Skip to content
This repository was archived by the owner on Oct 31, 2025. It is now read-only.

Refactoring to 1.5.1 #1792

Open
mmarcinmichal wants to merge 1,764 commits into
base: master
Choose a base branch
from
Open

Refactoring to 1.5.1 #1792

mmarcinmichal wants to merge 1,764 commits into from

Conversation

@mmarcinmichal
Copy link

@mmarcinmichal mmarcinmichal commented Nov 27, 2023

The code, originally in version 1.4.1, has been upgraded to version 1.5.1. The main changes include:

  • Compatibility with Tensorflow 2.13.0
  • Compatibility with jax==0.4.20
  • Update of dependencies
  • Bringing the code to a Build passing status (running tests, with minor exclusions for terraformer and RL tests).

henrykmichalewski and others added 30 commits July 15, 2021 14:22
@henrykmichalewski @copybara-github
MathQA-Python generation colab.
501edf9 
PiperOrigin-RevId: 385008224
@sebastianjaszczur @copybara-github
Added a Python script running predict mode of pretrained Terraformer …
6202671 
…model.

PiperOrigin-RevId: 385184130
@kkimdev @copybara-github
Fix testTensordot test by adjusting tol.
78142d8 
PiperOrigin-RevId: 386141282
@lukaszkaiser
Add eval mode to decoding, no-positional encoding option.
b6075dd
@copybara-github
Merge pull request google#1676 from lukaszkaiser:lk1
3012a17 
PiperOrigin-RevId: 388370105
@lukaszkaiser
Scientific papers decoding in eval mode.
6d99c42
@lukaszkaiser
Merge branch 'master' into lk1
216f90f
@copybara-github
Merge pull request google#1677 from lukaszkaiser:lk1
5dbfbf2 
PiperOrigin-RevId: 388585118
@lukaszkaiser
Nicer printing in Terraformer predict script.
88e2785
@copybara-github
Merge pull request google#1659 from manifest:checkpoint
6ecf3bd 
PiperOrigin-RevId: 389321901
@copybara-github
Merge pull request google#1678 from lukaszkaiser:lk1
1be2dfe 
PiperOrigin-RevId: 389322099
@copybara-github
Merge pull request google#1658 from manifest:data_dup
db21384 
PiperOrigin-RevId: 389322729
@copybara-github
Merge pull request google#1357 from manifest:masked_self_attention
e1f727d 
PiperOrigin-RevId: 389322833
@copybara-github
Merge pull request google#1651 from manifest:bidirectional
565de00 
PiperOrigin-RevId: 389325185
@hawkinsp @copybara-github
[JAX] Fix test failure in Trax due to upcoming change to JAX.
27178b1 
An upcoming change to JAX wraps `jit` decorators around more standard library functions, including division (/). jit-decorated functions will error if passed a large Python scalar integer that overflows the range of a int32 or int64. The workaround is to explicitly case the Python scalar to a specific type, here np.float32.

PiperOrigin-RevId: 390152127
@syzymon
Add current version of RelformerLM
ffb5368 
__________________________________
Fix variable names + formatting

Add rel attention wrapper layer

Gin configurable resampling layers

Add unit tests

Import the current model

Add rel attn inputs

Overwrite rel attention by the current version
@syzymon
Resampling fixes
e54a982
@syzymon
Gin configs for enwik8 and imagenet
dd7529e 
Updated gin for imagenet32

Fix formatting

Enwik8 and imagenet64 configs

Remove assert

Rename + fix formatting

Add docstrings
@mtyrolski @syzymon
CIFAR10 config for Relformer
1ac17b2
Yaml update
fa73ae9
memory saving implemented
f51bb53
@syzymon
Use rotary positional embeddings in relformer_imagenet64.gin config
5aa60c1
No need for im64 sweep
d9fca55
@syzymon
Disable predict tests
3317c54
@syzymon
Add an option to disable residual upsampling
b9c5838
Changed enwik8 sweep, limited 16,4 jobs in favour of 16/12, 2
eab694b
@syzymon
Increase number of layers in cifar10 experiment
91c96c9
@syzymon
Fix max_len in the cifar10 Transformer config
725d651
@syzymon
Add wiki40b experiment
e6a2056
@syzymon
Update imagenet32/64 parameters
b3cfaac
@mmarcinmichal
Avoid caching during trainer loss computation
2a8d8ef
@mmarcinmichal
Examples refactoring
c98a2a5
@mmarcinmichal
Examples refactoring
4f92d97
@mmarcinmichal
Refactor training orchestration services
3a28f73
@mmarcinmichal
Add checkpoint and metrics interfaces
4c4692a
@mmarcinmichal
Examples refactoring
a7c3364
@mmarcinmichal
Examples refactoring
fb68408
@mmarcinmichal
Refactor training loop orchestration services
6441a4a
@mmarcinmichal
Ensure orchestrator reshapes batches via device manager
a624172
@mmarcinmichal
Merge branch 'master' into codex/refactor-training-loop-with-separate…
619af5e 
…-components-swos8s
@mmarcinmichal
Deprecate trainer_lib in favor of Loop
b9f8baf
@mmarcinmichal
Merge remote-tracking branch 'origin/codex/refactor-training-loop-wit…
074cd4e 
…h-separate-components-swos8s'
@mmarcinmichal
Remove trainer_lib in favor of loop wrappers
c08ce3f
@mmarcinmichal
Examples refactoring
4a7debd
@mmarcinmichal
Refactoring
080645d
@mmarcinmichal
Refactoring
7391523
@mmarcinmichal
data
b4af8b5 
fastmath
@mmarcinmichal
data
835e41b 
fastmath
layers

optimizer

tf
utils
@mmarcinmichal
data
153b968 
fastmath
layers

optimizer

tf
utils
@mmarcinmichal
data
919ff5d 
fastmath
layers

optimizer

tf
utils
@mmarcinmichal
data
b56063a 
fastmath
layers

optimizer

tf
utils
@mmarcinmichal
data
e1b95cc 
fastmath
layers

optimizer

tf
utils
@mmarcinmichal
New version
cf819da 
optimizer

tf
utils
@mmarcinmichal
Big refactor to 3.0.
19b6cf6
@mmarcinmichal
Big refactor to 3.0.
a06e07f
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.