New TF model design#7753

Closed

jplu wants to merge 78 commits intohuggingface:masterfrom

jplu:tf-bert-pretraining

Commits on Nov 11, 2020

Merge changes
jplu
committed
Merge changes
jplu
committed
Fix conflict
jplu
committed
Fix conflict
jplu
committed
Fix conflict
jplu
committed
Fix test
jplu
committed
Fix conflict
jplu
committed
Fix conflict
jplu
committed
Make return_dict the default behavior and display a warning message
jplu
committed
Revert
jplu
committed
Replace wrong keyword
jplu
committed
Revert code
jplu
committed
Add eisum layers
jplu
committed
Add eisum layers to Electra, Longformer and Roberta
jplu
committed
Rework a bit Longformer
jplu
committed
Fix rebase issue
jplu
committed
Fix rebase issue
jplu
committed
Apply style
jplu
committed
Better model design for BERT + LM pretraining for BERT + more robust custom TF weights loading
jplu
committed
Fix return_dict to True
jplu
committed
Make the layers created or not on the fly
jplu
committed
changes for return_dict
jplu
committed
Implement new output logic
jplu
committed
Finish new TF model outputs
jplu
committed
Improve TF models input
jplu
committed
Inputs parsing function now returns list of values ordered by keys
jplu
committed
Apply style
jplu
committed
Fix check_copies
jplu
committed
fix Boolean
jplu
committed
Rollback on naming
jplu
committed
Address few Sylvain's comments
jplu
committed
Update src/transformers/modeling_tf_bert.py

jplu
and
sgugger
committed
Finish to address Sylvain's comments
jplu
committed
Several bug fix
jplu
committed
Address Patrick's comments
jplu
committed
Put the einsum equations in lower case to make ONNX happy
jplu
committed
Apply style
jplu
committed
Revert the ModelOutput
jplu
committed
Update src/transformers/modeling_tf_bert.py

jplu
and
LysandreJik
committed
Fix style
jplu
committed
Fix the tests
jplu
committed
Apply style
jplu
committed
Fix tests
jplu
committed
Improve embeddings computation
jplu
committed
Apply style
jplu
committed
Fix tests
jplu
committed
Fix tests
jplu
committed
Improve the perf of the TFBertLayer
jplu
committed
Optimize MaskedLM computation
jplu
committed
Review boolean handling
jplu
committed
Fix tests
jplu
committed
Apply style
jplu
committed
Apply style + fix ONNX tests
jplu
committed
Use the embeddings as utils
jplu
committed
Fix conflict
jplu
committed
Apply style
jplu
committed
Apply style
jplu
committed
Reput loss to next sentence output
jplu
committed
Remove useless print
jplu
committed
Apply style
jplu
committed
Make return_dict=True by default in PreTrainedConfig
jplu
committed
Remove duplicate layers
jplu
committed
Fix TFLongformer tests
jplu
committed
Apply style
jplu
committed
Fix some Flax/PyTorch tests
jplu
committed
Fix some Flax/PyTorch tests
jplu
committed
Remove naming exceptions
jplu
committed
Apply style
jplu
committed
Add a warning message + revert to global default return_dict behavior (False)
jplu
committed
Fix few forgotten tests
jplu
committed
Apply style
jplu
committed
Fix Longformer tests
jplu
committed
Fix T5
jplu
committed
Fix conflict
jplu
committed
Restore example
jplu
committed
Restore TF Electra
jplu
committed
Restore longformer
jplu
committed
Rebase
jplu
committed