* Report value for a step instead of epoch.
Report an objective function value for a step instead of epoch to optuna.
I made this modification for the following reason:
If "eval_steps" is less than steps per epoch, there maybe warnings like this: "optuna/trial/_trial.py:592: UserWarning: The reported value is ignored because this `step` 0 is already reported.". So "step" are more appropriate than "epoch" here.
* MOD: make style.
Co-authored-by: zhaowei01 <zhaowei01@yuanfudao.com>
* fix tolerance for a bloom slow test
* enhance alibi padding
- get rid of for loops
- deals better with padded batched input
- avoid useless cpu/gpu communication when creating alibi
Co-authored-by: justheuristic <justheuristic@gmail.com>
* optimize attention mask
* fix scaled softmax limit values
* optimize building alibi tensor
Co-authored-by: Younes Belkada <younesbelkada@users.noreply.github.com>
* fix attention_mask shape when it's None
* minor fixes
- fix docstring + arg names
* remove colons in docstring
* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* apply suggestion
* remove unsued arg
* refactor a bit
- use [:, None] for consistency
* refactor attention block
Co-authored-by: Nouamane Tazi <nouamane98@gmail.com>
* quick fixes
* first attempt
* refactor attention block and fix all tests except "test_simple_generation"
- added comments to better explain attention block
* remove debug lines and add TODO comment
* change `torch.bmm` to `torch.baddbmm`
- fixes `test_simple_generation`but breaks `test_batch_generation_padd`
* styling
* all tests are passing now
- use `bmm`
- add explanation for `allow_fp16_reduced_precision_reduction`
Co-authored-by: Younes Belkada <younesbelkada@users.noreply.github.com>
* styling
Co-authored-by: Younes Belkada <younesbelkada@users.noreply.github.com>
* fix support for accelerate
Co-authored-by: Younes Belkada <younesbelkada@users.noreply.github.com>
* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* remove attn softmax in fp32
* refactor comments
* refactor a bit
- remove warning message
- remove print on test
* refer to pytorch t5
* change the slow tests
- do the tests in fp32
- remove some comments
- keep large comments
* update expected output for `test_simple_generation`
- we now test using fp32
* make style + change comments a bit
* fix dtype padd test
Co-authored-by: justheuristic <justheuristic@gmail.com>
Co-authored-by: Nouamane Tazi <nouamane98@gmail.com>
Co-authored-by: Younes Belkada <younesbelkada@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Fix RESOURCE_EXHAUSTED error for large datasets on Flax example scripts
* using np.permutation for creating batch_idx
* train_samples_idx -> training_samples_idx
* fix type hints
* Fix type issue in using bucketing with Trainer
- Fix type issues in LengthGrouperSampler,
DistributedLengthGroupedSampler
refs: #18003
* Change logging type in LengthGroupedSampler
- Change `logger.warning` to `logger.info`
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Change logging type in DistributedLengthGroupedSampler
- Change `logger.warning` to `logger.info`
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Remove adundant clause in LengthGroupedSampler
- Use `elif`
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Remove adundant clause in DistributedLengthGroupedSampler
- Use `elif`
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Apply black, isort to modified codes in the script
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Drop columns after loading samples, rather than before, to avoid breaking transforms
* make fixup
* Add workaround so this PR can work with current datasets version
* Return scalar losses instead of per-sample means
* Make loss shape (1,) instead of scalar
* Allow scalar losses in test_loss_computation
* Allow scalar losses in test_loss_computation
* Allow scalar losses in test_loss_computation
* Remove XLA loss function for RAG
* Refactor to inherit from nn.Module instead of nn.ModuleList
* Fix typo
* Empty to trigger CI re-run
Blender Bot tests failing (should be unrelated to this PR) and pass locally). I don't have sufficient permisisons to re-run the CI workflow (totally or from failed)
* Rought TF conversion outline
* Tidy up
* Fix padding differences between layers
* Add back embedder - whoops
* Match test file to main
* Match upstream test file
* Correctly pass and assign image_size parameter
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* Add in MainLayer
* Correctly name layer
* Tidy up AdaptivePooler
* Small tidy-up
More accurate type hints and remove whitespaces
* Change AdaptiveAvgPool
Use the AdaptiveAvgPool implementation by @Rocketknight1, which correctly pools if the output shape does not evenly divide by input shape c.f. 9e26607e22 (r900109509)
Co-authored-by: From: matt <rocketknight1@gmail.com>
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
* Use updated AdaptiveAvgPool
Co-authored-by: matt <rocketknight1@gmail.com>
* Make AdaptiveAvgPool compatible with CPU
* Remove image_size from configuration
* Fixup
* Tensorflow -> TensorFlow
* Fix pt references in tests
* Apply suggestions from code review - grammar and wording
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
* Add TFResNet to doc tests
* PR comments - GlobalAveragePooling and clearer comments
* Remove unused import
* Add in keepdims argument
* Add num_channels check
* grammar fix: by -> of
Co-authored-by: matt <rocketknight1@gmail.com>
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
* Remove transposes - keep NHWC throughout forward pass
* Fixup look sharp
* Add missing layer names
* Final tidy up - remove from_pt now weights on hub
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
Co-authored-by: matt <rocketknight1@gmail.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
* Exclude Databricks from notebook env only if the runtime is below 11.0
* Dummy commit to trigger CI
* Empty commit to trigger CI
* Empty commit to trigger CI
* Empty commit to trigger CI
* Empty commit to trigger CI
* Empty commit to trigger CI
* Empty commit to trigger CI
* Empty commit to trigger CI
* Shifting labels for causal LM when using label smoother
When training CausalLM, loss is computed within model's foward() function and
labels are shifted internally. However, if label smoothing is applied, loss is
computed in trainer's compute_loss function and labels are not shifted.
This causes unintended confusion during the alignment of labels and corresponding
inputs. This commit is for resolving this confusion.
Resolves#17960
On branch shift_labels_for_causalLM
Changes to be committed:
modified: src/transformers/trainer.py
modified: src/transformers/trainer_pt_utils.py
* Update trainer.py
* Update src/transformers/trainer.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
