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 +1,187 @@
|
|
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.start_end_packer import StartEndPacker
|
20
|
+
from keras_hub.src.models.gpt2.gpt2_tokenizer import GPT2Tokenizer
|
21
|
+
from keras_hub.src.models.preprocessor import Preprocessor
|
22
|
+
from keras_hub.src.utils.keras_utils import (
|
23
|
+
convert_inputs_to_list_of_tensor_segments,
|
24
|
+
)
|
25
|
+
|
26
|
+
|
27
|
+
@keras_hub_export("keras_hub.models.GPT2Preprocessor")
|
28
|
+
class GPT2Preprocessor(Preprocessor):
|
29
|
+
"""GPT2 preprocessing layer which tokenizes and packs inputs.
|
30
|
+
|
31
|
+
This preprocessing layer will do 2 things:
|
32
|
+
|
33
|
+
- Tokenize the inputs using the `tokenizer`.
|
34
|
+
- Construct a dictionary with keys `"token_ids"`, `"padding_mask"`, that can
|
35
|
+
be passed directly to a `keras_hub.models.GPT2Backbone`.
|
36
|
+
|
37
|
+
This layer can be used directly with `tf.data.Dataset.map` to preprocess
|
38
|
+
string data in the `(x, y, sample_weight)` format used by
|
39
|
+
`keras.Model.fit`.
|
40
|
+
|
41
|
+
The call method of this layer accepts three arguments, `x`, `y`, and
|
42
|
+
`sample_weight`. `x` can be a python string or tensor representing a single
|
43
|
+
segment, a list of python strings representing a batch of single segments,
|
44
|
+
or a list of tensors representing multiple segments to be packed together.
|
45
|
+
`y` and `sample_weight` are both optional, can have any format, and will be
|
46
|
+
passed through unaltered.
|
47
|
+
|
48
|
+
`GPT2Preprocessor` forces the input to have only one segment, as GPT2 is
|
49
|
+
mainly used for generation tasks. For tasks having multi-segment inputs
|
50
|
+
like "glue/mnli", please use a model designed for classification purposes
|
51
|
+
such as BERT or RoBERTa.
|
52
|
+
|
53
|
+
Args:
|
54
|
+
tokenizer: A `keras_hub.models.GPT2Tokenizer` instance.
|
55
|
+
sequence_length: The length of the packed inputs.
|
56
|
+
add_start_token: If `True`, the preprocessor will prepend the tokenizer
|
57
|
+
start token to each input sequence.
|
58
|
+
add_end_token: If `True`, the preprocessor will append the tokenizer
|
59
|
+
end token to each input sequence.
|
60
|
+
|
61
|
+
Call arguments:
|
62
|
+
x: A string, `tf.Tensor` or list of python strings.
|
63
|
+
y: Any label data. Will be passed through unaltered.
|
64
|
+
sample_weight: Any label weight data. Will be passed through unaltered.
|
65
|
+
sequence_length: Pass to override the configured `sequence_length` of
|
66
|
+
the layer.
|
67
|
+
|
68
|
+
Examples:
|
69
|
+
|
70
|
+
Directly calling the layer on data.
|
71
|
+
```python
|
72
|
+
preprocessor = keras_hub.models.GPT2Preprocessor.from_preset("gpt2_base_en")
|
73
|
+
|
74
|
+
# Tokenize and pack a single sentence.
|
75
|
+
preprocessor("The quick brown fox jumped.")
|
76
|
+
|
77
|
+
# Tokenize a batch of single sentences.
|
78
|
+
preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
|
79
|
+
|
80
|
+
# Custom vocabulary.
|
81
|
+
features = ["a quick fox.", "a fox quick."]
|
82
|
+
vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
|
83
|
+
merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
|
84
|
+
merges += ["Ġ f", "o x", "Ġf ox"]
|
85
|
+
tokenizer = keras_hub.models.GPT2Tokenizer(
|
86
|
+
vocabulary=vocab,
|
87
|
+
merges=merges,
|
88
|
+
)
|
89
|
+
preprocessor = keras_hub.models.GPT2Preprocessor(tokenizer=tokenizer)
|
90
|
+
preprocessor("The quick brown fox jumped.")
|
91
|
+
```
|
92
|
+
|
93
|
+
Mapping with `tf.data.Dataset`.
|
94
|
+
```python
|
95
|
+
preprocessor = keras_hub.models.GPT2Preprocessor.from_preset("gpt2_base_en")
|
96
|
+
|
97
|
+
text = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
|
98
|
+
label = tf.constant([1, 1])
|
99
|
+
|
100
|
+
# Map labeled single sentences.
|
101
|
+
ds = tf.data.Dataset.from_tensor_slices((text, label))
|
102
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
103
|
+
|
104
|
+
# Map unlabeled single sentences.
|
105
|
+
ds = tf.data.Dataset.from_tensor_slices(text)
|
106
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
107
|
+
```
|
108
|
+
"""
|
109
|
+
|
110
|
+
tokenizer_cls = GPT2Tokenizer
|
111
|
+
|
112
|
+
def __init__(
|
113
|
+
self,
|
114
|
+
tokenizer,
|
115
|
+
sequence_length=1024,
|
116
|
+
add_start_token=True,
|
117
|
+
add_end_token=True,
|
118
|
+
**kwargs,
|
119
|
+
):
|
120
|
+
super().__init__(**kwargs)
|
121
|
+
self.tokenizer = tokenizer
|
122
|
+
self.packer = None
|
123
|
+
self.sequence_length = sequence_length
|
124
|
+
self.add_start_token = add_start_token
|
125
|
+
self.add_end_token = add_end_token
|
126
|
+
|
127
|
+
def build(self, input_shape):
|
128
|
+
# Defer packer creation to `build()` so that we can be sure tokenizer
|
129
|
+
# assets have loaded when restoring a saved model.
|
130
|
+
self.packer = StartEndPacker(
|
131
|
+
start_value=self.tokenizer.start_token_id,
|
132
|
+
end_value=self.tokenizer.end_token_id,
|
133
|
+
pad_value=self.tokenizer.pad_token_id,
|
134
|
+
sequence_length=self.sequence_length,
|
135
|
+
return_padding_mask=True,
|
136
|
+
)
|
137
|
+
self.built = True
|
138
|
+
|
139
|
+
def call(
|
140
|
+
self,
|
141
|
+
x,
|
142
|
+
y=None,
|
143
|
+
sample_weight=None,
|
144
|
+
sequence_length=None,
|
145
|
+
):
|
146
|
+
x = convert_inputs_to_list_of_tensor_segments(x)
|
147
|
+
if len(x) != 1:
|
148
|
+
raise ValueError(
|
149
|
+
"GPT2 requires each input feature to contain only "
|
150
|
+
f"one segment, but received {len(x)}. If you are using GPT2 "
|
151
|
+
"for a multi-segment classification task, please refer to "
|
152
|
+
"classification models like BERT or RoBERTa."
|
153
|
+
)
|
154
|
+
sequence_length = sequence_length or self.sequence_length
|
155
|
+
token_ids, padding_mask = self.packer(
|
156
|
+
self.tokenizer(x[0]),
|
157
|
+
sequence_length=sequence_length,
|
158
|
+
add_start_value=self.add_start_token,
|
159
|
+
add_end_value=self.add_end_token,
|
160
|
+
)
|
161
|
+
x = {
|
162
|
+
"token_ids": token_ids,
|
163
|
+
"padding_mask": padding_mask,
|
164
|
+
}
|
165
|
+
return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
|
166
|
+
|
167
|
+
def get_config(self):
|
168
|
+
config = super().get_config()
|
169
|
+
config.update(
|
170
|
+
{
|
171
|
+
"sequence_length": self.sequence_length,
|
172
|
+
"add_start_token": self.add_start_token,
|
173
|
+
"add_end_token": self.add_end_token,
|
174
|
+
}
|
175
|
+
)
|
176
|
+
return config
|
177
|
+
|
178
|
+
@property
|
179
|
+
def sequence_length(self):
|
180
|
+
"""The padded length of model input sequences."""
|
181
|
+
return self._sequence_length
|
182
|
+
|
183
|
+
@sequence_length.setter
|
184
|
+
def sequence_length(self, value):
|
185
|
+
self._sequence_length = value
|
186
|
+
if self.packer is not None:
|
187
|
+
self.packer.sequence_length = value
|
@@ -0,0 +1,82 @@
|
|
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
|
+
"""GPT-2 model preset configurations."""
|
15
|
+
|
16
|
+
# Metadata for loading pretrained model weights.
|
17
|
+
backbone_presets = {
|
18
|
+
"gpt2_base_en": {
|
19
|
+
"metadata": {
|
20
|
+
"description": (
|
21
|
+
"12-layer GPT-2 model where case is maintained. "
|
22
|
+
"Trained on WebText."
|
23
|
+
),
|
24
|
+
"params": 124439808,
|
25
|
+
"official_name": "GPT-2",
|
26
|
+
"path": "gpt2",
|
27
|
+
"model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
|
28
|
+
},
|
29
|
+
"kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_base_en/2",
|
30
|
+
},
|
31
|
+
"gpt2_medium_en": {
|
32
|
+
"metadata": {
|
33
|
+
"description": (
|
34
|
+
"24-layer GPT-2 model where case is maintained. "
|
35
|
+
"Trained on WebText."
|
36
|
+
),
|
37
|
+
"params": 354823168,
|
38
|
+
"official_name": "GPT-2",
|
39
|
+
"path": "gpt2",
|
40
|
+
"model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
|
41
|
+
},
|
42
|
+
"kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_medium_en/2",
|
43
|
+
},
|
44
|
+
"gpt2_large_en": {
|
45
|
+
"metadata": {
|
46
|
+
"description": (
|
47
|
+
"36-layer GPT-2 model where case is maintained. "
|
48
|
+
"Trained on WebText."
|
49
|
+
),
|
50
|
+
"params": 774030080,
|
51
|
+
"official_name": "GPT-2",
|
52
|
+
"path": "gpt2",
|
53
|
+
"model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
|
54
|
+
},
|
55
|
+
"kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_large_en/2",
|
56
|
+
},
|
57
|
+
"gpt2_extra_large_en": {
|
58
|
+
"metadata": {
|
59
|
+
"description": (
|
60
|
+
"48-layer GPT-2 model where case is maintained. "
|
61
|
+
"Trained on WebText."
|
62
|
+
),
|
63
|
+
"params": 1557611200,
|
64
|
+
"official_name": "GPT-2",
|
65
|
+
"path": "gpt2",
|
66
|
+
"model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
|
67
|
+
},
|
68
|
+
"kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_extra_large_en/2",
|
69
|
+
},
|
70
|
+
"gpt2_base_en_cnn_dailymail": {
|
71
|
+
"metadata": {
|
72
|
+
"description": (
|
73
|
+
"12-layer GPT-2 model where case is maintained. "
|
74
|
+
"Finetuned on the CNN/DailyMail summarization dataset."
|
75
|
+
),
|
76
|
+
"params": 124439808,
|
77
|
+
"official_name": "GPT-2",
|
78
|
+
"path": "gpt2",
|
79
|
+
},
|
80
|
+
"kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_base_en_cnn_dailymail/2",
|
81
|
+
},
|
82
|
+
}
|
@@ -0,0 +1,110 @@
|
|
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.GPT2Tokenizer")
|
21
|
+
class GPT2Tokenizer(BytePairTokenizer):
|
22
|
+
"""A GPT-2 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 GPT-2
|
27
|
+
models and provides a `from_preset()` method to automatically download
|
28
|
+
a matching vocabulary for a GPT-2 preset.
|
29
|
+
|
30
|
+
This tokenizer does not provide truncation or padding of inputs.
|
31
|
+
|
32
|
+
If input is a batch of strings (rank > 0), the layer will output a
|
33
|
+
`tf.RaggedTensor` where the last dimension of the output is ragged.
|
34
|
+
|
35
|
+
If input is a scalar string (rank == 0), the layer will output a dense
|
36
|
+
`tf.Tensor` with static shape `[None]`.
|
37
|
+
|
38
|
+
Args:
|
39
|
+
vocabulary: string or dict, maps token to integer ids. If it is a
|
40
|
+
string, it should be the file path to a json file.
|
41
|
+
merges: string or list, contains the merge rule. If it is a string,
|
42
|
+
it should be the file path to merge rules. The merge rule file
|
43
|
+
should have one merge rule per line. Every merge rule contains
|
44
|
+
merge entities separated by a space.
|
45
|
+
|
46
|
+
Examples:
|
47
|
+
|
48
|
+
```python
|
49
|
+
# Unbatched input.
|
50
|
+
tokenizer = keras_hub.models.GPT2Tokenizer.from_preset("gpt2_base_en")
|
51
|
+
tokenizer("The quick brown fox jumped.")
|
52
|
+
|
53
|
+
# Batched input.
|
54
|
+
tokenizer(["The quick brown fox jumped.", "The fox slept."])
|
55
|
+
|
56
|
+
# Detokenization.
|
57
|
+
tokenizer.detokenize(tokenizer("The quick brown fox jumped."))
|
58
|
+
|
59
|
+
# Custom vocabulary.
|
60
|
+
vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
|
61
|
+
merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
|
62
|
+
merges += ["Ġ f", "o x", "Ġf ox"]
|
63
|
+
tokenizer = keras_hub.models.GPT2Tokenizer(vocabulary=vocab, merges=merges)
|
64
|
+
tokenizer("a quick fox.")
|
65
|
+
```
|
66
|
+
"""
|
67
|
+
|
68
|
+
def __init__(
|
69
|
+
self,
|
70
|
+
vocabulary=None,
|
71
|
+
merges=None,
|
72
|
+
**kwargs,
|
73
|
+
):
|
74
|
+
# GPT2 uses the same start as end token, i.e., "<|endoftext|>".
|
75
|
+
self.end_token = self.start_token = "<|endoftext|>"
|
76
|
+
|
77
|
+
super().__init__(
|
78
|
+
vocabulary=vocabulary,
|
79
|
+
merges=merges,
|
80
|
+
unsplittable_tokens=[self.end_token],
|
81
|
+
**kwargs,
|
82
|
+
)
|
83
|
+
|
84
|
+
def set_vocabulary_and_merges(self, vocabulary, merges):
|
85
|
+
super().set_vocabulary_and_merges(vocabulary, merges)
|
86
|
+
|
87
|
+
if vocabulary is not None:
|
88
|
+
# Check for necessary special tokens.
|
89
|
+
if self.end_token not in self.get_vocabulary():
|
90
|
+
raise ValueError(
|
91
|
+
f"Cannot find token `'{self.end_token}'` in the provided "
|
92
|
+
f"`vocabulary`. Please provide `'{self.end_token}'` in "
|
93
|
+
"your `vocabulary` or use a pretrained `vocabulary` name."
|
94
|
+
)
|
95
|
+
|
96
|
+
self.end_token_id = self.token_to_id(self.end_token)
|
97
|
+
self.start_token_id = self.end_token_id
|
98
|
+
self.pad_token_id = 0
|
99
|
+
else:
|
100
|
+
self.end_token_id = None
|
101
|
+
self.start_token_id = None
|
102
|
+
self.pad_token_id = None
|
103
|
+
|
104
|
+
def get_config(self):
|
105
|
+
config = super().get_config()
|
106
|
+
# In the constructor, we pass the list of special tokens to the
|
107
|
+
# `unsplittable_tokens` arg of the superclass' constructor. Hence, we
|
108
|
+
# delete it from the config here.
|
109
|
+
del config["unsplittable_tokens"]
|
110
|
+
return config
|
@@ -0,0 +1,13 @@
|
|
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.
|
@@ -0,0 +1,251 @@
|
|
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
|
+
from keras import ops
|
17
|
+
|
18
|
+
from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
|
19
|
+
from keras_hub.src.utils.keras_utils import clone_initializer
|
20
|
+
|
21
|
+
|
22
|
+
class GPTNeoXAttention(keras.layers.Layer):
|
23
|
+
"""GPTNeoXAttention layer.
|
24
|
+
|
25
|
+
This is an implementation of attention layer as described in the
|
26
|
+
paper ["GPT-NeoX-20B: An Open-Source Autoregressive Language Model"](https://arxiv.org/abs/2204.06745).
|
27
|
+
Effectively, this layer implements Multi-Head Self Attention with a rotary
|
28
|
+
embedding for encoding position information.
|
29
|
+
|
30
|
+
Args:
|
31
|
+
num_heads: int. Number of attention heads.
|
32
|
+
hidden_dim: int. Hidden dimension of the input, i.e., `hidden_states`.
|
33
|
+
bucket_size: int. The size of the relative position
|
34
|
+
buckets. Generally equal to `max_sequence_length // 2`.
|
35
|
+
dropout: float. Dropout probability.
|
36
|
+
kernel_initializer: string or `keras.initializers` initializer.
|
37
|
+
The kernel initializer for the dense layers.
|
38
|
+
bias_initializer: string or `keras.initializers` initializer.
|
39
|
+
The bias initializer for the dense layers.
|
40
|
+
rotary_percentage: float. The percentage by which query, key, value
|
41
|
+
matrices are to be rotated.
|
42
|
+
rotary_max_wavelength: int. The maximum angular wavelength of the
|
43
|
+
sine/cosine curves, for rotary embeddings.
|
44
|
+
max_sequence_length: int. The maximum input sequence length.
|
45
|
+
"""
|
46
|
+
|
47
|
+
def __init__(
|
48
|
+
self,
|
49
|
+
num_heads,
|
50
|
+
hidden_dim,
|
51
|
+
dropout=0.0,
|
52
|
+
kernel_initializer="glorot_uniform",
|
53
|
+
bias_initializer="zeros",
|
54
|
+
rotary_percentage=0.25,
|
55
|
+
rotary_max_wavelength=10000,
|
56
|
+
max_sequence_length=512,
|
57
|
+
**kwargs,
|
58
|
+
):
|
59
|
+
super().__init__(**kwargs)
|
60
|
+
self.num_heads = num_heads
|
61
|
+
self.hidden_dim = hidden_dim
|
62
|
+
self.rotary_percentage = rotary_percentage
|
63
|
+
self.dropout = dropout
|
64
|
+
self.attn_head_size = hidden_dim // num_heads
|
65
|
+
self.rotary_max_wavelength = rotary_max_wavelength
|
66
|
+
self.rotary_dim = int(self.attn_head_size * rotary_percentage)
|
67
|
+
self.rotary_embedding_layer = RotaryEmbedding(
|
68
|
+
max_wavelength=rotary_max_wavelength,
|
69
|
+
dtype=self.dtype_policy,
|
70
|
+
)
|
71
|
+
self.kernel_initializer = keras.initializers.get(kernel_initializer)
|
72
|
+
self.bias_initializer = keras.initializers.get(bias_initializer)
|
73
|
+
self.max_sequence_length = max_sequence_length
|
74
|
+
|
75
|
+
def build(self, input_shape):
|
76
|
+
self._qkv_dense = keras.layers.EinsumDense(
|
77
|
+
equation="abc,cde->abde",
|
78
|
+
output_shape=(None, self.num_heads, 3 * self.attn_head_size),
|
79
|
+
bias_axes="de",
|
80
|
+
**self._get_common_kwargs_for_sublayer(use_bias=True),
|
81
|
+
dtype=self.dtype_policy,
|
82
|
+
name="query_key_value",
|
83
|
+
)
|
84
|
+
self._qkv_dense.build(input_shape)
|
85
|
+
|
86
|
+
self._attn_dropout_layer = keras.layers.Dropout(
|
87
|
+
self.dropout,
|
88
|
+
dtype=self.dtype_policy,
|
89
|
+
name="attention_dropout",
|
90
|
+
)
|
91
|
+
|
92
|
+
self._softmax = keras.layers.Softmax(
|
93
|
+
axis=-1,
|
94
|
+
dtype="float32",
|
95
|
+
name="attention_softmax",
|
96
|
+
)
|
97
|
+
|
98
|
+
# Output.
|
99
|
+
self._output_dense = keras.layers.EinsumDense(
|
100
|
+
equation="abc,cd->abd",
|
101
|
+
output_shape=(None, self.hidden_dim),
|
102
|
+
bias_axes="d",
|
103
|
+
**self._get_common_kwargs_for_sublayer(use_bias=True),
|
104
|
+
dtype=self.dtype_policy,
|
105
|
+
name="attention_output",
|
106
|
+
)
|
107
|
+
|
108
|
+
self._output_dense.build(input_shape)
|
109
|
+
self.built = True
|
110
|
+
|
111
|
+
def _get_common_kwargs_for_sublayer(self, use_bias=True):
|
112
|
+
common_kwargs = {}
|
113
|
+
|
114
|
+
kernel_initializer = clone_initializer(self.kernel_initializer)
|
115
|
+
bias_initializer = clone_initializer(self.bias_initializer)
|
116
|
+
|
117
|
+
common_kwargs["kernel_initializer"] = kernel_initializer
|
118
|
+
if use_bias:
|
119
|
+
common_kwargs["bias_initializer"] = bias_initializer
|
120
|
+
|
121
|
+
return common_kwargs
|
122
|
+
|
123
|
+
def _masked_softmax(self, attention_scores, attention_mask=None):
|
124
|
+
if attention_mask is not None:
|
125
|
+
mask_expansion_axis = -3
|
126
|
+
for _ in range(
|
127
|
+
len(attention_scores.shape) - len(attention_mask.shape)
|
128
|
+
):
|
129
|
+
attention_mask = ops.expand_dims(
|
130
|
+
attention_mask, axis=mask_expansion_axis
|
131
|
+
)
|
132
|
+
return self._softmax(attention_scores, attention_mask)
|
133
|
+
|
134
|
+
def _compute_attention(
|
135
|
+
self, query, key, value, attention_mask=None, training=None
|
136
|
+
):
|
137
|
+
attention_scores = ops.einsum("aecd,abcd->acbe", key, query)
|
138
|
+
|
139
|
+
norm_factor = ops.sqrt(
|
140
|
+
ops.convert_to_tensor(self.attn_head_size, self.compute_dtype)
|
141
|
+
)
|
142
|
+
|
143
|
+
attention_scores /= norm_factor
|
144
|
+
|
145
|
+
attention_scores = self._masked_softmax(
|
146
|
+
attention_scores, attention_mask
|
147
|
+
)
|
148
|
+
attention_scores = self._attn_dropout_layer(
|
149
|
+
attention_scores, training=training
|
150
|
+
)
|
151
|
+
attention_output = ops.einsum(
|
152
|
+
"acbe,aecd->abcd", attention_scores, value
|
153
|
+
)
|
154
|
+
|
155
|
+
return attention_output
|
156
|
+
|
157
|
+
def call(
|
158
|
+
self,
|
159
|
+
hidden_states,
|
160
|
+
attention_mask=None,
|
161
|
+
cache=None,
|
162
|
+
cache_update_index=None,
|
163
|
+
training=None,
|
164
|
+
):
|
165
|
+
query_key_value = self._qkv_dense(hidden_states)
|
166
|
+
|
167
|
+
query = query_key_value[..., : self.attn_head_size]
|
168
|
+
|
169
|
+
if cache is not None:
|
170
|
+
key_cache = cache[:, 0, ...]
|
171
|
+
value_cache = cache[:, 1, ...]
|
172
|
+
if cache_update_index is None:
|
173
|
+
key = key_cache
|
174
|
+
value = value_cache
|
175
|
+
else:
|
176
|
+
key_update = query_key_value[
|
177
|
+
..., self.attn_head_size : 2 * self.attn_head_size
|
178
|
+
]
|
179
|
+
value_update = query_key_value[..., 2 * self.attn_head_size :]
|
180
|
+
start = [0, cache_update_index, 0, 0]
|
181
|
+
key = ops.slice_update(key_cache, start, key_update)
|
182
|
+
value = ops.slice_update(value_cache, start, value_update)
|
183
|
+
cache = ops.stack((key, value), axis=1)
|
184
|
+
else:
|
185
|
+
if cache_update_index is not None:
|
186
|
+
raise ValueError(
|
187
|
+
"`cache_update_index` should not be set if `cache` is "
|
188
|
+
f"`None`. Received: cache={cache}, "
|
189
|
+
f"cache_update_index={cache_update_index}"
|
190
|
+
)
|
191
|
+
key = query_key_value[
|
192
|
+
..., self.attn_head_size : 2 * self.attn_head_size
|
193
|
+
]
|
194
|
+
value = query_key_value[..., 2 * self.attn_head_size :]
|
195
|
+
|
196
|
+
query_rot, query_pass = (
|
197
|
+
query[..., : self.rotary_dim],
|
198
|
+
query[..., self.rotary_dim :],
|
199
|
+
)
|
200
|
+
key_rot, key_pass = (
|
201
|
+
key[..., : self.rotary_dim],
|
202
|
+
key[..., self.rotary_dim :],
|
203
|
+
)
|
204
|
+
|
205
|
+
query_rot = self.rotary_embedding_layer(query_rot)
|
206
|
+
key_rot = self.rotary_embedding_layer(key_rot)
|
207
|
+
|
208
|
+
query = ops.concatenate((query_rot, query_pass), axis=-1)
|
209
|
+
key = ops.concatenate((key_rot, key_pass), axis=-1)
|
210
|
+
|
211
|
+
attention_output = self._compute_attention(
|
212
|
+
query=query,
|
213
|
+
key=key,
|
214
|
+
value=value,
|
215
|
+
attention_mask=attention_mask,
|
216
|
+
training=training,
|
217
|
+
)
|
218
|
+
|
219
|
+
# Reshape `attention_output` to `(batch_size, sequence_length, hidden_dim)`.
|
220
|
+
attention_output = ops.reshape(
|
221
|
+
attention_output,
|
222
|
+
[
|
223
|
+
ops.shape(attention_output)[0],
|
224
|
+
ops.shape(attention_output)[1],
|
225
|
+
self.hidden_dim,
|
226
|
+
],
|
227
|
+
)
|
228
|
+
|
229
|
+
attention_output = self._output_dense(attention_output)
|
230
|
+
|
231
|
+
return attention_output, cache
|
232
|
+
|
233
|
+
def get_config(self):
|
234
|
+
config = super().get_config()
|
235
|
+
config.update(
|
236
|
+
{
|
237
|
+
"num_heads": self.num_heads,
|
238
|
+
"hidden_dim": self.hidden_dim,
|
239
|
+
"dropout": self.dropout,
|
240
|
+
"kernel_initializer": keras.initializers.serialize(
|
241
|
+
self.kernel_initializer
|
242
|
+
),
|
243
|
+
"bias_initializer": keras.initializers.serialize(
|
244
|
+
self.bias_initializer
|
245
|
+
),
|
246
|
+
"rotary_percentage": self.rotary_percentage,
|
247
|
+
"rotary_max_wavelength": self.rotary_max_wavelength,
|
248
|
+
"max_sequence_length": self.max_sequence_length,
|
249
|
+
}
|
250
|
+
)
|
251
|
+
return config
|