Code

[WIP] Export to tflite int8#9

Code

Open

kyakuno wants to merge 28 commits intoonnx_2_1axinc-ai/segment-anything-2:onnx_2_1from tflite_int8axinc-ai/segment-anything-2:tflite_int8Copy head branch name to clipboard

+1,079-251

Commits

Commits on Dec 20, 2024

• Export to tflite int8

  kyakuno

  committed
  93666c1

  Copy full SHA for 93666c1

• Fix int8 inference

  kyakuno

  committed
  3ab2f02

  Copy full SHA for 3ab2f02

Commits on Dec 21, 2024

• Channel wise quantization

  kyakuno

  committed
  f4d50c9

  Copy full SHA for f4d50c9

• Export mask decoder

  kyakuno

  committed
  d83fcf8

  Copy full SHA for d83fcf8

• Fix data supply

  kyakuno

  committed
  1ad8e46

  Copy full SHA for 1ad8e46

• Export memory attention to int8

  kyakuno

  committed
  c0e95b8

  Copy full SHA for c0e95b8

• Disable per channel

  kyakuno

  committed
  a23c412

  Copy full SHA for a23c412

• Tensorflow quantization

  kyakuno

  committed
  ff185b2

  Copy full SHA for ff185b2

• Export mask decoder

  kyakuno

  committed
  d72daa1

  Copy full SHA for d72daa1

• Export prompt encoder

  kyakuno

  committed
  466e660

  Copy full SHA for 466e660

Commits on Dec 22, 2024

• Export memory attention and memory encoder

  kyakuno

  committed
  2bab821

  Copy full SHA for 2bab821

• Export memory encoder

  kyakuno

  committed
  c998c4c

  Copy full SHA for c998c4c

• Export projection

  kyakuno

  committed
  056f594

  Copy full SHA for 056f594

• Implement calibration mode

  kyakuno

  committed
  445f141

  Copy full SHA for 445f141

• Calibration for memory attention

  kyakuno

  committed
  45cd9b3

  Copy full SHA for 45cd9b3

• Int8 switch

  kyakuno

  committed
  37290b2

  Copy full SHA for 37290b2

• Implement calibration

  kyakuno

  committed
  5de0f24

  Copy full SHA for 5de0f24

• Calibrate prompt encoder

  kyakuno

  committed
  7a765c6

  Copy full SHA for 7a765c6

• Temporally use float model for prompt encoder

  kyakuno

  committed
  ee2f66e

  Copy full SHA for ee2f66e

• Use large dataset

  kyakuno

  committed
  c10b194

  Copy full SHA for c10b194

• Added quantize script

  kyakuno

  committed
  1cc5769

  Copy full SHA for 1cc5769

• Use large data for calibration

  kyakuno

  committed
  4d78e77

  Copy full SHA for 4d78e77

Commits on Dec 24, 2024

• Convert to numpy for ailia

  kyakuno

  committed
  951c6c6

  Copy full SHA for 951c6c6

Commits on Dec 25, 2024

• Fix tflite int8 for image encoder of video predictor

  kyakuno

  committed
  2f46cf1

  Copy full SHA for 2f46cf1

• Fix tflite_int8 for prompt_encoder and mask_decoder for video predictor

  kyakuno

  committed
  4ad6b4f

  Copy full SHA for 4ad6b4f

Commits on Dec 27, 2024

• Added torch quantize support

  kyakuno

  committed
  6341bca

  Copy full SHA for 6341bca

Commits on Jan 21, 2025

• Added mixed precision support

  kyakuno

  committed
  a292627

  Copy full SHA for a292627

• Upload mixed precision model

  kyakuno

  committed
  16af7f1

  Copy full SHA for 16af7f1