Commits

Branch selector

 master

User selector

renjie-liu

Datepicker

All time

Commit History

Commits on Mar 20, 2020

• Fix typos: should be ChannelDimIndexInterface

  Show description for a5e787c
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  a5e787c

  Copy full SHA for a5e787c

Commits on Mar 19, 2020

• remove bidi rnn seq_len test, we didn't support seq_len anyway in tflite kernel anyway

  Show description for 7fb2f93
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  7fb2f93

  Copy full SHA for 7fb2f93

• Support negative axis for reverse_v2.

  Show description for 6eca562
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  6eca562

  Copy full SHA for 6eca562

Commits on Mar 16, 2020

• Disable flaky unidirectional_sequence_lstm_test

  Show description for e27d8bb
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  e27d8bb

  Copy full SHA for e27d8bb

Commits on Mar 14, 2020

• Fix flaky lstm test. The cell_clip should be 0.

  Show description for 3ba639f
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  3ba639f

  Copy full SHA for 3ba639f

Commits on Mar 13, 2020

• update time major: we won't transpose inputs & outputs if it's not time-majored but only set the time_majored attribute. This will make bidi-lstm fusion easier.

  Show description for d8fdc31
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  d8fdc31

  Copy full SHA for d8fdc31

• Add bidirectional_sequence_lstm op to tfl dialect.

  Show description for dde815f
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  dde815f

  Copy full SHA for dde815f

Commits on Mar 10, 2020

• Update fuse logic for keras lstm to allow last output being consumed.

  Show description for 94db791
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  94db791

  Copy full SHA for 94db791

• Add support for last output for keras lstm fusion.

  Show description for f2b115d
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  f2b115d

  Copy full SHA for f2b115d

Commits on Mar 7, 2020

• add a e2e test for keras lstm fusion.

  Show description for 56ee67f
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  56ee67f

  Copy full SHA for 56ee67f

Commits on Mar 6, 2020

• Use i32 instead of i64 for transpose perm. TFL currently does not support i64 perm for transpose op right now.

  Show description for 14b3001
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  14b3001

  Copy full SHA for 14b3001

• Update entry function: if we don't see a main func, we will look for the function that has tf.entry_function attribute and set that one to be main.

  Show description for 8f50df8
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  8f50df8

  Copy full SHA for 8f50df8

• Add shape inference pass to optimize away the unnecessary casts.

  Show description for 629e3a8
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  629e3a8

  Copy full SHA for 629e3a8

• update time stamp for function signature: due to the while loop, the lstm function signature will have like tensor<1x?x10xf32>, but if we know the input, we can derive the timestamp.

  Show description for 1113b22
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  1113b22

  Copy full SHA for 1113b22

• fix lstm non-time major inputs.

  Show description for 692e9bf
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  692e9bf

  Copy full SHA for 692e9bf

Commits on Mar 5, 2020

• Turn on saved_model related passes if we're importing from saved_model.

  Show description for 09f617f
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  09f617f

  Copy full SHA for 09f617f

• fix tflite lstm & rnn tests for tf v2.

  Show description for a2dac1d
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  a2dac1d

  Copy full SHA for a2dac1d

Commits on Mar 4, 2020

• Expose a c++ api for graph_def_to_tfl_flatbuffer for python converter to convert saved_model directly.

  Show description for ef95ed6
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  ef95ed6

  Copy full SHA for ef95ed6

Commits on Mar 3, 2020

• [NFC] Refactor graphdef_to_tfl_flatbuffer logic: factor out populate_quantization_specs,

  Show description for 188c063
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  188c063

  Copy full SHA for 188c063

Commits on Feb 29, 2020

• Legalize to uni seq lstm instead of lstm as keras lstm have sequences.

  Show description for f31dc6d
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  f31dc6d

  Copy full SHA for f31dc6d

• Fuse keras lstm conditionally，due to Keras LSTM & TFL fused LSTM difference, we only fuse keras lstm only if the full sequences are consumed.

  Show description for 03d5fe9
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  03d5fe9

  Copy full SHA for 03d5fe9

Commits on Feb 28, 2020

• Avoid updating keras lstm function signature

  Show description for 6739ada
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  6739ada

  Copy full SHA for 6739ada

Commits on Feb 26, 2020

• Add go_backwards support for keras fused lstm

  Show description for 6b8469f
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  6b8469f

  Copy full SHA for 6b8469f

Commits on Feb 25, 2020

• Add time-major support for keras_lstm fused ops.

  Show description for be63241
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  be63241

  Copy full SHA for be63241

• Add unidirectional sequence rnn op_def to graphdef_to_flatbuffer and also add a e2e test.

  Show description for 24b2d02
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  24b2d02

  Copy full SHA for 24b2d02

• Legalize ophint converted unidirectional sequence rnn in mlir.

  Show description for c3267a9
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  c3267a9

  Copy full SHA for c3267a9

• Add ophint converted unidirectional sequence lstm opdef to graphdef_to_flatbuffer, so usl will be legalized as customized op and also add the e2e test.

  Show description for aa69f34
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  aa69f34

  Copy full SHA for aa69f34

• Make sure ophint converted graph does not have unknown shapes.

  Show description for 25a7219
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  25a7219

  Copy full SHA for 25a7219

• Update depthwise_conv 3x3 test setup.

  Show description for 1933f1b
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  1933f1b

  Copy full SHA for 1933f1b

Commits on Feb 21, 2020

• Update ophint optional inputs logic (fill in optional input with constant nodes), also update the toco unidirectional lstm legalization.

  Show description for 57c49e1
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  57c49e1

  Copy full SHA for 57c49e1

Commits on Feb 20, 2020

• Legalize ophint converted node in mlir. (this is the first cl of a series of cls for import ophint python directly converted model).

  Show description for 5ce0816
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  5ce0816

  Copy full SHA for 5ce0816

• Add function to load saved model for tflite mlir converter.

  Show description for f8b2a05
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  f8b2a05

  Copy full SHA for f8b2a05

• In verify ophint extraction: instead of error out, we should just don't do anything.

  Show description for 1f8af07
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  1f8af07

  Copy full SHA for 1f8af07

Commits on Feb 19, 2020

• Remove the depthwise conv 3x3 output shift > 0 restriction for per-channel. (it's a obsolete restriction), we're using sqrshl, so it's fine to handle output_shift <= 0 case.

  Show description for ea13922
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  ea13922

  Copy full SHA for ea13922

Commits on Feb 13, 2020

• Add ellipsis mask transformation in mlir for tflite.

  Show description for f6a8bcd
  

  renjie-liu

  authored and

  tensorflower-gardener

  committed
  f6a8bcd

  Copy full SHA for f6a8bcd

Pagination

Previous

Next