Skip to content

Correct tie_embedding_weights and add logit checking #1288

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
merged 6 commits into from
Oct 28, 2023
Merged

Correct tie_embedding_weights and add logit checking #1288

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
f97cb90
Add tie_embedding_weights and logit checking
nkovela1 Oct 27, 2023
c1edcad
Fix formatting
nkovela1 Oct 27, 2023
2eec377
Remove comments
nkovela1 Oct 27, 2023
a8178ff
Change backbone default
nkovela1 Oct 27, 2023
e68d815
Change weights hashes
nkovela1 Oct 27, 2023
05ace8d
Fix first hash
nkovela1 Oct 28, 2023
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: 1 addition & 1 deletion keras_nlp/models/t5/t5_backbone.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,7 +82,7 @@ def __init__(
activation="relu",
use_gated_activation=True,
layer_norm_epsilon=1e-06,
tie_embedding_weights=False,
tie_embedding_weights=True,
**kwargs,
):
# Encoder inputs
Expand Down
18 changes: 12 additions & 6 deletions keras_nlp/models/t5/t5_presets.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,10 +36,11 @@
"activation": "relu",
"use_gated_activation": False,
"layer_norm_epsilon": 1e-06,
"tie_embedding_weights": True,
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

let's update the default for T5Backbone to True

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Changed the default, thanks!

},
"preprocessor_config": {},
"weights_url": "https://storage.googleapis.com/keras-nlp/models/t5_small_multi/v1/model.weights.h5",
"weights_hash": "5a241ea61142eaf96ac1805898a2f2d1",
"weights_hash": "2e10b5f72405d464ee55026b07e60741",
"vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/t5_small_multi/v1/vocab.spm",
"vocabulary_hash": "9d15ef55d09d5a425ceb63fa31f7cae3",
},
Expand All @@ -64,10 +65,11 @@
"activation": "relu",
"use_gated_activation": False,
"layer_norm_epsilon": 1e-06,
"tie_embedding_weights": True,
},
"preprocessor_config": {},
"weights_url": "https://storage.googleapis.com/keras-nlp/models/t5_base_multi/v1/model.weights.h5",
"weights_hash": "9bef4c6650d91d1ea438ee4a2bea47ad",
"weights_hash": "bed6ef276cfe83d1323467051211978d",
"vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/t5_base_multi/v1/vocab.spm",
"vocabulary_hash": "9d15ef55d09d5a425ceb63fa31f7cae3",
},
Expand All @@ -92,10 +94,11 @@
"activation": "relu",
"use_gated_activation": False,
"layer_norm_epsilon": 1e-06,
"tie_embedding_weights": True,
},
"preprocessor_config": {},
"weights_url": "https://storage.googleapis.com/keras-nlp/models/t5_large_multi/v1/model.weights.h5",
"weights_hash": "eab8eee1bad033e65324a71cd6e5a8e9",
"weights_hash": "7854a05c2e6812899bf6f0f104792cda",
"vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/t5_large_multi/v1/vocab.spm",
"vocabulary_hash": "9d15ef55d09d5a425ceb63fa31f7cae3",
},
Expand All @@ -121,10 +124,11 @@
"activation": "keras_nlp>gelu_approximate",
"use_gated_activation": True,
"layer_norm_epsilon": 1e-06,
"tie_embedding_weights": False,
},
"preprocessor_config": {},
"weights_url": "https://storage.googleapis.com/keras-nlp/models/flan_small_multi/v1/model.weights.h5",
"weights_hash": "4e39b0bab56606a9ab2b8e52a6bc7a9f",
"weights_hash": "aa0fbaddb1759ef313bbc4f9e4f1e197",
"vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/flan_small_multi/v1/vocab.spm",
"vocabulary_hash": "9d15ef55d09d5a425ceb63fa31f7cae3",
},
Expand All @@ -149,10 +153,11 @@
"activation": "keras_nlp>gelu_approximate",
"use_gated_activation": True,
"layer_norm_epsilon": 1e-06,
"tie_embedding_weights": False,
},
"preprocessor_config": {},
"weights_url": "https://storage.googleapis.com/keras-nlp/models/flan_base_multi/v1/model.weights.h5",
"weights_hash": "b529270c5361db36d359a46403532b5c",
"weights_hash": "84a10bec83fd093931bb2a6264115d31",
"vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/flan_base_multi/v1/vocab.spm",
"vocabulary_hash": "9d15ef55d09d5a425ceb63fa31f7cae3",
},
Expand All @@ -177,10 +182,11 @@
"activation": "keras_nlp>gelu_approximate",
"use_gated_activation": True,
"layer_norm_epsilon": 1e-06,
"tie_embedding_weights": False,
},
"preprocessor_config": {},
"weights_url": "https://storage.googleapis.com/keras-nlp/models/flan_large_multi/v1/model.weights.h5",
"weights_hash": "50b8d3c88fc10db07e495d79ff29a1b6",
"weights_hash": "513f530ce790efa7e261c0ef965f3697",
"vocabulary_url": "https://storage.googleapis.com/keras-nlp/models/flan_large_multi/v1/vocab.spm",
"vocabulary_hash": "9d15ef55d09d5a425ceb63fa31f7cae3",
},
Expand Down
36 changes: 29 additions & 7 deletions tools/checkpoint_conversion/convert_t5_checkpoints.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -81,6 +81,12 @@ def convert_checkpoints(hf_model):
keras_nlp_model.get_layer("token_embedding").embeddings.assign(
hf_wts[f"{section}.embed_tokens.weight"]
)
if not keras_nlp_model.tie_embedding_weights:
keras_nlp_model.get_layer(
"token_embedding"
).reverse_embeddings.assign(
hf_wts["lm_head.weight"].transpose(1, 0).numpy()
)

# Query, key, value, and output projectors in self-attention
keras_nlp_model.get_layer(
Expand Down Expand Up @@ -308,17 +314,18 @@ def check_output(
print(k, v)

# Forward pass
keras_outputs = keras_model(keras_inputs)
hf_outputs = hf_model(**hf_inputs)
keras_out = keras_model(keras_inputs)
hf_out = hf_model(**hf_inputs, output_hidden_states=True)

# Only compare non-padded token ids.
keras_outputs = keras_outputs["decoder_sequence_output"]
keras_hidden_states = keras_out["decoder_sequence_output"]
hf_hidden_states = hf_out.decoder_hidden_states[-1]

keras_outputs = ops.take_along_axis(
keras_outputs, ops.where(decoder_padding_mask)
keras_hidden_states, ops.where(decoder_padding_mask)
)
hf_outputs = hf_outputs.last_hidden_state
hf_outputs = ops.take_along_axis(
hf_outputs, ops.where(decoder_padding_mask)
hf_hidden_states, ops.where(decoder_padding_mask)
)

print("-> KerasNLP output:", keras_outputs[0:5])
Expand All @@ -327,6 +334,21 @@ def check_output(
keras_outputs.detach().numpy(), hf_outputs.detach().numpy(), atol=1e-5
)

if keras_model.tie_embedding_weights:
keras_hidden_states = keras_hidden_states * (
keras_model.hidden_dim**-0.5
)

keras_logits = keras_model.token_embedding(
keras_hidden_states, reverse=True
)
hf_logits = hf_out.logits
print("-> KerasNLP logits:", keras_logits[0:5])
print("-> HF logits:", hf_logits[0:5])
np.testing.assert_allclose(
keras_logits.detach().numpy(), hf_logits.detach().numpy(), atol=1e-3
)


def count_params(weights):
shapes = [v.shape for v in weights]
Expand All @@ -339,7 +361,7 @@ def main(_):
os.mkdir(f"./{FLAGS.preset}")

print("\n-> Convert weights.")
hf_model = transformers.AutoModel.from_pretrained(hf_id)
hf_model = transformers.T5ForConditionalGeneration.from_pretrained(hf_id)
keras_model = convert_checkpoints(hf_model)

# Save the model.
Expand Down