* [T5 Tokenizer] Model has no fixed position ids - there is no hardcoded max length
* [T5 Tokenizer] Model has no fixed position ids - there is no hardcoded max length
* correct t5 tokenizer
* correct t5 tokenizer
* fix test
* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* finish
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* First draft
* Add YolosForObjectDetection
* Make forward pass work
* Add mid position embeddings
* Add interpolation of position encodings
* Add expected values
* Add YOLOS to tests
* Add integration test
* Support tiny model as well
* Support all models in conversion script
* Remove mid_pe_size attribute
* Make more tests pass
* Add model to README and fix config
* Add copied from statements
* Rename base_model_prefix to vit
* Add missing YOLOS_PRETRAINED_CONFIG_ARCHIVE_MAP
* Apply suggestions from code review
* Apply more suggestions from code review
* Convert remaining checkpoints
* Improve docstrings
* Add YolosFeatureExtractor
* Add feature extractor to docs
* Add corresponding tests
* Fix style
* Fix docs
* Apply suggestion from code review
* Fix bad rebase
* Fix some more bad rebase
* Fix missing character
* Improve docs and variable names
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
* Add meta proxy
* Uses meta data to trace data dependent control-flow
* Remove commented class
* Handles torch creating functions
* Added type annotation to fix tracing
* Tracing works for everything but T5 and GPT-J
* Almost all previously supported models pass
* All architectures can be traced except T5
* Intermediate commit to have a trace of the comparison operators for HFProxy
* Everything works, except loss computation
* Everything works
* Removed unused import
* Overriden methods do not use underlying ops (linear and torch.matmul), and model attributes are copied to the traced version
* Fix torch_matmul_override
* Change attributes reference to deepcopy
* Remove breakpoint and add torch_index_override
* Small fix
* Fix typo
* Replace asserts by explicit exceptions
The emoji version must be either 0.5.4 or 0.6.0. Newer emoji versions have been updated to newer versions of the Emoji Charts, thus not consistent with the one used for pre-processing the pre-training Tweet corpus (i.e. not consistent with the vocab).
1. Fixes evaluation errors popping up when you train/eval on squad v2 (one was newly encountered and one that was previously reported Running SQuAD 1.0 sample command raises IndexError #15401 but not completely fixed).
2. Removes boolean arguments that don't use store_true. Please, don't use these: *ANY non-empty string is being converted to True in this case and this clearly is not the desired behavior (and it creates a LOT of confusion).
3. All no-trainer test scripts are now saving metric values in the same way (with the right prefix eval_), which is consistent with the trainer-based versions.
4. Adds forgotten model.eval() in the no-trainer versions. This improved some results, but not everything (see the discussion in the end). Please, see the F1 scores and the discussion below.
* Add first draft
* Improve script and README
* Improve README
* Apply suggestions from code review
* Improve script, add link to resulting model
* Add corresponding test
* Adjust learning rate
