NielsRogge/transformers

Pull Requests Commits

Using chunk instead of split to avoid accessing instance variable.

mfuntowicz committed 5 years ago

b55bc33b

Prefer using tuples instead of list for return types.

mfuntowicz committed 5 years ago

f604cbd0

Use constant value instead of math.sqrt(2.0 / pi) in gelu

mfuntowicz committed 5 years ago

30ef0ad0

Use positive indexing for flatten so it can be exported to ONNX.

mfuntowicz committed 5 years ago

5d9a2273

Avoid accessing shape and slices by using flatten in merge_heads

mfuntowicz committed 5 years ago

630adb2b

Avoid reshaping operations in Conv1D by using baddbmm

mfuntowicz committed 5 years ago

23bb7b94

Make sure gelu_new component are not represented as integers.

mfuntowicz committed 5 years ago

2e36f6d3

Revert "Use torch/tf rsqrt to scale attentions instead of div and math"

mfuntowicz committed 5 years ago

3ceb6833

Use torch/tf rsqrt to scale attentions instead of div and math

mfuntowicz committed 5 years ago

5b016216

Permuting only 2 axis should use transpose, as it always returns a view.

mfuntowicz committed 5 years ago

dd132be1

Do not allocate Softmax every forward call, use functional.

mfuntowicz committed 5 years ago

3989328c

Enable model export through ONNX runtime which some adjustments.

mfuntowicz committed 5 years ago

252f635e

julien-c committed 5 years ago

3e0f0621

Fix: one more try

julien-c committed 5 years ago

fc2a4c88

Same fix for `addcmul_`

julien-c committed 5 years ago

55bda525

Fix UserWarning: This overload of add_ is deprecated in pytorch==1.5.0

julien-c committed 5 years ago

ad02c961

[skip ci] remove local rank

julien-c committed 5 years ago

15550ce0

rerun notebook 02-transformers (#4341)

nikitajz committed 5 years ago

Verified 62427d08

Allow for None gradients in GradientAccumulator. (#4372)

jarednielsen committed 5 years ago

Verified 34706ba0

Should return overflowing information for the log (#4385)

LysandreJik committed 5 years ago

Verified edf9ac11

Attempt to unpin torch version for Github Action. (#4384)

mfuntowicz committed 5 years ago

Verified b908f2e9

[examples] Streamline doc

julien-c committed 5 years ago

af2e6bf8

p_mask in SQuAD pre-processing (#4049)

LysandreJik committed 5 years ago

Verified 7defc667

Updated ONNX notebook link in README.

mfuntowicz committed 5 years ago

84894974

Conversion script to export transformers models to ONNX IR. (#4253)

mfuntowicz committed 5 years ago

Verified db0076a9

Fix trainer evaluation (#4363)

patil-suraj committed 5 years ago

Verified 2d054801

Create README.md (#4359)

savasy committed 5 years ago

Verified 035678ef

Create README.md (#4357)

sy-wada committed 5 years ago

Verified b9c9e053

Tokenizer.batch_decode convenience method (#4159)

sshleifer committed 5 years ago

Verified 9535bf19

[tests] make pipelines tests faster with smaller models (#4238)

sshleifer committed 5 years ago

Verified 7822cd38

Older