keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- keras_hub/__init__.py +52 -0
- keras_hub/api/__init__.py +27 -0
- keras_hub/api/layers/__init__.py +47 -0
- keras_hub/api/metrics/__init__.py +24 -0
- keras_hub/api/models/__init__.py +249 -0
- keras_hub/api/samplers/__init__.py +29 -0
- keras_hub/api/tokenizers/__init__.py +35 -0
- keras_hub/src/__init__.py +13 -0
- keras_hub/src/api_export.py +53 -0
- keras_hub/src/layers/__init__.py +13 -0
- keras_hub/src/layers/modeling/__init__.py +13 -0
- keras_hub/src/layers/modeling/alibi_bias.py +143 -0
- keras_hub/src/layers/modeling/cached_multi_head_attention.py +137 -0
- keras_hub/src/layers/modeling/f_net_encoder.py +200 -0
- keras_hub/src/layers/modeling/masked_lm_head.py +239 -0
- keras_hub/src/layers/modeling/position_embedding.py +123 -0
- keras_hub/src/layers/modeling/reversible_embedding.py +311 -0
- keras_hub/src/layers/modeling/rotary_embedding.py +169 -0
- keras_hub/src/layers/modeling/sine_position_encoding.py +108 -0
- keras_hub/src/layers/modeling/token_and_position_embedding.py +150 -0
- keras_hub/src/layers/modeling/transformer_decoder.py +496 -0
- keras_hub/src/layers/modeling/transformer_encoder.py +262 -0
- keras_hub/src/layers/modeling/transformer_layer_utils.py +106 -0
- keras_hub/src/layers/preprocessing/__init__.py +13 -0
- keras_hub/src/layers/preprocessing/masked_lm_mask_generator.py +220 -0
- keras_hub/src/layers/preprocessing/multi_segment_packer.py +319 -0
- keras_hub/src/layers/preprocessing/preprocessing_layer.py +62 -0
- keras_hub/src/layers/preprocessing/random_deletion.py +271 -0
- keras_hub/src/layers/preprocessing/random_swap.py +267 -0
- keras_hub/src/layers/preprocessing/start_end_packer.py +219 -0
- keras_hub/src/metrics/__init__.py +13 -0
- keras_hub/src/metrics/bleu.py +394 -0
- keras_hub/src/metrics/edit_distance.py +197 -0
- keras_hub/src/metrics/perplexity.py +181 -0
- keras_hub/src/metrics/rouge_base.py +204 -0
- keras_hub/src/metrics/rouge_l.py +97 -0
- keras_hub/src/metrics/rouge_n.py +125 -0
- keras_hub/src/models/__init__.py +13 -0
- keras_hub/src/models/albert/__init__.py +20 -0
- keras_hub/src/models/albert/albert_backbone.py +267 -0
- keras_hub/src/models/albert/albert_classifier.py +202 -0
- keras_hub/src/models/albert/albert_masked_lm.py +129 -0
- keras_hub/src/models/albert/albert_masked_lm_preprocessor.py +194 -0
- keras_hub/src/models/albert/albert_preprocessor.py +206 -0
- keras_hub/src/models/albert/albert_presets.py +70 -0
- keras_hub/src/models/albert/albert_tokenizer.py +119 -0
- keras_hub/src/models/backbone.py +311 -0
- keras_hub/src/models/bart/__init__.py +20 -0
- keras_hub/src/models/bart/bart_backbone.py +261 -0
- keras_hub/src/models/bart/bart_preprocessor.py +276 -0
- keras_hub/src/models/bart/bart_presets.py +74 -0
- keras_hub/src/models/bart/bart_seq_2_seq_lm.py +490 -0
- keras_hub/src/models/bart/bart_seq_2_seq_lm_preprocessor.py +262 -0
- keras_hub/src/models/bart/bart_tokenizer.py +124 -0
- keras_hub/src/models/bert/__init__.py +23 -0
- keras_hub/src/models/bert/bert_backbone.py +227 -0
- keras_hub/src/models/bert/bert_classifier.py +183 -0
- keras_hub/src/models/bert/bert_masked_lm.py +131 -0
- keras_hub/src/models/bert/bert_masked_lm_preprocessor.py +198 -0
- keras_hub/src/models/bert/bert_preprocessor.py +184 -0
- keras_hub/src/models/bert/bert_presets.py +147 -0
- keras_hub/src/models/bert/bert_tokenizer.py +112 -0
- keras_hub/src/models/bloom/__init__.py +20 -0
- keras_hub/src/models/bloom/bloom_attention.py +186 -0
- keras_hub/src/models/bloom/bloom_backbone.py +173 -0
- keras_hub/src/models/bloom/bloom_causal_lm.py +298 -0
- keras_hub/src/models/bloom/bloom_causal_lm_preprocessor.py +176 -0
- keras_hub/src/models/bloom/bloom_decoder.py +206 -0
- keras_hub/src/models/bloom/bloom_preprocessor.py +185 -0
- keras_hub/src/models/bloom/bloom_presets.py +121 -0
- keras_hub/src/models/bloom/bloom_tokenizer.py +116 -0
- keras_hub/src/models/causal_lm.py +383 -0
- keras_hub/src/models/classifier.py +109 -0
- keras_hub/src/models/csp_darknet/__init__.py +13 -0
- keras_hub/src/models/csp_darknet/csp_darknet_backbone.py +410 -0
- keras_hub/src/models/csp_darknet/csp_darknet_image_classifier.py +133 -0
- keras_hub/src/models/deberta_v3/__init__.py +24 -0
- keras_hub/src/models/deberta_v3/deberta_v3_backbone.py +210 -0
- keras_hub/src/models/deberta_v3/deberta_v3_classifier.py +228 -0
- keras_hub/src/models/deberta_v3/deberta_v3_masked_lm.py +135 -0
- keras_hub/src/models/deberta_v3/deberta_v3_masked_lm_preprocessor.py +191 -0
- keras_hub/src/models/deberta_v3/deberta_v3_preprocessor.py +206 -0
- keras_hub/src/models/deberta_v3/deberta_v3_presets.py +82 -0
- keras_hub/src/models/deberta_v3/deberta_v3_tokenizer.py +155 -0
- keras_hub/src/models/deberta_v3/disentangled_attention_encoder.py +227 -0
- keras_hub/src/models/deberta_v3/disentangled_self_attention.py +412 -0
- keras_hub/src/models/deberta_v3/relative_embedding.py +94 -0
- keras_hub/src/models/densenet/__init__.py +13 -0
- keras_hub/src/models/densenet/densenet_backbone.py +210 -0
- keras_hub/src/models/densenet/densenet_image_classifier.py +131 -0
- keras_hub/src/models/distil_bert/__init__.py +26 -0
- keras_hub/src/models/distil_bert/distil_bert_backbone.py +187 -0
- keras_hub/src/models/distil_bert/distil_bert_classifier.py +208 -0
- keras_hub/src/models/distil_bert/distil_bert_masked_lm.py +137 -0
- keras_hub/src/models/distil_bert/distil_bert_masked_lm_preprocessor.py +194 -0
- keras_hub/src/models/distil_bert/distil_bert_preprocessor.py +175 -0
- keras_hub/src/models/distil_bert/distil_bert_presets.py +57 -0
- keras_hub/src/models/distil_bert/distil_bert_tokenizer.py +114 -0
- keras_hub/src/models/electra/__init__.py +20 -0
- keras_hub/src/models/electra/electra_backbone.py +247 -0
- keras_hub/src/models/electra/electra_preprocessor.py +154 -0
- keras_hub/src/models/electra/electra_presets.py +95 -0
- keras_hub/src/models/electra/electra_tokenizer.py +104 -0
- keras_hub/src/models/f_net/__init__.py +20 -0
- keras_hub/src/models/f_net/f_net_backbone.py +236 -0
- keras_hub/src/models/f_net/f_net_classifier.py +154 -0
- keras_hub/src/models/f_net/f_net_masked_lm.py +132 -0
- keras_hub/src/models/f_net/f_net_masked_lm_preprocessor.py +196 -0
- keras_hub/src/models/f_net/f_net_preprocessor.py +177 -0
- keras_hub/src/models/f_net/f_net_presets.py +43 -0
- keras_hub/src/models/f_net/f_net_tokenizer.py +95 -0
- keras_hub/src/models/falcon/__init__.py +20 -0
- keras_hub/src/models/falcon/falcon_attention.py +156 -0
- keras_hub/src/models/falcon/falcon_backbone.py +164 -0
- keras_hub/src/models/falcon/falcon_causal_lm.py +291 -0
- keras_hub/src/models/falcon/falcon_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/falcon/falcon_preprocessor.py +187 -0
- keras_hub/src/models/falcon/falcon_presets.py +30 -0
- keras_hub/src/models/falcon/falcon_tokenizer.py +110 -0
- keras_hub/src/models/falcon/falcon_transformer_decoder.py +255 -0
- keras_hub/src/models/feature_pyramid_backbone.py +73 -0
- keras_hub/src/models/gemma/__init__.py +20 -0
- keras_hub/src/models/gemma/gemma_attention.py +250 -0
- keras_hub/src/models/gemma/gemma_backbone.py +316 -0
- keras_hub/src/models/gemma/gemma_causal_lm.py +448 -0
- keras_hub/src/models/gemma/gemma_causal_lm_preprocessor.py +167 -0
- keras_hub/src/models/gemma/gemma_decoder_block.py +241 -0
- keras_hub/src/models/gemma/gemma_preprocessor.py +191 -0
- keras_hub/src/models/gemma/gemma_presets.py +248 -0
- keras_hub/src/models/gemma/gemma_tokenizer.py +103 -0
- keras_hub/src/models/gemma/rms_normalization.py +40 -0
- keras_hub/src/models/gpt2/__init__.py +20 -0
- keras_hub/src/models/gpt2/gpt2_backbone.py +199 -0
- keras_hub/src/models/gpt2/gpt2_causal_lm.py +437 -0
- keras_hub/src/models/gpt2/gpt2_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/gpt2/gpt2_preprocessor.py +187 -0
- keras_hub/src/models/gpt2/gpt2_presets.py +82 -0
- keras_hub/src/models/gpt2/gpt2_tokenizer.py +110 -0
- keras_hub/src/models/gpt_neo_x/__init__.py +13 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py +251 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_backbone.py +175 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm.py +201 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm_preprocessor.py +141 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_decoder.py +258 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_preprocessor.py +145 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_tokenizer.py +88 -0
- keras_hub/src/models/image_classifier.py +90 -0
- keras_hub/src/models/llama/__init__.py +20 -0
- keras_hub/src/models/llama/llama_attention.py +225 -0
- keras_hub/src/models/llama/llama_backbone.py +188 -0
- keras_hub/src/models/llama/llama_causal_lm.py +327 -0
- keras_hub/src/models/llama/llama_causal_lm_preprocessor.py +170 -0
- keras_hub/src/models/llama/llama_decoder.py +246 -0
- keras_hub/src/models/llama/llama_layernorm.py +48 -0
- keras_hub/src/models/llama/llama_preprocessor.py +189 -0
- keras_hub/src/models/llama/llama_presets.py +80 -0
- keras_hub/src/models/llama/llama_tokenizer.py +84 -0
- keras_hub/src/models/llama3/__init__.py +20 -0
- keras_hub/src/models/llama3/llama3_backbone.py +84 -0
- keras_hub/src/models/llama3/llama3_causal_lm.py +46 -0
- keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/llama3/llama3_preprocessor.py +21 -0
- keras_hub/src/models/llama3/llama3_presets.py +69 -0
- keras_hub/src/models/llama3/llama3_tokenizer.py +63 -0
- keras_hub/src/models/masked_lm.py +101 -0
- keras_hub/src/models/mistral/__init__.py +20 -0
- keras_hub/src/models/mistral/mistral_attention.py +238 -0
- keras_hub/src/models/mistral/mistral_backbone.py +203 -0
- keras_hub/src/models/mistral/mistral_causal_lm.py +328 -0
- keras_hub/src/models/mistral/mistral_causal_lm_preprocessor.py +175 -0
- keras_hub/src/models/mistral/mistral_layer_norm.py +48 -0
- keras_hub/src/models/mistral/mistral_preprocessor.py +190 -0
- keras_hub/src/models/mistral/mistral_presets.py +48 -0
- keras_hub/src/models/mistral/mistral_tokenizer.py +82 -0
- keras_hub/src/models/mistral/mistral_transformer_decoder.py +265 -0
- keras_hub/src/models/mix_transformer/__init__.py +13 -0
- keras_hub/src/models/mix_transformer/mix_transformer_backbone.py +181 -0
- keras_hub/src/models/mix_transformer/mix_transformer_classifier.py +133 -0
- keras_hub/src/models/mix_transformer/mix_transformer_layers.py +300 -0
- keras_hub/src/models/opt/__init__.py +20 -0
- keras_hub/src/models/opt/opt_backbone.py +173 -0
- keras_hub/src/models/opt/opt_causal_lm.py +301 -0
- keras_hub/src/models/opt/opt_causal_lm_preprocessor.py +177 -0
- keras_hub/src/models/opt/opt_preprocessor.py +188 -0
- keras_hub/src/models/opt/opt_presets.py +72 -0
- keras_hub/src/models/opt/opt_tokenizer.py +116 -0
- keras_hub/src/models/pali_gemma/__init__.py +23 -0
- keras_hub/src/models/pali_gemma/pali_gemma_backbone.py +277 -0
- keras_hub/src/models/pali_gemma/pali_gemma_causal_lm.py +313 -0
- keras_hub/src/models/pali_gemma/pali_gemma_causal_lm_preprocessor.py +147 -0
- keras_hub/src/models/pali_gemma/pali_gemma_decoder_block.py +160 -0
- keras_hub/src/models/pali_gemma/pali_gemma_presets.py +78 -0
- keras_hub/src/models/pali_gemma/pali_gemma_tokenizer.py +79 -0
- keras_hub/src/models/pali_gemma/pali_gemma_vit.py +566 -0
- keras_hub/src/models/phi3/__init__.py +20 -0
- keras_hub/src/models/phi3/phi3_attention.py +260 -0
- keras_hub/src/models/phi3/phi3_backbone.py +224 -0
- keras_hub/src/models/phi3/phi3_causal_lm.py +218 -0
- keras_hub/src/models/phi3/phi3_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/phi3/phi3_decoder.py +260 -0
- keras_hub/src/models/phi3/phi3_layernorm.py +48 -0
- keras_hub/src/models/phi3/phi3_preprocessor.py +190 -0
- keras_hub/src/models/phi3/phi3_presets.py +50 -0
- keras_hub/src/models/phi3/phi3_rotary_embedding.py +137 -0
- keras_hub/src/models/phi3/phi3_tokenizer.py +94 -0
- keras_hub/src/models/preprocessor.py +207 -0
- keras_hub/src/models/resnet/__init__.py +13 -0
- keras_hub/src/models/resnet/resnet_backbone.py +612 -0
- keras_hub/src/models/resnet/resnet_image_classifier.py +136 -0
- keras_hub/src/models/roberta/__init__.py +20 -0
- keras_hub/src/models/roberta/roberta_backbone.py +184 -0
- keras_hub/src/models/roberta/roberta_classifier.py +209 -0
- keras_hub/src/models/roberta/roberta_masked_lm.py +136 -0
- keras_hub/src/models/roberta/roberta_masked_lm_preprocessor.py +198 -0
- keras_hub/src/models/roberta/roberta_preprocessor.py +192 -0
- keras_hub/src/models/roberta/roberta_presets.py +43 -0
- keras_hub/src/models/roberta/roberta_tokenizer.py +132 -0
- keras_hub/src/models/seq_2_seq_lm.py +54 -0
- keras_hub/src/models/t5/__init__.py +20 -0
- keras_hub/src/models/t5/t5_backbone.py +261 -0
- keras_hub/src/models/t5/t5_layer_norm.py +35 -0
- keras_hub/src/models/t5/t5_multi_head_attention.py +324 -0
- keras_hub/src/models/t5/t5_presets.py +95 -0
- keras_hub/src/models/t5/t5_tokenizer.py +100 -0
- keras_hub/src/models/t5/t5_transformer_layer.py +178 -0
- keras_hub/src/models/task.py +419 -0
- keras_hub/src/models/vgg/__init__.py +13 -0
- keras_hub/src/models/vgg/vgg_backbone.py +158 -0
- keras_hub/src/models/vgg/vgg_image_classifier.py +124 -0
- keras_hub/src/models/vit_det/__init__.py +13 -0
- keras_hub/src/models/vit_det/vit_det_backbone.py +204 -0
- keras_hub/src/models/vit_det/vit_layers.py +565 -0
- keras_hub/src/models/whisper/__init__.py +20 -0
- keras_hub/src/models/whisper/whisper_audio_feature_extractor.py +260 -0
- keras_hub/src/models/whisper/whisper_backbone.py +305 -0
- keras_hub/src/models/whisper/whisper_cached_multi_head_attention.py +153 -0
- keras_hub/src/models/whisper/whisper_decoder.py +141 -0
- keras_hub/src/models/whisper/whisper_encoder.py +106 -0
- keras_hub/src/models/whisper/whisper_preprocessor.py +326 -0
- keras_hub/src/models/whisper/whisper_presets.py +148 -0
- keras_hub/src/models/whisper/whisper_tokenizer.py +163 -0
- keras_hub/src/models/xlm_roberta/__init__.py +26 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_backbone.py +81 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_classifier.py +225 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm.py +141 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm_preprocessor.py +195 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_preprocessor.py +205 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_presets.py +43 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_tokenizer.py +191 -0
- keras_hub/src/models/xlnet/__init__.py +13 -0
- keras_hub/src/models/xlnet/relative_attention.py +459 -0
- keras_hub/src/models/xlnet/xlnet_backbone.py +222 -0
- keras_hub/src/models/xlnet/xlnet_content_and_query_embedding.py +133 -0
- keras_hub/src/models/xlnet/xlnet_encoder.py +378 -0
- keras_hub/src/samplers/__init__.py +13 -0
- keras_hub/src/samplers/beam_sampler.py +207 -0
- keras_hub/src/samplers/contrastive_sampler.py +231 -0
- keras_hub/src/samplers/greedy_sampler.py +50 -0
- keras_hub/src/samplers/random_sampler.py +77 -0
- keras_hub/src/samplers/sampler.py +237 -0
- keras_hub/src/samplers/serialization.py +97 -0
- keras_hub/src/samplers/top_k_sampler.py +92 -0
- keras_hub/src/samplers/top_p_sampler.py +113 -0
- keras_hub/src/tests/__init__.py +13 -0
- keras_hub/src/tests/test_case.py +608 -0
- keras_hub/src/tokenizers/__init__.py +13 -0
- keras_hub/src/tokenizers/byte_pair_tokenizer.py +638 -0
- keras_hub/src/tokenizers/byte_tokenizer.py +299 -0
- keras_hub/src/tokenizers/sentence_piece_tokenizer.py +267 -0
- keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py +150 -0
- keras_hub/src/tokenizers/tokenizer.py +235 -0
- keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py +355 -0
- keras_hub/src/tokenizers/word_piece_tokenizer.py +544 -0
- keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py +176 -0
- keras_hub/src/utils/__init__.py +13 -0
- keras_hub/src/utils/keras_utils.py +130 -0
- keras_hub/src/utils/pipeline_model.py +293 -0
- keras_hub/src/utils/preset_utils.py +621 -0
- keras_hub/src/utils/python_utils.py +21 -0
- keras_hub/src/utils/tensor_utils.py +206 -0
- keras_hub/src/utils/timm/__init__.py +13 -0
- keras_hub/src/utils/timm/convert.py +37 -0
- keras_hub/src/utils/timm/convert_resnet.py +171 -0
- keras_hub/src/utils/transformers/__init__.py +13 -0
- keras_hub/src/utils/transformers/convert.py +101 -0
- keras_hub/src/utils/transformers/convert_bert.py +173 -0
- keras_hub/src/utils/transformers/convert_distilbert.py +184 -0
- keras_hub/src/utils/transformers/convert_gemma.py +187 -0
- keras_hub/src/utils/transformers/convert_gpt2.py +186 -0
- keras_hub/src/utils/transformers/convert_llama3.py +136 -0
- keras_hub/src/utils/transformers/convert_pali_gemma.py +303 -0
- keras_hub/src/utils/transformers/safetensor_utils.py +97 -0
- keras_hub/src/version_utils.py +23 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/METADATA +34 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/RECORD +297 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/WHEEL +5 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/top_level.txt +1 -0
@@ -0,0 +1,262 @@
|
|
1
|
+
# Copyright 2024 The KerasHub Authors
|
2
|
+
#
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4
|
+
# you may not use this file except in compliance with the License.
|
5
|
+
# You may obtain a copy of the License at
|
6
|
+
#
|
7
|
+
# https://www.apache.org/licenses/LICENSE-2.0
|
8
|
+
#
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
+
# See the License for the specific language governing permissions and
|
13
|
+
# limitations under the License.
|
14
|
+
|
15
|
+
|
16
|
+
import keras
|
17
|
+
from absl import logging
|
18
|
+
|
19
|
+
from keras_hub.src.api_export import keras_hub_export
|
20
|
+
from keras_hub.src.models.bart.bart_preprocessor import BartPreprocessor
|
21
|
+
from keras_hub.src.utils.keras_utils import (
|
22
|
+
convert_inputs_to_list_of_tensor_segments,
|
23
|
+
)
|
24
|
+
from keras_hub.src.utils.tensor_utils import strip_to_ragged
|
25
|
+
|
26
|
+
try:
|
27
|
+
import tensorflow as tf
|
28
|
+
except ImportError:
|
29
|
+
tf = None
|
30
|
+
|
31
|
+
|
32
|
+
@keras_hub_export("keras_hub.models.BartSeq2SeqLMPreprocessor")
|
33
|
+
class BartSeq2SeqLMPreprocessor(BartPreprocessor):
|
34
|
+
"""BART Seq2Seq LM preprocessor.
|
35
|
+
|
36
|
+
This layer is used as preprocessor for seq2seq tasks using the BART model.
|
37
|
+
This class subclasses `keras_hub.models.BartPreprocessor` and keeps most of
|
38
|
+
its functionality. It has two changes from the superclass:
|
39
|
+
|
40
|
+
1. Sets the `y` (label) and `sample_weights` fields by shifting the
|
41
|
+
decoder input sequence one step towards the left. Both these fields are
|
42
|
+
inferred internally, and any passed values will be ignored.
|
43
|
+
2. Drops the last token from the decoder input sequence as it does not have
|
44
|
+
a successor.
|
45
|
+
|
46
|
+
Args:
|
47
|
+
tokenizer: A `keras_hub.models.BartTokenizer` instance.
|
48
|
+
encoder_sequence_length: The length of the packed encoder inputs.
|
49
|
+
decoder_sequence_length: The length of the packed decoder inputs.
|
50
|
+
|
51
|
+
Call arguments:
|
52
|
+
x: A dictionary with `encoder_text` and `decoder_text` as its keys.
|
53
|
+
Each value in the dictionary should be a tensor of single string
|
54
|
+
sequences. Inputs may be batched or unbatched. Raw python inputs
|
55
|
+
will be converted to tensors.
|
56
|
+
y: Label data. Should always be `None` as the layer generates labels by
|
57
|
+
shifting the decoder input sequence one step to the left.
|
58
|
+
sample_weight: Label weights. Should always be `None` as the layer
|
59
|
+
generates label weights by shifting the padding mask one step to the
|
60
|
+
left.
|
61
|
+
|
62
|
+
Examples:
|
63
|
+
|
64
|
+
Directly calling the layer on data
|
65
|
+
```python
|
66
|
+
preprocessor = keras_hub.models.BartPreprocessor.from_preset("bart_base_en")
|
67
|
+
|
68
|
+
# Preprocess unbatched inputs.
|
69
|
+
inputs = {
|
70
|
+
"encoder_text": "The fox was sleeping.",
|
71
|
+
"decoder_text": "The fox was awake."
|
72
|
+
}
|
73
|
+
preprocessor(inputs)
|
74
|
+
|
75
|
+
# Preprocess batched inputs.
|
76
|
+
inputs = {
|
77
|
+
"encoder_text": ["The fox was sleeping.", "The lion was quiet."],
|
78
|
+
"decoder_text": ["The fox was awake.", "The lion was roaring."]
|
79
|
+
}
|
80
|
+
preprocessor(inputs)
|
81
|
+
|
82
|
+
# Custom vocabulary.
|
83
|
+
vocab = {
|
84
|
+
"<s>": 0,
|
85
|
+
"<pad>": 1,
|
86
|
+
"</s>": 2,
|
87
|
+
"Ġafter": 5,
|
88
|
+
"noon": 6,
|
89
|
+
"Ġsun": 7,
|
90
|
+
}
|
91
|
+
merges = ["Ġ a", "Ġ s", "Ġ n", "e r", "n o", "o n", "Ġs u", "Ġa f", "no on"]
|
92
|
+
merges += ["Ġsu n", "Ġaf t", "Ġaft er"]
|
93
|
+
|
94
|
+
tokenizer = keras_hub.models.BartTokenizer(
|
95
|
+
vocabulary=vocab,
|
96
|
+
merges=merges,
|
97
|
+
)
|
98
|
+
preprocessor = keras_hub.models.BartPreprocessor(
|
99
|
+
tokenizer=tokenizer,
|
100
|
+
encoder_sequence_length=20,
|
101
|
+
decoder_sequence_length=10,
|
102
|
+
)
|
103
|
+
inputs = {
|
104
|
+
"encoder_text": "The fox was sleeping.",
|
105
|
+
"decoder_text": "The fox was awake."
|
106
|
+
}
|
107
|
+
preprocessor(inputs)
|
108
|
+
```
|
109
|
+
|
110
|
+
Mapping with `tf.data.Dataset`.
|
111
|
+
```python
|
112
|
+
preprocessor = keras_hub.models.BartPreprocessor.from_preset("bart_base_en")
|
113
|
+
|
114
|
+
# Map single sentences.
|
115
|
+
features = {
|
116
|
+
"encoder_text": tf.constant(
|
117
|
+
["The fox was sleeping.", "The lion was quiet."]
|
118
|
+
),
|
119
|
+
"decoder_text": tf.constant(
|
120
|
+
["The fox was awake.", "The lion was roaring."]
|
121
|
+
)
|
122
|
+
}
|
123
|
+
ds = tf.data.Dataset.from_tensor_slices(features)
|
124
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
125
|
+
```
|
126
|
+
"""
|
127
|
+
|
128
|
+
def call(
|
129
|
+
self,
|
130
|
+
x,
|
131
|
+
y=None,
|
132
|
+
sample_weight=None,
|
133
|
+
*,
|
134
|
+
encoder_sequence_length=None,
|
135
|
+
decoder_sequence_length=None,
|
136
|
+
# `sequence_length` is an alias for `decoder_sequence_length`
|
137
|
+
sequence_length=None,
|
138
|
+
):
|
139
|
+
if y is not None or sample_weight is not None:
|
140
|
+
logging.warning(
|
141
|
+
"`BartSeq2SeqLMPreprocessor` infers `y` and `sample_weight` "
|
142
|
+
"from the provided input data, i.e., `x`. However, non-`None`"
|
143
|
+
"values have been passed for `y` or `sample_weight` or both. "
|
144
|
+
"These values will be ignored."
|
145
|
+
)
|
146
|
+
|
147
|
+
if encoder_sequence_length is None:
|
148
|
+
encoder_sequence_length = self.encoder_sequence_length
|
149
|
+
decoder_sequence_length = decoder_sequence_length or sequence_length
|
150
|
+
if decoder_sequence_length is None:
|
151
|
+
decoder_sequence_length = self.decoder_sequence_length
|
152
|
+
|
153
|
+
x = super().call(
|
154
|
+
x,
|
155
|
+
encoder_sequence_length=encoder_sequence_length,
|
156
|
+
decoder_sequence_length=decoder_sequence_length + 1,
|
157
|
+
)
|
158
|
+
decoder_token_ids = x.pop("decoder_token_ids")
|
159
|
+
decoder_padding_mask = x.pop("decoder_padding_mask")
|
160
|
+
|
161
|
+
# The last token does not have a next token. Hence, we truncate it.
|
162
|
+
x = {
|
163
|
+
**x,
|
164
|
+
"decoder_token_ids": decoder_token_ids[..., :-1],
|
165
|
+
"decoder_padding_mask": decoder_padding_mask[..., :-1],
|
166
|
+
}
|
167
|
+
# Target `y` will be the decoder input sequence shifted one step to the
|
168
|
+
# left (i.e., the next token).
|
169
|
+
y = decoder_token_ids[..., 1:]
|
170
|
+
sample_weight = decoder_padding_mask[..., 1:]
|
171
|
+
return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
|
172
|
+
|
173
|
+
def generate_preprocess(
|
174
|
+
self,
|
175
|
+
x,
|
176
|
+
*,
|
177
|
+
encoder_sequence_length=None,
|
178
|
+
# `sequence_length` is an alias for `decoder_sequence_length`
|
179
|
+
decoder_sequence_length=None,
|
180
|
+
sequence_length=None,
|
181
|
+
):
|
182
|
+
"""Convert encoder and decoder input strings to integer token inputs for generation.
|
183
|
+
|
184
|
+
Similar to calling the layer for training, this method takes in a dict
|
185
|
+
containing `"encoder_text"` and `"decoder_text"`, with strings or tensor
|
186
|
+
strings for values, tokenizes and packs the input, and computes a
|
187
|
+
padding mask masking all inputs not filled in with a padded value.
|
188
|
+
|
189
|
+
Unlike calling the layer for training, this method does not compute
|
190
|
+
labels and will never append a tokenizer.end_token_id to the end of
|
191
|
+
the decoder sequence (as generation is expected to continue at the end
|
192
|
+
of the inputted decoder prompt).
|
193
|
+
"""
|
194
|
+
if not self.built:
|
195
|
+
self.build(None)
|
196
|
+
|
197
|
+
if isinstance(x, dict):
|
198
|
+
encoder_text = x["encoder_text"]
|
199
|
+
decoder_text = x["decoder_text"]
|
200
|
+
else:
|
201
|
+
encoder_text = x
|
202
|
+
# Initialize empty prompt for the decoder.
|
203
|
+
decoder_text = tf.fill((tf.shape(encoder_text)[0],), "")
|
204
|
+
|
205
|
+
if encoder_sequence_length is None:
|
206
|
+
encoder_sequence_length = self.encoder_sequence_length
|
207
|
+
decoder_sequence_length = decoder_sequence_length or sequence_length
|
208
|
+
if decoder_sequence_length is None:
|
209
|
+
decoder_sequence_length = self.decoder_sequence_length
|
210
|
+
|
211
|
+
# Tokenize and pack the encoder inputs.
|
212
|
+
# TODO: Remove `[0]` once we have shifted to `MultiSegmentPacker`.
|
213
|
+
encoder_text = convert_inputs_to_list_of_tensor_segments(encoder_text)[
|
214
|
+
0
|
215
|
+
]
|
216
|
+
encoder_token_ids = self.tokenizer(encoder_text)
|
217
|
+
encoder_token_ids, encoder_padding_mask = self.encoder_packer(
|
218
|
+
encoder_token_ids,
|
219
|
+
sequence_length=encoder_sequence_length,
|
220
|
+
)
|
221
|
+
|
222
|
+
# Tokenize and pack the decoder inputs.
|
223
|
+
decoder_text = convert_inputs_to_list_of_tensor_segments(decoder_text)[
|
224
|
+
0
|
225
|
+
]
|
226
|
+
decoder_token_ids = self.tokenizer(decoder_text)
|
227
|
+
decoder_token_ids, decoder_padding_mask = self.decoder_packer(
|
228
|
+
decoder_token_ids,
|
229
|
+
sequence_length=decoder_sequence_length,
|
230
|
+
add_end_value=False,
|
231
|
+
)
|
232
|
+
|
233
|
+
return {
|
234
|
+
"encoder_token_ids": encoder_token_ids,
|
235
|
+
"encoder_padding_mask": encoder_padding_mask,
|
236
|
+
"decoder_token_ids": decoder_token_ids,
|
237
|
+
"decoder_padding_mask": decoder_padding_mask,
|
238
|
+
}
|
239
|
+
|
240
|
+
def generate_postprocess(
|
241
|
+
self,
|
242
|
+
x,
|
243
|
+
):
|
244
|
+
"""Convert integer token output to strings for generation.
|
245
|
+
|
246
|
+
This method reverses `generate_preprocess()`, by first removing all
|
247
|
+
padding and start/end tokens, and then converting the integer sequence
|
248
|
+
back to a string.
|
249
|
+
"""
|
250
|
+
if not self.built:
|
251
|
+
self.build(None)
|
252
|
+
|
253
|
+
token_ids, padding_mask = (
|
254
|
+
x["decoder_token_ids"],
|
255
|
+
x["decoder_padding_mask"],
|
256
|
+
)
|
257
|
+
ids_to_strip = (
|
258
|
+
self.tokenizer.start_token_id,
|
259
|
+
self.tokenizer.end_token_id,
|
260
|
+
)
|
261
|
+
token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
|
262
|
+
return self.tokenizer.detokenize(token_ids)
|
@@ -0,0 +1,124 @@
|
|
1
|
+
# Copyright 2024 The KerasHub Authors
|
2
|
+
#
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4
|
+
# you may not use this file except in compliance with the License.
|
5
|
+
# You may obtain a copy of the License at
|
6
|
+
#
|
7
|
+
# https://www.apache.org/licenses/LICENSE-2.0
|
8
|
+
#
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
+
# See the License for the specific language governing permissions and
|
13
|
+
# limitations under the License.
|
14
|
+
|
15
|
+
|
16
|
+
from keras_hub.src.api_export import keras_hub_export
|
17
|
+
from keras_hub.src.tokenizers.byte_pair_tokenizer import BytePairTokenizer
|
18
|
+
|
19
|
+
|
20
|
+
@keras_hub_export("keras_hub.models.BartTokenizer")
|
21
|
+
class BartTokenizer(BytePairTokenizer):
|
22
|
+
"""A BART tokenizer using Byte-Pair Encoding subword segmentation.
|
23
|
+
|
24
|
+
This tokenizer class will tokenize raw strings into integer sequences and
|
25
|
+
is based on `keras_hub.tokenizers.BytePairTokenizer`. Unlike the
|
26
|
+
underlying tokenizer, it will check for all special tokens needed by BART
|
27
|
+
models and provides a `from_preset()` method to automatically download
|
28
|
+
a matching vocabulary for a BART preset.
|
29
|
+
|
30
|
+
This tokenizer does not provide truncation or padding of inputs. It can be
|
31
|
+
combined with a `keras_hub.models.BartPreprocessor` layer for input
|
32
|
+
packing.
|
33
|
+
|
34
|
+
If input is a batch of strings (rank > 0), the layer will output a
|
35
|
+
`tf.RaggedTensor` where the last dimension of the output is ragged.
|
36
|
+
|
37
|
+
If input is a scalar string (rank == 0), the layer will output a dense
|
38
|
+
`tf.Tensor` with static shape `[None]`.
|
39
|
+
|
40
|
+
Args:
|
41
|
+
vocabulary: string or dict, maps token to integer ids. If it is a
|
42
|
+
string, it should be the file path to a json file.
|
43
|
+
merges: string or list, contains the merge rule. If it is a string,
|
44
|
+
it should be the file path to merge rules. The merge rule file
|
45
|
+
should have one merge rule per line. Every merge rule contains
|
46
|
+
merge entities separated by a space.
|
47
|
+
|
48
|
+
Examples:
|
49
|
+
|
50
|
+
```python
|
51
|
+
# Unbatched input.
|
52
|
+
tokenizer = keras_hub.models.BartTokenizer.from_preset(
|
53
|
+
"bart_base_en",
|
54
|
+
)
|
55
|
+
tokenizer("The quick brown fox jumped.")
|
56
|
+
|
57
|
+
# Batched input.
|
58
|
+
tokenizer(["The quick brown fox jumped.", "The fox slept."])
|
59
|
+
|
60
|
+
# Detokenization.
|
61
|
+
tokenizer.detokenize(tokenizer("The quick brown fox jumped."))
|
62
|
+
|
63
|
+
# Custom vocabulary.
|
64
|
+
vocab = {"<s>": 0, "<pad>": 1, "</s>": 2, "<mask>": 3}
|
65
|
+
vocab = {**vocab, "a": 4, "Ġquick": 5, "Ġfox": 6}
|
66
|
+
merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
|
67
|
+
merges += ["Ġ f", "o x", "Ġf ox"]
|
68
|
+
tokenizer = keras_hub.models.BartTokenizer(
|
69
|
+
vocabulary=vocab,
|
70
|
+
merges=merges,
|
71
|
+
)
|
72
|
+
tokenizer("The quick brown fox jumped.")
|
73
|
+
```
|
74
|
+
"""
|
75
|
+
|
76
|
+
def __init__(
|
77
|
+
self,
|
78
|
+
vocabulary=None,
|
79
|
+
merges=None,
|
80
|
+
**kwargs,
|
81
|
+
):
|
82
|
+
self.start_token = "<s>"
|
83
|
+
self.pad_token = "<pad>"
|
84
|
+
self.end_token = "</s>"
|
85
|
+
|
86
|
+
super().__init__(
|
87
|
+
vocabulary=vocabulary,
|
88
|
+
merges=merges,
|
89
|
+
unsplittable_tokens=[
|
90
|
+
self.start_token,
|
91
|
+
self.pad_token,
|
92
|
+
self.end_token,
|
93
|
+
],
|
94
|
+
**kwargs,
|
95
|
+
)
|
96
|
+
|
97
|
+
def set_vocabulary_and_merges(self, vocabulary, merges):
|
98
|
+
super().set_vocabulary_and_merges(vocabulary, merges)
|
99
|
+
|
100
|
+
if vocabulary is not None:
|
101
|
+
# Check for necessary special tokens.
|
102
|
+
for token in [self.start_token, self.pad_token, self.end_token]:
|
103
|
+
if token not in self.vocabulary:
|
104
|
+
raise ValueError(
|
105
|
+
f"Cannot find token `'{token}'` in the provided "
|
106
|
+
f"`vocabulary`. Please provide `'{token}'` in your "
|
107
|
+
"`vocabulary` or use a pretrained `vocabulary` name."
|
108
|
+
)
|
109
|
+
|
110
|
+
self.start_token_id = self.token_to_id(self.start_token)
|
111
|
+
self.pad_token_id = self.token_to_id(self.pad_token)
|
112
|
+
self.end_token_id = self.token_to_id(self.end_token)
|
113
|
+
else:
|
114
|
+
self.start_token_id = None
|
115
|
+
self.pad_token_id = None
|
116
|
+
self.end_token_id = None
|
117
|
+
|
118
|
+
def get_config(self):
|
119
|
+
config = super().get_config()
|
120
|
+
# In the constructor, we pass the list of special tokens to the
|
121
|
+
# `unsplittable_tokens` arg of the superclass' constructor. Hence, we
|
122
|
+
# delete it from the config here.
|
123
|
+
del config["unsplittable_tokens"]
|
124
|
+
return config
|
@@ -0,0 +1,23 @@
|
|
1
|
+
# Copyright 2024 The KerasHub Authors
|
2
|
+
#
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4
|
+
# you may not use this file except in compliance with the License.
|
5
|
+
# You may obtain a copy of the License at
|
6
|
+
#
|
7
|
+
# https://www.apache.org/licenses/LICENSE-2.0
|
8
|
+
#
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
+
# See the License for the specific language governing permissions and
|
13
|
+
# limitations under the License.
|
14
|
+
|
15
|
+
from keras_hub.src.models.bert.bert_backbone import BertBackbone
|
16
|
+
from keras_hub.src.models.bert.bert_classifier import BertClassifier
|
17
|
+
from keras_hub.src.models.bert.bert_presets import backbone_presets
|
18
|
+
from keras_hub.src.models.bert.bert_presets import classifier_presets
|
19
|
+
from keras_hub.src.models.bert.bert_tokenizer import BertTokenizer
|
20
|
+
from keras_hub.src.utils.preset_utils import register_presets
|
21
|
+
|
22
|
+
register_presets(backbone_presets, (BertBackbone, BertTokenizer))
|
23
|
+
register_presets(classifier_presets, (BertClassifier, BertTokenizer))
|
@@ -0,0 +1,227 @@
|
|
1
|
+
# Copyright 2024 The KerasHub Authors
|
2
|
+
#
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4
|
+
# you may not use this file except in compliance with the License.
|
5
|
+
# You may obtain a copy of the License at
|
6
|
+
#
|
7
|
+
# https://www.apache.org/licenses/LICENSE-2.0
|
8
|
+
#
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
+
# See the License for the specific language governing permissions and
|
13
|
+
# limitations under the License.
|
14
|
+
|
15
|
+
import keras
|
16
|
+
|
17
|
+
from keras_hub.src.api_export import keras_hub_export
|
18
|
+
from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
|
19
|
+
from keras_hub.src.layers.modeling.reversible_embedding import (
|
20
|
+
ReversibleEmbedding,
|
21
|
+
)
|
22
|
+
from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
|
23
|
+
from keras_hub.src.models.backbone import Backbone
|
24
|
+
from keras_hub.src.utils.keras_utils import gelu_approximate
|
25
|
+
|
26
|
+
|
27
|
+
def bert_kernel_initializer(stddev=0.02):
|
28
|
+
return keras.initializers.TruncatedNormal(stddev=stddev)
|
29
|
+
|
30
|
+
|
31
|
+
@keras_hub_export("keras_hub.models.BertBackbone")
|
32
|
+
class BertBackbone(Backbone):
|
33
|
+
"""A BERT encoder network.
|
34
|
+
|
35
|
+
This class implements a bi-directional Transformer-based encoder as
|
36
|
+
described in ["BERT: Pre-training of Deep Bidirectional Transformers for
|
37
|
+
Language Understanding"](https://arxiv.org/abs/1810.04805). It includes the
|
38
|
+
embedding lookups and transformer layers, but not the masked language model
|
39
|
+
or next sentence prediction heads.
|
40
|
+
|
41
|
+
The default constructor gives a fully customizable, randomly initialized
|
42
|
+
BERT encoder with any number of layers, heads, and embedding dimensions. To
|
43
|
+
load preset architectures and weights, use the `from_preset()` constructor.
|
44
|
+
|
45
|
+
Disclaimer: Pre-trained models are provided on an "as is" basis, without
|
46
|
+
warranties or conditions of any kind.
|
47
|
+
|
48
|
+
Args:
|
49
|
+
vocabulary_size: int. The size of the token vocabulary.
|
50
|
+
num_layers: int. The number of transformer layers.
|
51
|
+
num_heads: int. The number of attention heads for each transformer.
|
52
|
+
The hidden size must be divisible by the number of attention heads.
|
53
|
+
hidden_dim: int. The size of the transformer encoding and pooler layers.
|
54
|
+
intermediate_dim: int. The output dimension of the first Dense layer in
|
55
|
+
a two-layer feedforward network for each transformer.
|
56
|
+
dropout: float. Dropout probability for the Transformer encoder.
|
57
|
+
max_sequence_length: int. The maximum sequence length that this encoder
|
58
|
+
can consume. If None, `max_sequence_length` uses the value from
|
59
|
+
sequence length. This determines the variable shape for positional
|
60
|
+
embeddings.
|
61
|
+
num_segments: int. The number of types that the 'segment_ids' input can
|
62
|
+
take.
|
63
|
+
dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
|
64
|
+
for model computations and weights. Note that some computations,
|
65
|
+
such as softmax and layer normalization, will always be done at
|
66
|
+
float32 precision regardless of dtype.
|
67
|
+
|
68
|
+
Examples:
|
69
|
+
```python
|
70
|
+
input_data = {
|
71
|
+
"token_ids": np.ones(shape=(1, 12), dtype="int32"),
|
72
|
+
"segment_ids": np.array([[0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0]]),
|
73
|
+
"padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
|
74
|
+
}
|
75
|
+
|
76
|
+
# Pretrained BERT encoder.
|
77
|
+
model = keras_hub.models.BertBackbone.from_preset("bert_base_en_uncased")
|
78
|
+
model(input_data)
|
79
|
+
|
80
|
+
# Randomly initialized BERT encoder with a custom config.
|
81
|
+
model = keras_hub.models.BertBackbone(
|
82
|
+
vocabulary_size=30552,
|
83
|
+
num_layers=4,
|
84
|
+
num_heads=4,
|
85
|
+
hidden_dim=256,
|
86
|
+
intermediate_dim=512,
|
87
|
+
max_sequence_length=128,
|
88
|
+
)
|
89
|
+
model(input_data)
|
90
|
+
```
|
91
|
+
"""
|
92
|
+
|
93
|
+
def __init__(
|
94
|
+
self,
|
95
|
+
vocabulary_size,
|
96
|
+
num_layers,
|
97
|
+
num_heads,
|
98
|
+
hidden_dim,
|
99
|
+
intermediate_dim,
|
100
|
+
dropout=0.1,
|
101
|
+
max_sequence_length=512,
|
102
|
+
num_segments=2,
|
103
|
+
dtype=None,
|
104
|
+
**kwargs,
|
105
|
+
):
|
106
|
+
# === Layers ===
|
107
|
+
self.token_embedding = ReversibleEmbedding(
|
108
|
+
input_dim=vocabulary_size,
|
109
|
+
output_dim=hidden_dim,
|
110
|
+
embeddings_initializer=bert_kernel_initializer(),
|
111
|
+
dtype=dtype,
|
112
|
+
name="token_embedding",
|
113
|
+
)
|
114
|
+
self.position_embedding = PositionEmbedding(
|
115
|
+
initializer=bert_kernel_initializer(),
|
116
|
+
sequence_length=max_sequence_length,
|
117
|
+
dtype=dtype,
|
118
|
+
name="position_embedding",
|
119
|
+
)
|
120
|
+
self.segment_embedding = keras.layers.Embedding(
|
121
|
+
input_dim=num_segments,
|
122
|
+
output_dim=hidden_dim,
|
123
|
+
embeddings_initializer=bert_kernel_initializer(),
|
124
|
+
dtype=dtype,
|
125
|
+
name="segment_embedding",
|
126
|
+
)
|
127
|
+
self.embeddings_add = keras.layers.Add(
|
128
|
+
dtype=dtype,
|
129
|
+
name="embeddings_add",
|
130
|
+
)
|
131
|
+
self.embeddings_layer_norm = keras.layers.LayerNormalization(
|
132
|
+
axis=-1,
|
133
|
+
epsilon=1e-12,
|
134
|
+
dtype=dtype,
|
135
|
+
name="embeddings_layer_norm",
|
136
|
+
)
|
137
|
+
self.embeddings_dropout = keras.layers.Dropout(
|
138
|
+
dropout,
|
139
|
+
dtype=dtype,
|
140
|
+
name="embeddings_dropout",
|
141
|
+
)
|
142
|
+
self.transformer_layers = []
|
143
|
+
for i in range(num_layers):
|
144
|
+
layer = TransformerEncoder(
|
145
|
+
num_heads=num_heads,
|
146
|
+
intermediate_dim=intermediate_dim,
|
147
|
+
activation=gelu_approximate,
|
148
|
+
dropout=dropout,
|
149
|
+
layer_norm_epsilon=1e-12,
|
150
|
+
kernel_initializer=bert_kernel_initializer(),
|
151
|
+
dtype=dtype,
|
152
|
+
name=f"transformer_layer_{i}",
|
153
|
+
)
|
154
|
+
self.transformer_layers.append(layer)
|
155
|
+
self.pooled_dense = keras.layers.Dense(
|
156
|
+
hidden_dim,
|
157
|
+
kernel_initializer=bert_kernel_initializer(),
|
158
|
+
activation="tanh",
|
159
|
+
dtype=dtype,
|
160
|
+
name="pooled_dense",
|
161
|
+
)
|
162
|
+
|
163
|
+
# === Functional Model ===
|
164
|
+
token_id_input = keras.Input(
|
165
|
+
shape=(None,), dtype="int32", name="token_ids"
|
166
|
+
)
|
167
|
+
segment_id_input = keras.Input(
|
168
|
+
shape=(None,), dtype="int32", name="segment_ids"
|
169
|
+
)
|
170
|
+
padding_mask_input = keras.Input(
|
171
|
+
shape=(None,), dtype="int32", name="padding_mask"
|
172
|
+
)
|
173
|
+
# Embed tokens, positions, and segment ids.
|
174
|
+
tokens = self.token_embedding(token_id_input)
|
175
|
+
positions = self.position_embedding(tokens)
|
176
|
+
segments = self.segment_embedding(segment_id_input)
|
177
|
+
# Sum, normalize and apply dropout to embeddings.
|
178
|
+
x = self.embeddings_add((tokens, positions, segments))
|
179
|
+
x = self.embeddings_layer_norm(x)
|
180
|
+
x = self.embeddings_dropout(x)
|
181
|
+
for transformer_layer in self.transformer_layers:
|
182
|
+
x = transformer_layer(x, padding_mask=padding_mask_input)
|
183
|
+
# Construct the two BERT outputs. The pooled output is a dense layer on
|
184
|
+
# top of the [CLS] token.
|
185
|
+
sequence_output = x
|
186
|
+
cls_token_index = 0
|
187
|
+
pooled_output = self.pooled_dense(x[:, cls_token_index, :])
|
188
|
+
super().__init__(
|
189
|
+
inputs={
|
190
|
+
"token_ids": token_id_input,
|
191
|
+
"segment_ids": segment_id_input,
|
192
|
+
"padding_mask": padding_mask_input,
|
193
|
+
},
|
194
|
+
outputs={
|
195
|
+
"sequence_output": sequence_output,
|
196
|
+
"pooled_output": pooled_output,
|
197
|
+
},
|
198
|
+
dtype=dtype,
|
199
|
+
**kwargs,
|
200
|
+
)
|
201
|
+
|
202
|
+
# === Config ===
|
203
|
+
self.vocabulary_size = vocabulary_size
|
204
|
+
self.num_layers = num_layers
|
205
|
+
self.num_heads = num_heads
|
206
|
+
self.hidden_dim = hidden_dim
|
207
|
+
self.intermediate_dim = intermediate_dim
|
208
|
+
self.dropout = dropout
|
209
|
+
self.max_sequence_length = max_sequence_length
|
210
|
+
self.num_segments = num_segments
|
211
|
+
self.cls_token_index = cls_token_index
|
212
|
+
|
213
|
+
def get_config(self):
|
214
|
+
config = super().get_config()
|
215
|
+
config.update(
|
216
|
+
{
|
217
|
+
"vocabulary_size": self.vocabulary_size,
|
218
|
+
"num_layers": self.num_layers,
|
219
|
+
"num_heads": self.num_heads,
|
220
|
+
"hidden_dim": self.hidden_dim,
|
221
|
+
"intermediate_dim": self.intermediate_dim,
|
222
|
+
"dropout": self.dropout,
|
223
|
+
"max_sequence_length": self.max_sequence_length,
|
224
|
+
"num_segments": self.num_segments,
|
225
|
+
}
|
226
|
+
)
|
227
|
+
return config
|