Code

RADE V2 prototyping#42

Code

Open

drowe67 wants to merge 37 commits intomaindrowe67/radae:mainfrom dr-radev2drowe67/radae:dr-radev2Copy head branch name to clipboard

Commits

Commits on Jan 13, 2025

• wip BPF filter in training loop

  drowe67

  committed
  4c19d43

  Copy full SHA for 4c19d43

Commits on Jan 14, 2025

• BPF filter in training loop, trains up OK, loss a bit high, spectrum looks good, PAPR 2dB with 2nd clipper

  drowe67

  committed
  4adfc92

  Copy full SHA for 4adfc92

Commits on Jan 16, 2025

• wip multiple clip/filter stages

  drowe67

  committed
  1acf8b6

  Copy full SHA for 1acf8b6

• 3 stage clip-filter gives us a reasonable 99% power BW of < 1300 Hz

  drowe67

  committed
  9a8b175

  Copy full SHA for 9a8b175

• clean up of 3 stage BPF-clip, 99% varies between 1250 and 1450Hz on diff training runs

  drowe67

  committed
  10bde8d

  Copy full SHA for 10bde8d

• 51 tap filter works OK

  drowe67

  committed
  0ffa3ff

  Copy full SHA for 0ffa3ff

Commits on Jan 20, 2025

• supporting plots for RADE V2 PAPR/bandwidth paper

  drowe67

  committed
  dce6360

  Copy full SHA for dce6360

• AWGN/MPP plots for RADE V2 PAPR/bandwidth paper

  drowe67

  committed
  c1899e6

  Copy full SHA for c1899e6

• 250117_test model used for paper curves

  drowe67

  committed
  d5f1383

  Copy full SHA for d5f1383

Commits on Jan 26, 2025

• wip first pass ML EQ

  drowe67

  committed
  a1167fe

  Copy full SHA for a1167fe

• wip - correcting a random phase offset

  drowe67

  committed
  52ecb9e

  Copy full SHA for 52ecb9e

• moved channel model into forward(), inference section plots scatter

  drowe67

  committed
  39bb89d

  Copy full SHA for 39bb89d

Commits on Jan 27, 2025

• way to bypass training/EQ to test vanilla channel simulation

  drowe67

  committed
  6e65d7a

  Copy full SHA for 6e65d7a

• train with phase based loss function; classical lin DSP option for comparison

  drowe67

  committed
  772c237

  Copy full SHA for 772c237

Commits on Jan 28, 2025

• load/save models, Eb/No curves

  drowe67

  committed
  db09c52

  Copy full SHA for db09c52

• top level script to generate curves

  drowe67

  committed
  4bd38b9

  Copy full SHA for 4bd38b9

Commits on Jan 29, 2025

• refactored to do de-framing, similar results

  drowe67

  committed
  86bd4ff

  Copy full SHA for 86bd4ff

• wip framer class

  drowe67

  committed
  3ebdfe3

  Copy full SHA for 3ebdfe3

• generic complex to (real,img) float pairs

  drowe67

  committed
  1f2f29a

  Copy full SHA for 1f2f29a

Commits on Jan 30, 2025

• dsp equaliser in framer

  drowe67

  committed
  d04142b

  Copy full SHA for d04142b

• reshaped frames to (batch,Nc,Ns)

  drowe67

  committed
  629b249

  Copy full SHA for 629b249

• wip frame 2, working for bypass and dsp

  drowe67

  committed
  c2d1ffb

  Copy full SHA for c2d1ffb

• wip frame 2, converging well with ML and low BER

  drowe67

  committed
  b4482a8

  Copy full SHA for b4482a8

Commits on Jan 31, 2025

• mleq_cvurves.sh set up for different frames, loss_phase working again

  drowe67

  committed
  a77e1c9

  Copy full SHA for a77e1c9

Commits on Feb 3, 2025

• getting good results from frame2 with constant phase offset

  drowe67

  committed
  3ec7a49

  Copy full SHA for 3ec7a49

• added frequency offset to training, working quite well

  drowe67

  committed
  9c09401

  Copy full SHA for 9c09401

Commits on Feb 5, 2025

• first pass at training with complex h

  drowe67

  committed
  ef67028

  Copy full SHA for ef67028

Commits on Feb 6, 2025

• complex H option for inference

  drowe67

  committed
  0b3a297

  Copy full SHA for 0b3a297

• pilot puncturing option, but poor results when training

  drowe67

  committed
  ad08b3e

  Copy full SHA for ad08b3e

Commits on Feb 11, 2025

• hard clipper in time domain training, evauluating 250207_test

  drowe67

  committed
  7e25849

  Copy full SHA for 7e25849

Commits on Feb 13, 2025

• added random per sequnce fine time offset (linear phase) and freq offsets)

  drowe67

  committed
  d7a049f

  Copy full SHA for d7a049f

Commits on Feb 21, 2025

• wip bottelneck 2 training for comparison

  drowe67

  committed
  756232c

  Copy full SHA for 756232c

Commits on Feb 25, 2025

• after trying 250225

  drowe67

  committed
  9856344

  Copy full SHA for 9856344

Commits on Feb 26, 2025

• Feb 250225 write up in hf3

  drowe67

  committed
  74b7278

  Copy full SHA for 74b7278

Commits on Mar 1, 2025

• --correct_time_offset, corrects phase shift due to --time_offset -> effectively genie coarse time est, tested on model05 with digital symbols

  drowe67

  committed
  4dbe7bf

  Copy full SHA for 4dbe7bf

Commits on Mar 3, 2025

• 250227b plots

  drowe67

  committed
  3063d04

  Copy full SHA for 3063d04

• --correct_time_offset now has an indep argument to --time_offset, to allow adjustment for various time shifts

  drowe67

  committed
  20cf162

  Copy full SHA for 20cf162