Commits

Branch selector

 master

User selector

All users

Datepicker

All time

Commit History

Commits on Jan 25, 2026

• Merge pull request #189 from GotthardtZ/master

  Show description for a09b7bf
  hxim

  authored
  a09b7bf

  Copy full SHA for a09b7bf

Commits on Jan 24, 2026

• paq8px_v210

  GotthardtZ

  committed
  ed2d607

  Copy full SHA for ed2d607

• LSTM: tidy-up: fixed build on linux and other small issues. Updated READM.md

  GotthardtZ

  committed
  cb5e493

  Copy full SHA for cb5e493

Commits on Jan 21, 2026

• LSTM: VectorFunctions_AVX2::Softmax

  GotthardtZ

  committed
  74da656

  Copy full SHA for 74da656

Commits on Jan 20, 2026

• LSTM: parameter optimization

  GotthardtZ

  committed
  3680e5a

  Copy full SHA for 3680e5a

Commits on Jan 19, 2026

• LSTM: check prediction result (NaN, Inf).

  GotthardtZ

  committed
  8f866e8

  Copy full SHA for 8f866e8

• LSTM: new command line option: -lstmonly. Rewritten help screen.

  GotthardtZ

  committed
  59849dd

  Copy full SHA for 59849dd

Commits on Jan 12, 2026

• LSTM: naming cosmetics.

  GotthardtZ

  committed
  a6f1a34

  Copy full SHA for a6f1a34

• LSTM: do memcpy and memset with vector functions.

  GotthardtZ

  committed
  3c649b1

  Copy full SHA for 3c649b1

• Updated help screen with -savelstm:text and -loadstm:text

  GotthardtZ

  committed
  5511189

  Copy full SHA for 5511189

Commits on Jan 11, 2026

• LSTM: Cosmetics.

  GotthardtZ

  committed
  0d4894c

  Copy full SHA for 0d4894c

• LSTM: Initialize output layer's weights. But leave component weights as zeroes.

  GotthardtZ

  committed
  3b48e8e

  Copy full SHA for 3b48e8e

• LSTM: Cosmetics: centralized eps; added assert.

  GotthardtZ

  committed
  0a85772

  Copy full SHA for 0a85772

Commits on Jan 10, 2026

• LSTM: Simplified PolynomialDecay, now it's SqrtLearningRateDecay.

  GotthardtZ

  committed
  6c2c21f

  Copy full SHA for 6c2c21f

• LSTM: use clearer naming here and there.

  GotthardtZ

  committed
  fe6d07c

  Copy full SHA for fe6d07c

• LSTM: removed the unnecessary input_gate_complement storage. this is now always calculated on the fly. Other cosmetics.

  GotthardtZ

  committed
  4ec86f0

  Copy full SHA for 4ec86f0

• LSTM: use clearer naming here and there.

  GotthardtZ

  committed
  0797641

  Copy full SHA for 0797641

• LSTM: Adam: replaced bias correction and fixed beta2 with dynamic beta2

  GotthardtZ

  committed
  6802589

  Copy full SHA for 6802589

• LSTM: Adam: rescale the second moment values when changing sequence length.

  GotthardtZ

  committed
  1e2e767

  Copy full SHA for 1e2e767

Commits on Jan 9, 2026

• LSTM: split learning rates to base_lr and lr_scale, using a specific base learning rate in each optimizer.

  GotthardtZ

  committed
  263b93b

  Copy full SHA for 263b93b

• LSTM: epoch -> sequence_position; norm -> pre_activation; state -> cell_state; state_error -> cell_state_gradient; input_node -> input_gate; state -> gate_outputs; error -> gate_gradient, Layer -> …

  Show description for bed86f1
  GotthardtZ

  committed
  bed86f1

  Copy full SHA for bed86f1

Commits on Jan 8, 2026

• LSTM: added tuning_param to LstmLayer.

  GotthardtZ

  committed
  24c7a38

  Copy full SHA for 24c7a38

• LSTM: cleanup, consolidation, naming

  GotthardtZ

  committed
  9f3a904

  Copy full SHA for 9f3a904

• LSTM: added command line options to load/save model parameters from/to a file

  GotthardtZ

  committed
  e44aa10

  Copy full SHA for e44aa10

• Introduced hidden command line parameter: 'param' for model tuning

  GotthardtZ

  committed
  6255dab

  Copy full SHA for 6255dab

Commits on Jan 6, 2026

• LSTM: several speed-optimizations and small refactoring.

  GotthardtZ

  committed
  20a25e1

  Copy full SHA for 20a25e1

Commits on Jan 5, 2026

• LSTM: epoch management is now centralized.

  GotthardtZ

  committed
  32fe153

  Copy full SHA for 32fe153

Commits on Jan 4, 2026

• LSTM: time_step a.k.a update_step is now centralized.

  GotthardtZ

  committed
  8a13152

  Copy full SHA for 8a13152

• LSTM: don't allocate more memory than needed for layer_input.

  GotthardtZ

  committed
  7420c15

  Copy full SHA for 7420c15

• LSTM: Adam: added SSE2 support

  GotthardtZ

  committed
  f16d85a

  Copy full SHA for f16d85a

• LSTM: speedup: AccumulateLayerGradients

  GotthardtZ

  committed
  f0f7ca5

  Copy full SHA for f0f7ca5

Commits on Jan 3, 2026

• LSTM: speedup: AccumulateOutputLayerGradients

  GotthardtZ

  committed
  b723feb

  Copy full SHA for b723feb

• LSTM: output layer: use single shared weight set, not weight sets per epoch + changed SDG to Adam optimizer.

  GotthardtZ

  committed
  097cecc

  Copy full SHA for 097cecc

• LSTM: re-added weigth biases

  GotthardtZ

  committed
  834c90c

  Copy full SHA for 834c90c

• LSTM: re-added biases to output.

  GotthardtZ

  committed
  75f1ab7

  Copy full SHA for 75f1ab7

Pagination

Previous

Next