Skip to content

[T5] Fix init in TF and Flax for pretraining #17294

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
patrickvonplaten merged 6 commits into from
May 18, 2022
Merged

[T5] Fix init in TF and Flax for pretraining #17294

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c7a8c11
fix init
patrickvonplaten May 16, 2022
621b8dd
Apply suggestions from code review
patrickvonplaten May 16, 2022
7ea09bc
fix
patrickvonplaten May 17, 2022
983aa34
Merge branch 'fix_init_t5' of https://github.com/patrickvonplaten/tra…
patrickvonplaten May 17, 2022
dbd9a75
finish
patrickvonplaten May 17, 2022
0308cad
Update src/transformers/modeling_tf_utils.py
patrickvonplaten May 17, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions src/transformers/models/t5/modeling_t5.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -768,6 +768,8 @@ def _init_weights(self, module):
# Mesh TensorFlow embeddings initialization
# See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)
if hasattr(module, "lm_head") and not self.config.tie_word_embeddings:
module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)
elif isinstance(module, T5DenseReluDense):
# Mesh TensorFlow FF initialization
# See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
Expand Down
13 changes: 9 additions & 4 deletions src/transformers/models/t5/modeling_tf_t5.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1112,7 +1112,9 @@ def _shift_right(self, input_ids):
class TFT5Model(TFT5PreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model, name="shared")
self.shared = TFSharedEmbeddings(
config.vocab_size, config.d_model, name="shared", initializer_range=self.config.initializer_factor
)

# retrieve correct absolute scope for embed token wrapper
with tf.compat.v1.variable_scope("shared") as shared_abs_scope_name:
Expand Down Expand Up @@ -1259,8 +1261,9 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel, TFCausalLanguageModeling
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.model_dim = config.d_model

self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model, name="shared")
self.shared = TFSharedEmbeddings(
config.vocab_size, config.d_model, name="shared", initializer_range=self.config.initializer_factor
)

# retrieve correct absolute scope for embed token wrapper
with tf.compat.v1.variable_scope("shared") as shared_abs_scope_name:
Expand Down Expand Up @@ -1600,7 +1603,9 @@ def _reorder_cache(self, past, beam_idx):
class TFT5EncoderModel(TFT5PreTrainedModel):
def __init__(self, config, *inputs, **kwargs):
super().__init__(config, *inputs, **kwargs)
self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model, name="shared")
self.shared = TFSharedEmbeddings(
config.vocab_size, config.d_model, name="shared", initializer_range=self.config.initializer_factor
)

# retrieve correct absolute scope for embed token wrapper
with tf.compat.v1.variable_scope("shared") as shared_abs_scope_name:
Expand Down