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 +1,175 @@
|
|
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
|
+
|
18
|
+
from keras_hub.src.api_export import keras_hub_export
|
19
|
+
from keras_hub.src.layers.preprocessing.multi_segment_packer import (
|
20
|
+
MultiSegmentPacker,
|
21
|
+
)
|
22
|
+
from keras_hub.src.models.distil_bert.distil_bert_tokenizer import (
|
23
|
+
DistilBertTokenizer,
|
24
|
+
)
|
25
|
+
from keras_hub.src.models.preprocessor import Preprocessor
|
26
|
+
from keras_hub.src.utils.keras_utils import (
|
27
|
+
convert_inputs_to_list_of_tensor_segments,
|
28
|
+
)
|
29
|
+
|
30
|
+
|
31
|
+
@keras_hub_export("keras_hub.models.DistilBertPreprocessor")
|
32
|
+
class DistilBertPreprocessor(Preprocessor):
|
33
|
+
"""A DistilBERT preprocessing layer which tokenizes and packs inputs.
|
34
|
+
|
35
|
+
This preprocessing layer will do three things:
|
36
|
+
|
37
|
+
1. Tokenize any number of input segments using the `tokenizer`.
|
38
|
+
2. Pack the inputs together using a `keras_hub.layers.MultiSegmentPacker`.
|
39
|
+
with the appropriate `"[CLS]"`, `"[SEP]"` and `"[PAD]"` tokens.
|
40
|
+
3. Construct a dictionary of with keys `"token_ids"` and `"padding_mask"`,
|
41
|
+
that can be passed directly to a DistilBERT model.
|
42
|
+
|
43
|
+
This layer can be used directly with `tf.data.Dataset.map` to preprocess
|
44
|
+
string data in the `(x, y, sample_weight)` format used by
|
45
|
+
`keras.Model.fit`.
|
46
|
+
|
47
|
+
Args:
|
48
|
+
tokenizer: A `keras_hub.models.DistilBertTokenizer` instance.
|
49
|
+
sequence_length: The length of the packed inputs.
|
50
|
+
truncate: string. The algorithm to truncate a list of batched segments
|
51
|
+
to fit within `sequence_length`. The value can be either
|
52
|
+
`round_robin` or `waterfall`:
|
53
|
+
- `"round_robin"`: Available space is assigned one token at a
|
54
|
+
time in a round-robin fashion to the inputs that still need
|
55
|
+
some, until the limit is reached.
|
56
|
+
- `"waterfall"`: The allocation of the budget is done using a
|
57
|
+
"waterfall" algorithm that allocates quota in a
|
58
|
+
left-to-right manner and fills up the buckets until we run
|
59
|
+
out of budget. It supports an arbitrary number of segments.
|
60
|
+
|
61
|
+
Call arguments:
|
62
|
+
x: A tensor of single string sequences, or a tuple of multiple
|
63
|
+
tensor sequences to be packed together. Inputs may be batched or
|
64
|
+
unbatched. For single sequences, raw python inputs will be converted
|
65
|
+
to tensors. For multiple sequences, pass tensors directly.
|
66
|
+
y: Any label data. Will be passed through unaltered.
|
67
|
+
sample_weight: Any label weight data. Will be passed through unaltered.
|
68
|
+
|
69
|
+
Examples:
|
70
|
+
|
71
|
+
Directly calling the layer on data.
|
72
|
+
```python
|
73
|
+
preprocessor = keras_hub.models.DistilBertPreprocessor.from_preset(
|
74
|
+
"distil_bert_base_en_uncased"
|
75
|
+
)
|
76
|
+
preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
|
77
|
+
|
78
|
+
# Custom vocabulary.
|
79
|
+
vocab = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
|
80
|
+
vocab += ["The", "quick", "brown", "fox", "jumped", "."]
|
81
|
+
tokenizer = keras_hub.models.DistilBertTokenizer(vocabulary=vocab)
|
82
|
+
preprocessor = keras_hub.models.DistilBertPreprocessor(tokenizer)
|
83
|
+
preprocessor("The quick brown fox jumped.")
|
84
|
+
```
|
85
|
+
|
86
|
+
Mapping with `tf.data.Dataset`.
|
87
|
+
```python
|
88
|
+
preprocessor = keras_hub.models.DistilBertPreprocessor.from_preset(
|
89
|
+
"distil_bert_base_en_uncased"
|
90
|
+
)
|
91
|
+
|
92
|
+
first = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
|
93
|
+
second = tf.constant(["The fox tripped.", "Oh look, a whale."])
|
94
|
+
label = tf.constant([1, 1])
|
95
|
+
# Map labeled single sentences.
|
96
|
+
ds = tf.data.Dataset.from_tensor_slices((first, label))
|
97
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
98
|
+
|
99
|
+
|
100
|
+
# Map unlabeled single sentences.
|
101
|
+
ds = tf.data.Dataset.from_tensor_slices(first)
|
102
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
103
|
+
|
104
|
+
# Map labeled sentence pairs.
|
105
|
+
ds = tf.data.Dataset.from_tensor_slices(((first, second), label))
|
106
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
107
|
+
# Map unlabeled sentence pairs.
|
108
|
+
ds = tf.data.Dataset.from_tensor_slices((first, second))
|
109
|
+
|
110
|
+
# Watch out for tf.data's default unpacking of tuples here!
|
111
|
+
# Best to invoke the `preprocessor` directly in this case.
|
112
|
+
ds = ds.map(
|
113
|
+
lambda first, second: preprocessor(x=(first, second)),
|
114
|
+
num_parallel_calls=tf.data.AUTOTUNE,
|
115
|
+
)
|
116
|
+
```
|
117
|
+
"""
|
118
|
+
|
119
|
+
tokenizer_cls = DistilBertTokenizer
|
120
|
+
|
121
|
+
def __init__(
|
122
|
+
self,
|
123
|
+
tokenizer,
|
124
|
+
sequence_length=512,
|
125
|
+
truncate="round_robin",
|
126
|
+
**kwargs,
|
127
|
+
):
|
128
|
+
super().__init__(**kwargs)
|
129
|
+
self.tokenizer = tokenizer
|
130
|
+
self.packer = None
|
131
|
+
self.sequence_length = sequence_length
|
132
|
+
self.truncate = truncate
|
133
|
+
|
134
|
+
def build(self, input_shape):
|
135
|
+
super().build(input_shape)
|
136
|
+
# Defer masker creation to `build()` so that we can be sure tokenizer
|
137
|
+
# assets have loaded when restoring a saved model.
|
138
|
+
self.packer = MultiSegmentPacker(
|
139
|
+
start_value=self.tokenizer.cls_token_id,
|
140
|
+
end_value=self.tokenizer.sep_token_id,
|
141
|
+
pad_value=self.tokenizer.pad_token_id,
|
142
|
+
truncate=self.truncate,
|
143
|
+
sequence_length=self.sequence_length,
|
144
|
+
)
|
145
|
+
|
146
|
+
def call(self, x, y=None, sample_weight=None):
|
147
|
+
x = convert_inputs_to_list_of_tensor_segments(x)
|
148
|
+
x = [self.tokenizer(segment) for segment in x]
|
149
|
+
token_ids, _ = self.packer(x)
|
150
|
+
x = {
|
151
|
+
"token_ids": token_ids,
|
152
|
+
"padding_mask": token_ids != self.tokenizer.pad_token_id,
|
153
|
+
}
|
154
|
+
return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
|
155
|
+
|
156
|
+
def get_config(self):
|
157
|
+
config = super().get_config()
|
158
|
+
config.update(
|
159
|
+
{
|
160
|
+
"sequence_length": self.sequence_length,
|
161
|
+
"truncate": self.truncate,
|
162
|
+
}
|
163
|
+
)
|
164
|
+
return config
|
165
|
+
|
166
|
+
@property
|
167
|
+
def sequence_length(self):
|
168
|
+
"""The padded length of model input sequences."""
|
169
|
+
return self._sequence_length
|
170
|
+
|
171
|
+
@sequence_length.setter
|
172
|
+
def sequence_length(self, value):
|
173
|
+
self._sequence_length = value
|
174
|
+
if self.packer is not None:
|
175
|
+
self.packer.sequence_length = value
|
@@ -0,0 +1,57 @@
|
|
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
|
+
"""DistilBERT model preset configurations."""
|
15
|
+
|
16
|
+
backbone_presets = {
|
17
|
+
"distil_bert_base_en_uncased": {
|
18
|
+
"metadata": {
|
19
|
+
"description": (
|
20
|
+
"6-layer DistilBERT model where all input is lowercased. "
|
21
|
+
"Trained on English Wikipedia + BooksCorpus using BERT as the "
|
22
|
+
"teacher model."
|
23
|
+
),
|
24
|
+
"params": 66362880,
|
25
|
+
"official_name": "DistilBERT",
|
26
|
+
"path": "distil_bert",
|
27
|
+
"model_card": "https://huggingface.co/distilbert-base-uncased",
|
28
|
+
},
|
29
|
+
"kaggle_handle": "kaggle://keras/distil_bert/keras/distil_bert_base_en_uncased/2",
|
30
|
+
},
|
31
|
+
"distil_bert_base_en": {
|
32
|
+
"metadata": {
|
33
|
+
"description": (
|
34
|
+
"6-layer DistilBERT model where case is maintained. "
|
35
|
+
"Trained on English Wikipedia + BooksCorpus using BERT as the "
|
36
|
+
"teacher model."
|
37
|
+
),
|
38
|
+
"params": 65190912,
|
39
|
+
"official_name": "DistilBERT",
|
40
|
+
"path": "distil_bert",
|
41
|
+
"model_card": "https://huggingface.co/distilbert-base-cased",
|
42
|
+
},
|
43
|
+
"kaggle_handle": "kaggle://keras/distil_bert/keras/distil_bert_base_en/2",
|
44
|
+
},
|
45
|
+
"distil_bert_base_multi": {
|
46
|
+
"metadata": {
|
47
|
+
"description": (
|
48
|
+
"6-layer DistilBERT model where case is maintained. Trained on Wikipedias of 104 languages"
|
49
|
+
),
|
50
|
+
"params": 134734080,
|
51
|
+
"official_name": "DistilBERT",
|
52
|
+
"path": "distil_bert",
|
53
|
+
"model_card": "https://huggingface.co/distilbert-base-multilingual-cased",
|
54
|
+
},
|
55
|
+
"kaggle_handle": "kaggle://keras/distil_bert/keras/distil_bert_base_multi/2",
|
56
|
+
},
|
57
|
+
}
|
@@ -0,0 +1,114 @@
|
|
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.word_piece_tokenizer import WordPieceTokenizer
|
18
|
+
|
19
|
+
|
20
|
+
@keras_hub_export("keras_hub.models.DistilBertTokenizer")
|
21
|
+
class DistilBertTokenizer(WordPieceTokenizer):
|
22
|
+
"""A DistilBERT tokenizer using WordPiece subword segmentation.
|
23
|
+
|
24
|
+
This tokenizer class will tokenize raw strings into integer sequences and
|
25
|
+
is based on `keras_hub.tokenizers.WordPieceTokenizer`. Unlike the
|
26
|
+
underlying tokenizer, it will check for all special tokens needed by DistilBERT
|
27
|
+
models and provides a `from_preset()` method to automatically download
|
28
|
+
a matching vocabulary for a DistilBERT preset.
|
29
|
+
|
30
|
+
This tokenizer does not provide truncation or padding of inputs. It can be
|
31
|
+
combined with a `keras_hub.models.DistilBertPreprocessor` layer for input packing.
|
32
|
+
|
33
|
+
If input is a batch of strings (rank > 0), the layer will output a
|
34
|
+
`tf.RaggedTensor` where the last dimension of the output is ragged.
|
35
|
+
|
36
|
+
If input is a scalar string (rank == 0), the layer will output a dense
|
37
|
+
`tf.Tensor` with static shape `[None]`.
|
38
|
+
|
39
|
+
Args:
|
40
|
+
vocabulary: A list of strings or a string filename path. If
|
41
|
+
passing a list, each element of the list should be a single word
|
42
|
+
piece token string. If passing a filename, the file should be a
|
43
|
+
plain text file containing a single word piece token per line.
|
44
|
+
lowercase: If `True`, the input text will be first lowered before
|
45
|
+
tokenization.
|
46
|
+
special_tokens_in_strings: bool. A bool to indicate if the tokenizer
|
47
|
+
should expect special tokens in input strings that should be
|
48
|
+
tokenized and mapped correctly to their ids. Defaults to False.
|
49
|
+
|
50
|
+
Examples:
|
51
|
+
|
52
|
+
```python
|
53
|
+
# Unbatched input.
|
54
|
+
tokenizer = keras_hub.models.DistilBertTokenizer.from_preset(
|
55
|
+
"distil_bert_base_en_uncased",
|
56
|
+
)
|
57
|
+
tokenizer("The quick brown fox jumped.")
|
58
|
+
|
59
|
+
# Batched input.
|
60
|
+
tokenizer(["The quick brown fox jumped.", "The fox slept."])
|
61
|
+
|
62
|
+
# Detokenization.
|
63
|
+
tokenizer.detokenize(tokenizer("The quick brown fox jumped."))
|
64
|
+
|
65
|
+
# Custom vocabulary.
|
66
|
+
vocab = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
|
67
|
+
vocab += ["The", "quick", "brown", "fox", "jumped", "."]
|
68
|
+
tokenizer = keras_hub.models.DistilBertTokenizer(vocabulary=vocab)
|
69
|
+
tokenizer("The quick brown fox jumped.")
|
70
|
+
```
|
71
|
+
"""
|
72
|
+
|
73
|
+
def __init__(
|
74
|
+
self,
|
75
|
+
vocabulary,
|
76
|
+
lowercase=False,
|
77
|
+
special_tokens_in_strings=False,
|
78
|
+
**kwargs,
|
79
|
+
):
|
80
|
+
self.cls_token = "[CLS]"
|
81
|
+
self.sep_token = "[SEP]"
|
82
|
+
self.pad_token = "[PAD]"
|
83
|
+
self.mask_token = "[MASK]"
|
84
|
+
super().__init__(
|
85
|
+
vocabulary=vocabulary,
|
86
|
+
lowercase=lowercase,
|
87
|
+
special_tokens=[
|
88
|
+
self.cls_token,
|
89
|
+
self.sep_token,
|
90
|
+
self.pad_token,
|
91
|
+
self.mask_token,
|
92
|
+
],
|
93
|
+
special_tokens_in_strings=special_tokens_in_strings,
|
94
|
+
**kwargs,
|
95
|
+
)
|
96
|
+
|
97
|
+
def set_vocabulary(self, vocabulary):
|
98
|
+
super().set_vocabulary(vocabulary)
|
99
|
+
|
100
|
+
if vocabulary is not None:
|
101
|
+
self.cls_token_id = self.token_to_id(self.cls_token)
|
102
|
+
self.sep_token_id = self.token_to_id(self.sep_token)
|
103
|
+
self.pad_token_id = self.token_to_id(self.pad_token)
|
104
|
+
self.mask_token_id = self.token_to_id(self.mask_token)
|
105
|
+
else:
|
106
|
+
self.cls_token_id = None
|
107
|
+
self.sep_token_id = None
|
108
|
+
self.pad_token_id = None
|
109
|
+
self.mask_token_id = None
|
110
|
+
|
111
|
+
def get_config(self):
|
112
|
+
config = super().get_config()
|
113
|
+
del config["special_tokens"] # Not configurable; set in __init__.
|
114
|
+
return config
|
@@ -0,0 +1,20 @@
|
|
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.electra.electra_backbone import ElectraBackbone
|
16
|
+
from keras_hub.src.models.electra.electra_presets import backbone_presets
|
17
|
+
from keras_hub.src.models.electra.electra_tokenizer import ElectraTokenizer
|
18
|
+
from keras_hub.src.utils.preset_utils import register_presets
|
19
|
+
|
20
|
+
register_presets(backbone_presets, (ElectraBackbone, ElectraTokenizer))
|
@@ -0,0 +1,247 @@
|
|
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 electra_kernel_initializer(stddev=0.02):
|
28
|
+
return keras.initializers.TruncatedNormal(stddev=stddev)
|
29
|
+
|
30
|
+
|
31
|
+
@keras_hub_export("keras_hub.models.ElectraBackbone")
|
32
|
+
class ElectraBackbone(Backbone):
|
33
|
+
"""A Electra encoder network.
|
34
|
+
|
35
|
+
This network implements a bidirectional Transformer-based encoder as
|
36
|
+
described in ["Electra: Pre-training Text Encoders as Discriminators Rather
|
37
|
+
Than Generators"](https://arxiv.org/abs/2003.10555). It includes the
|
38
|
+
embedding lookups and transformer layers, but not the masked language model
|
39
|
+
or classification task networks.
|
40
|
+
|
41
|
+
The default constructor gives a fully customizable, randomly initialized
|
42
|
+
ELECTRA encoder with any number of layers, heads, and embedding
|
43
|
+
dimensions. To load preset architectures and weights, use the
|
44
|
+
`from_preset()` constructor.
|
45
|
+
|
46
|
+
Disclaimer: Pre-trained models are provided on an "as is" basis, without
|
47
|
+
warranties or conditions of any kind. The underlying model is provided by a
|
48
|
+
third party and subject to a separate license, available
|
49
|
+
[here](https://huggingface.co/docs/transformers/model_doc/electra#overview).
|
50
|
+
|
51
|
+
Args:
|
52
|
+
vocabulary_size: int. The size of the token vocabulary.
|
53
|
+
num_layers: int. The number of transformer layers.
|
54
|
+
num_heads: int. The number of attention heads for each transformer.
|
55
|
+
The hidden size must be divisible by the number of attention heads.
|
56
|
+
hidden_dim: int. The size of the transformer encoding and pooler layers.
|
57
|
+
embedding_dim: int. The size of the token embeddings.
|
58
|
+
intermediate_dim: int. The output dimension of the first Dense layer in
|
59
|
+
a two-layer feedforward network for each transformer.
|
60
|
+
dropout: float. Dropout probability for the Transformer encoder.
|
61
|
+
max_sequence_length: int. The maximum sequence length that this encoder
|
62
|
+
can consume. If None, `max_sequence_length` uses the value from
|
63
|
+
sequence length. This determines the variable shape for positional
|
64
|
+
embeddings.
|
65
|
+
dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
|
66
|
+
for model computations and weights. Note that some computations,
|
67
|
+
such as softmax and layer normalization, will always be done at
|
68
|
+
float32 precision regardless of dtype.
|
69
|
+
|
70
|
+
Example:
|
71
|
+
```python
|
72
|
+
input_data = {
|
73
|
+
"token_ids": np.ones(shape=(1, 12), dtype="int32"),
|
74
|
+
"segment_ids": np.array([[0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0]]),
|
75
|
+
"padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
|
76
|
+
}
|
77
|
+
|
78
|
+
# Pre-trained ELECTRA encoder.
|
79
|
+
model = keras_hub.models.ElectraBackbone.from_preset(
|
80
|
+
"electra_base_discriminator_en"
|
81
|
+
)
|
82
|
+
model(input_data)
|
83
|
+
|
84
|
+
# Randomly initialized Electra encoder
|
85
|
+
backbone = keras_hub.models.ElectraBackbone(
|
86
|
+
vocabulary_size=1000,
|
87
|
+
num_layers=2,
|
88
|
+
num_heads=2,
|
89
|
+
hidden_dim=32,
|
90
|
+
intermediate_dim=64,
|
91
|
+
dropout=0.1,
|
92
|
+
max_sequence_length=512,
|
93
|
+
)
|
94
|
+
# Returns sequence and pooled outputs.
|
95
|
+
sequence_output, pooled_output = backbone(input_data)
|
96
|
+
```
|
97
|
+
"""
|
98
|
+
|
99
|
+
def __init__(
|
100
|
+
self,
|
101
|
+
vocab_size,
|
102
|
+
num_layers,
|
103
|
+
num_heads,
|
104
|
+
hidden_dim,
|
105
|
+
embedding_dim,
|
106
|
+
intermediate_dim,
|
107
|
+
dropout=0.1,
|
108
|
+
max_sequence_length=512,
|
109
|
+
num_segments=2,
|
110
|
+
dtype=None,
|
111
|
+
**kwargs,
|
112
|
+
):
|
113
|
+
# === Layers ===
|
114
|
+
self.token_embedding = ReversibleEmbedding(
|
115
|
+
input_dim=vocab_size,
|
116
|
+
output_dim=embedding_dim,
|
117
|
+
embeddings_initializer=electra_kernel_initializer(),
|
118
|
+
dtype=dtype,
|
119
|
+
name="token_embedding",
|
120
|
+
)
|
121
|
+
self.position_embedding = PositionEmbedding(
|
122
|
+
initializer=electra_kernel_initializer(),
|
123
|
+
sequence_length=max_sequence_length,
|
124
|
+
dtype=dtype,
|
125
|
+
name="position_embedding",
|
126
|
+
)
|
127
|
+
self.segment_embedding = keras.layers.Embedding(
|
128
|
+
input_dim=num_segments,
|
129
|
+
output_dim=embedding_dim,
|
130
|
+
embeddings_initializer=electra_kernel_initializer(),
|
131
|
+
dtype=dtype,
|
132
|
+
name="segment_embedding",
|
133
|
+
)
|
134
|
+
self.embeddings_add = keras.layers.Add(
|
135
|
+
dtype=dtype,
|
136
|
+
name="embeddings_add",
|
137
|
+
)
|
138
|
+
self.embeddings_layer_norm = keras.layers.LayerNormalization(
|
139
|
+
axis=-1,
|
140
|
+
epsilon=1e-12,
|
141
|
+
dtype=dtype,
|
142
|
+
name="embeddings_layer_norm",
|
143
|
+
)
|
144
|
+
self.embeddings_dropout = keras.layers.Dropout(
|
145
|
+
dropout,
|
146
|
+
dtype=dtype,
|
147
|
+
name="embeddings_dropout",
|
148
|
+
)
|
149
|
+
if hidden_dim != embedding_dim:
|
150
|
+
self.embeddings_projection = keras.layers.Dense(
|
151
|
+
hidden_dim,
|
152
|
+
kernel_initializer=electra_kernel_initializer(),
|
153
|
+
dtype=dtype,
|
154
|
+
name="embeddings_projection",
|
155
|
+
)
|
156
|
+
self.transformer_layers = []
|
157
|
+
for i in range(num_layers):
|
158
|
+
layer = TransformerEncoder(
|
159
|
+
num_heads=num_heads,
|
160
|
+
intermediate_dim=intermediate_dim,
|
161
|
+
activation=gelu_approximate,
|
162
|
+
dropout=dropout,
|
163
|
+
layer_norm_epsilon=1e-12,
|
164
|
+
kernel_initializer=electra_kernel_initializer(),
|
165
|
+
dtype=dtype,
|
166
|
+
name=f"transformer_layer_{i}",
|
167
|
+
)
|
168
|
+
self.transformer_layers.append(layer)
|
169
|
+
self.pooled_dense = keras.layers.Dense(
|
170
|
+
hidden_dim,
|
171
|
+
kernel_initializer=electra_kernel_initializer(),
|
172
|
+
activation="tanh",
|
173
|
+
dtype=dtype,
|
174
|
+
name="pooled_dense",
|
175
|
+
)
|
176
|
+
|
177
|
+
# === Functional Model ===
|
178
|
+
token_id_input = keras.Input(
|
179
|
+
shape=(None,), dtype="int32", name="token_ids"
|
180
|
+
)
|
181
|
+
segment_id_input = keras.Input(
|
182
|
+
shape=(None,), dtype="int32", name="segment_ids"
|
183
|
+
)
|
184
|
+
padding_mask_input = keras.Input(
|
185
|
+
shape=(None,), dtype="int32", name="padding_mask"
|
186
|
+
)
|
187
|
+
# Embed tokens, positions, and segment ids.
|
188
|
+
tokens = self.token_embedding(token_id_input)
|
189
|
+
positions = self.position_embedding(tokens)
|
190
|
+
segments = self.segment_embedding(segment_id_input)
|
191
|
+
# Add all embeddings together.
|
192
|
+
x = self.embeddings_add((tokens, positions, segments))
|
193
|
+
x = self.embeddings_layer_norm(x)
|
194
|
+
x = self.embeddings_dropout(x)
|
195
|
+
if hidden_dim != embedding_dim:
|
196
|
+
x = self.embeddings_projection(x)
|
197
|
+
# Apply successive transformer encoder blocks.
|
198
|
+
for transformer_layer in self.transformer_layers:
|
199
|
+
x = transformer_layer(x, padding_mask=padding_mask_input)
|
200
|
+
# Index of classification token in the vocabulary
|
201
|
+
cls_token_index = 0
|
202
|
+
sequence_output = x
|
203
|
+
# Construct the two ELECTRA outputs. The pooled output is a dense layer on
|
204
|
+
# top of the [CLS] token.
|
205
|
+
pooled_output = self.pooled_dense(x[:, cls_token_index, :])
|
206
|
+
super().__init__(
|
207
|
+
inputs={
|
208
|
+
"token_ids": token_id_input,
|
209
|
+
"segment_ids": segment_id_input,
|
210
|
+
"padding_mask": padding_mask_input,
|
211
|
+
},
|
212
|
+
outputs={
|
213
|
+
"sequence_output": sequence_output,
|
214
|
+
"pooled_output": pooled_output,
|
215
|
+
},
|
216
|
+
dtype=dtype,
|
217
|
+
**kwargs,
|
218
|
+
)
|
219
|
+
|
220
|
+
# === Config ===
|
221
|
+
self.vocab_size = vocab_size
|
222
|
+
self.num_layers = num_layers
|
223
|
+
self.num_heads = num_heads
|
224
|
+
self.hidden_dim = hidden_dim
|
225
|
+
self.embedding_dim = embedding_dim
|
226
|
+
self.intermediate_dim = intermediate_dim
|
227
|
+
self.dropout = dropout
|
228
|
+
self.max_sequence_length = max_sequence_length
|
229
|
+
self.num_segments = num_segments
|
230
|
+
self.cls_token_index = cls_token_index
|
231
|
+
|
232
|
+
def get_config(self):
|
233
|
+
config = super().get_config()
|
234
|
+
config.update(
|
235
|
+
{
|
236
|
+
"vocab_size": self.vocab_size,
|
237
|
+
"num_layers": self.num_layers,
|
238
|
+
"num_heads": self.num_heads,
|
239
|
+
"hidden_dim": self.hidden_dim,
|
240
|
+
"embedding_dim": self.embedding_dim,
|
241
|
+
"intermediate_dim": self.intermediate_dim,
|
242
|
+
"dropout": self.dropout,
|
243
|
+
"max_sequence_length": self.max_sequence_length,
|
244
|
+
"num_segments": self.num_segments,
|
245
|
+
}
|
246
|
+
)
|
247
|
+
return config
|