branch 2 times, most recently from 3b64bce to 2ec968a Compare April 11, 2022 12:24

bo3z force-pushed the quartus-matmult-quantization branch from 2ec968a to d646f37 Compare April 19, 2022 14:12

bo3z added 9 commits April 21, 2022 19:52

Custom multiplication for Quartus

a98f3f9

Support for different product types on Quartus

ff56b1d

Quartus custom multiplication for batch normalisation

95d4aaf

Added quantization template for Quartus

8a76368

Moved BatchNormalizationQuantizedTanh to common FPGA folder - used by…

fbe7cb8

… both Quartus and Vivado

Added quantization pass for Quartus

49975f9

Parameterized qkeras tests for Quartus

df206d2

Quartus test for Softmax

0b7eb38

Common FPGA passes, for both Quartus and Vivado

c7a8888

bo3z force-pushed the quartus-matmult-quantization branch from d646f37 to c7a8888 Compare April 26, 2022 13:15

vloncar approved these changes Apr 29, 2022

View reviewed changes

vloncar merged commit c96a7bc into fastmachinelearning:master Apr 29, 2022

thesps mentioned this pull request Jul 18, 2022

Override parent backend optimizer passes with derived backend passes #597

Merged

Provide feedback