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 +1,81 @@
|
|
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.models.roberta import roberta_backbone
|
18
|
+
|
19
|
+
|
20
|
+
@keras_hub_export("keras_hub.models.XLMRobertaBackbone")
|
21
|
+
class XLMRobertaBackbone(roberta_backbone.RobertaBackbone):
|
22
|
+
"""An XLM-RoBERTa encoder network.
|
23
|
+
|
24
|
+
This class implements a bi-directional Transformer-based encoder as
|
25
|
+
described in ["Unsupervised Cross-lingual Representation Learning at Scale"](https://arxiv.org/abs/1911.02116).
|
26
|
+
It includes the embedding lookups and transformer layers, but it does not
|
27
|
+
include the masked language modeling head used during pretraining.
|
28
|
+
|
29
|
+
The default constructor gives a fully customizable, randomly initialized
|
30
|
+
RoBERTa encoder with any number of layers, heads, and embedding dimensions.
|
31
|
+
To load preset architectures and weights, use the `from_preset()`
|
32
|
+
constructor.
|
33
|
+
|
34
|
+
Disclaimer: Pre-trained models are provided on an "as is" basis, without
|
35
|
+
warranties or conditions of any kind. The underlying model is provided by a
|
36
|
+
third party and subject to a separate license, available
|
37
|
+
[here](https://github.com/facebookresearch/fairseq).
|
38
|
+
|
39
|
+
Args:
|
40
|
+
vocabulary_size: int. The size of the token vocabulary.
|
41
|
+
num_layers: int. The number of transformer layers.
|
42
|
+
num_heads: int. The number of attention heads for each transformer.
|
43
|
+
The hidden size must be divisible by the number of attention heads.
|
44
|
+
hidden_dim: int. The size of the transformer encoding layer.
|
45
|
+
intermediate_dim: int. The output dimension of the first Dense layer in
|
46
|
+
a two-layer feedforward network for each transformer.
|
47
|
+
dropout: float. Dropout probability for the Transformer encoder.
|
48
|
+
max_sequence_length: int. The maximum sequence length this encoder can
|
49
|
+
consume. The sequence length of the input must be less than
|
50
|
+
`max_sequence_length` default value. This determines the variable
|
51
|
+
shape for positional embeddings.
|
52
|
+
dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
|
53
|
+
for model computations and weights. Note that some computations,
|
54
|
+
such as softmax and layer normalization, will always be done at
|
55
|
+
float32 precision regardless of dtype.
|
56
|
+
|
57
|
+
Examples:
|
58
|
+
```python
|
59
|
+
input_data = {
|
60
|
+
"token_ids": np.ones(shape=(1, 12), dtype="int32"),
|
61
|
+
"padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
|
62
|
+
}
|
63
|
+
|
64
|
+
# Pretrained XLM-R encoder.
|
65
|
+
model = keras_hub.models.XLMRobertaBackbone.from_preset(
|
66
|
+
"xlm_roberta_base_multi",
|
67
|
+
)
|
68
|
+
model(input_data)
|
69
|
+
|
70
|
+
# Randomly initialized XLM-R model with custom config.
|
71
|
+
model = keras_hub.models.XLMRobertaBackbone(
|
72
|
+
vocabulary_size=250002,
|
73
|
+
num_layers=4,
|
74
|
+
num_heads=4,
|
75
|
+
hidden_dim=256,
|
76
|
+
intermediate_dim=512,
|
77
|
+
max_sequence_length=128
|
78
|
+
)
|
79
|
+
model(input_data)
|
80
|
+
```
|
81
|
+
"""
|
@@ -0,0 +1,225 @@
|
|
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.models.classifier import Classifier
|
20
|
+
from keras_hub.src.models.roberta.roberta_backbone import (
|
21
|
+
roberta_kernel_initializer,
|
22
|
+
)
|
23
|
+
from keras_hub.src.models.xlm_roberta.xlm_roberta_backbone import (
|
24
|
+
XLMRobertaBackbone,
|
25
|
+
)
|
26
|
+
from keras_hub.src.models.xlm_roberta.xlm_roberta_preprocessor import (
|
27
|
+
XLMRobertaPreprocessor,
|
28
|
+
)
|
29
|
+
|
30
|
+
|
31
|
+
@keras_hub_export("keras_hub.models.XLMRobertaClassifier")
|
32
|
+
class XLMRobertaClassifier(Classifier):
|
33
|
+
"""An end-to-end XLM-RoBERTa model for classification tasks.
|
34
|
+
|
35
|
+
This model attaches a classification head to a
|
36
|
+
`keras_hub.model.XLMRobertaBackbone` instance, mapping from the backbone
|
37
|
+
outputs to logits suitable for a classification task. For usage of
|
38
|
+
this model with pre-trained weights, see the `from_preset()` constructor.
|
39
|
+
|
40
|
+
This model can optionally be configured with a `preprocessor` layer, in
|
41
|
+
which case it will automatically apply preprocessing to raw inputs during
|
42
|
+
`fit()`, `predict()`, and `evaluate()`. This is done by default when
|
43
|
+
creating the model with `from_preset()`.
|
44
|
+
|
45
|
+
Disclaimer: Pre-trained models are provided on an "as is" basis, without
|
46
|
+
warranties or conditions of any kind. The underlying model is provided by a
|
47
|
+
third party and subject to a separate license, available
|
48
|
+
[here](https://github.com/facebookresearch/fairseq).
|
49
|
+
|
50
|
+
Args:
|
51
|
+
backbone: A `keras_hub.models.XLMRobertaBackbone` instance.
|
52
|
+
num_classes: int. Number of classes to predict.
|
53
|
+
preprocessor: A `keras_hub.models.XLMRobertaPreprocessor` or `None`. If
|
54
|
+
`None`, this model will not apply preprocessing, and inputs should
|
55
|
+
be preprocessed before calling the model.
|
56
|
+
activation: Optional `str` or callable. The activation function to use
|
57
|
+
on the model outputs. Set `activation="softmax"` to return output
|
58
|
+
probabilities. Defaults to `None`.
|
59
|
+
hidden_dim: int. The size of the pooler layer.
|
60
|
+
dropout: float. The dropout probability value, applied to the pooled
|
61
|
+
output, and after the first dense layer.
|
62
|
+
|
63
|
+
Examples:
|
64
|
+
|
65
|
+
Raw string data.
|
66
|
+
```python
|
67
|
+
features = ["The quick brown fox jumped.", "نسيت الواجب"]
|
68
|
+
labels = [0, 3]
|
69
|
+
|
70
|
+
# Pretrained classifier.
|
71
|
+
classifier = keras_hub.models.XLMRobertaClassifier.from_preset(
|
72
|
+
"xlm_roberta_base_multi",
|
73
|
+
num_classes=4,
|
74
|
+
)
|
75
|
+
classifier.fit(x=features, y=labels, batch_size=2)
|
76
|
+
classifier.predict(x=features, batch_size=2)
|
77
|
+
|
78
|
+
# Re-compile (e.g., with a new learning rate).
|
79
|
+
classifier.compile(
|
80
|
+
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
|
81
|
+
optimizer=keras.optimizers.Adam(5e-5),
|
82
|
+
jit_compile=True,
|
83
|
+
)
|
84
|
+
# Access backbone programmatically (e.g., to change `trainable`).
|
85
|
+
classifier.backbone.trainable = False
|
86
|
+
# Fit again.
|
87
|
+
classifier.fit(x=features, y=labels, batch_size=2)
|
88
|
+
```
|
89
|
+
|
90
|
+
Preprocessed integer data.
|
91
|
+
```python
|
92
|
+
features = {
|
93
|
+
"token_ids": np.ones(shape=(2, 12), dtype="int32"),
|
94
|
+
"padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]] * 2),
|
95
|
+
}
|
96
|
+
labels = [0, 3]
|
97
|
+
|
98
|
+
# Pretrained classifier without preprocessing.
|
99
|
+
classifier = keras_hub.models.XLMRobertaClassifier.from_preset(
|
100
|
+
"xlm_roberta_base_multi",
|
101
|
+
num_classes=4,
|
102
|
+
preprocessor=None,
|
103
|
+
)
|
104
|
+
classifier.fit(x=features, y=labels, batch_size=2)
|
105
|
+
```
|
106
|
+
|
107
|
+
Custom backbone and vocabulary.
|
108
|
+
```python
|
109
|
+
features = ["The quick brown fox jumped.", "نسيت الواجب"]
|
110
|
+
labels = [0, 3]
|
111
|
+
|
112
|
+
def train_sentencepiece(ds, vocab_size):
|
113
|
+
bytes_io = io.BytesIO()
|
114
|
+
sentencepiece.SentencePieceTrainer.train(
|
115
|
+
sentence_iterator=ds.as_numpy_iterator(),
|
116
|
+
model_writer=bytes_io,
|
117
|
+
vocab_size=vocab_size,
|
118
|
+
model_type="WORD",
|
119
|
+
unk_id=0,
|
120
|
+
bos_id=1,
|
121
|
+
eos_id=2,
|
122
|
+
)
|
123
|
+
return bytes_io.getvalue()
|
124
|
+
ds = tf.data.Dataset.from_tensor_slices(
|
125
|
+
["the quick brown fox", "the earth is round"]
|
126
|
+
)
|
127
|
+
proto = train_sentencepiece(ds, vocab_size=10)
|
128
|
+
tokenizer = keras_hub.models.XLMRobertaTokenizer(
|
129
|
+
proto=proto
|
130
|
+
)
|
131
|
+
preprocessor = keras_hub.models.XLMRobertaPreprocessor(
|
132
|
+
tokenizer,
|
133
|
+
sequence_length=128,
|
134
|
+
)
|
135
|
+
backbone = keras_hub.models.XLMRobertaBackbone(
|
136
|
+
vocabulary_size=250002,
|
137
|
+
num_layers=4,
|
138
|
+
num_heads=4,
|
139
|
+
hidden_dim=256,
|
140
|
+
intermediate_dim=512,
|
141
|
+
max_sequence_length=128,
|
142
|
+
)
|
143
|
+
classifier = keras_hub.models.XLMRobertaClassifier(
|
144
|
+
backbone=backbone,
|
145
|
+
preprocessor=preprocessor,
|
146
|
+
num_classes=4,
|
147
|
+
)
|
148
|
+
classifier.fit(x=features, y=labels, batch_size=2)
|
149
|
+
```
|
150
|
+
"""
|
151
|
+
|
152
|
+
backbone_cls = XLMRobertaBackbone
|
153
|
+
preprocessor_cls = XLMRobertaPreprocessor
|
154
|
+
|
155
|
+
def __init__(
|
156
|
+
self,
|
157
|
+
backbone,
|
158
|
+
num_classes,
|
159
|
+
preprocessor=None,
|
160
|
+
activation=None,
|
161
|
+
hidden_dim=None,
|
162
|
+
dropout=0.0,
|
163
|
+
**kwargs,
|
164
|
+
):
|
165
|
+
# === Layers ===
|
166
|
+
self.backbone = backbone
|
167
|
+
self.preprocessor = preprocessor
|
168
|
+
self.pooled_dropout = keras.layers.Dropout(
|
169
|
+
dropout,
|
170
|
+
dtype=backbone.dtype_policy,
|
171
|
+
name="pooled_dropout",
|
172
|
+
)
|
173
|
+
hidden_dim = hidden_dim or backbone.hidden_dim
|
174
|
+
self.pooled_dense = keras.layers.Dense(
|
175
|
+
hidden_dim,
|
176
|
+
activation="tanh",
|
177
|
+
dtype=backbone.dtype_policy,
|
178
|
+
name="pooled_dense",
|
179
|
+
)
|
180
|
+
self.output_dropout = keras.layers.Dropout(
|
181
|
+
dropout,
|
182
|
+
dtype=backbone.dtype_policy,
|
183
|
+
name="output_dropout",
|
184
|
+
)
|
185
|
+
self.output_dense = keras.layers.Dense(
|
186
|
+
num_classes,
|
187
|
+
kernel_initializer=roberta_kernel_initializer(),
|
188
|
+
activation=activation,
|
189
|
+
dtype=backbone.dtype_policy,
|
190
|
+
name="logits",
|
191
|
+
)
|
192
|
+
|
193
|
+
# === Functional Model ===
|
194
|
+
inputs = backbone.input
|
195
|
+
x = backbone(inputs)[:, backbone.start_token_index, :]
|
196
|
+
x = self.pooled_dropout(x)
|
197
|
+
x = self.pooled_dense(x)
|
198
|
+
x = self.output_dropout(x)
|
199
|
+
outputs = self.output_dense(x)
|
200
|
+
# Instantiate using Functional API Model constructor
|
201
|
+
super().__init__(
|
202
|
+
inputs=inputs,
|
203
|
+
outputs=outputs,
|
204
|
+
**kwargs,
|
205
|
+
)
|
206
|
+
|
207
|
+
# === Config ===
|
208
|
+
self.backbone = backbone
|
209
|
+
self.preprocessor = preprocessor
|
210
|
+
self.num_classes = num_classes
|
211
|
+
self.activation = keras.activations.get(activation)
|
212
|
+
self.hidden_dim = hidden_dim
|
213
|
+
self.dropout = dropout
|
214
|
+
|
215
|
+
def get_config(self):
|
216
|
+
config = super().get_config()
|
217
|
+
config.update(
|
218
|
+
{
|
219
|
+
"num_classes": self.num_classes,
|
220
|
+
"activation": keras.activations.serialize(self.activation),
|
221
|
+
"hidden_dim": self.hidden_dim,
|
222
|
+
"dropout": self.dropout,
|
223
|
+
}
|
224
|
+
)
|
225
|
+
return config
|
@@ -0,0 +1,141 @@
|
|
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.modeling.masked_lm_head import MaskedLMHead
|
20
|
+
from keras_hub.src.models.masked_lm import MaskedLM
|
21
|
+
from keras_hub.src.models.roberta.roberta_backbone import (
|
22
|
+
roberta_kernel_initializer,
|
23
|
+
)
|
24
|
+
from keras_hub.src.models.xlm_roberta.xlm_roberta_backbone import (
|
25
|
+
XLMRobertaBackbone,
|
26
|
+
)
|
27
|
+
from keras_hub.src.models.xlm_roberta.xlm_roberta_masked_lm_preprocessor import (
|
28
|
+
XLMRobertaMaskedLMPreprocessor,
|
29
|
+
)
|
30
|
+
|
31
|
+
|
32
|
+
@keras_hub_export("keras_hub.models.XLMRobertaMaskedLM")
|
33
|
+
class XLMRobertaMaskedLM(MaskedLM):
|
34
|
+
"""An end-to-end XLM-RoBERTa model for the masked language modeling task.
|
35
|
+
|
36
|
+
This model will train XLM-RoBERTa on a masked language modeling task.
|
37
|
+
The model will predict labels for a number of masked tokens in the
|
38
|
+
input data. For usage of this model with pre-trained weights, see the
|
39
|
+
`from_preset()` method.
|
40
|
+
|
41
|
+
This model can optionally be configured with a `preprocessor` layer, in
|
42
|
+
which case inputs can be raw string features during `fit()`, `predict()`,
|
43
|
+
and `evaluate()`. Inputs will be tokenized and dynamically masked during
|
44
|
+
training and evaluation. This is done by default when creating the model
|
45
|
+
with `from_preset()`.
|
46
|
+
|
47
|
+
Disclaimer: Pre-trained models are provided on an "as is" basis, without
|
48
|
+
warranties or conditions of any kind. The underlying model is provided by a
|
49
|
+
third party and subject to a separate license, available
|
50
|
+
[here](https://github.com/facebookresearch/fairseq).
|
51
|
+
|
52
|
+
Args:
|
53
|
+
backbone: A `keras_hub.models.XLMRobertaBackbone` instance.
|
54
|
+
preprocessor: A `keras_hub.models.XLMRobertaMaskedLMPreprocessor` or
|
55
|
+
`None`. If `None`, this model will not apply preprocessing, and
|
56
|
+
inputs should be preprocessed before calling the model.
|
57
|
+
|
58
|
+
Examples:
|
59
|
+
|
60
|
+
Raw string inputs and pretrained backbone.
|
61
|
+
```python
|
62
|
+
# Create a dataset with raw string features. Labels are inferred.
|
63
|
+
features = ["The quick brown fox jumped.", "I forgot my homework."]
|
64
|
+
|
65
|
+
# Pretrained language model
|
66
|
+
# on an MLM task.
|
67
|
+
masked_lm = keras_hub.models.XLMRobertaMaskedLM.from_preset(
|
68
|
+
"xlm_roberta_base_multi",
|
69
|
+
)
|
70
|
+
masked_lm.fit(x=features, batch_size=2)
|
71
|
+
```
|
72
|
+
|
73
|
+
# Re-compile (e.g., with a new learning rate).
|
74
|
+
masked_lm.compile(
|
75
|
+
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
|
76
|
+
optimizer=keras.optimizers.Adam(5e-5),
|
77
|
+
jit_compile=True,
|
78
|
+
)
|
79
|
+
# Access backbone programmatically (e.g., to change `trainable`).
|
80
|
+
masked_lm.backbone.trainable = False
|
81
|
+
# Fit again.
|
82
|
+
masked_lm.fit(x=features, batch_size=2)
|
83
|
+
```
|
84
|
+
|
85
|
+
Preprocessed integer data.
|
86
|
+
```python
|
87
|
+
# Create a preprocessed dataset where 0 is the mask token.
|
88
|
+
features = {
|
89
|
+
"token_ids": np.array([[1, 2, 0, 4, 0, 6, 7, 8]] * 2),
|
90
|
+
"padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1]] * 2),
|
91
|
+
"mask_positions": np.array([[2, 4]] * 2)
|
92
|
+
}
|
93
|
+
# Labels are the original masked values.
|
94
|
+
labels = [[3, 5]] * 2
|
95
|
+
|
96
|
+
masked_lm = keras_hub.models.XLMRobertaMaskedLM.from_preset(
|
97
|
+
"xlm_roberta_base_multi",
|
98
|
+
preprocessor=None,
|
99
|
+
)
|
100
|
+
|
101
|
+
masked_lm.fit(x=features, y=labels, batch_size=2)
|
102
|
+
```
|
103
|
+
"""
|
104
|
+
|
105
|
+
backbone_cls = XLMRobertaBackbone
|
106
|
+
preprocessor_cls = XLMRobertaMaskedLMPreprocessor
|
107
|
+
|
108
|
+
def __init__(
|
109
|
+
self,
|
110
|
+
backbone,
|
111
|
+
preprocessor=None,
|
112
|
+
**kwargs,
|
113
|
+
):
|
114
|
+
# === Layers ===
|
115
|
+
self.backbone = backbone
|
116
|
+
self.preprocessor = preprocessor
|
117
|
+
self.masked_lm_head = MaskedLMHead(
|
118
|
+
vocabulary_size=backbone.vocabulary_size,
|
119
|
+
token_embedding=backbone.token_embedding,
|
120
|
+
intermediate_activation="gelu",
|
121
|
+
kernel_initializer=roberta_kernel_initializer(),
|
122
|
+
dtype=backbone.dtype_policy,
|
123
|
+
name="mlm_head",
|
124
|
+
)
|
125
|
+
|
126
|
+
# === Functional Model ===
|
127
|
+
inputs = {
|
128
|
+
**backbone.input,
|
129
|
+
"mask_positions": keras.Input(
|
130
|
+
shape=(None,), dtype="int32", name="mask_positions"
|
131
|
+
),
|
132
|
+
}
|
133
|
+
backbone_outputs = backbone(backbone.input)
|
134
|
+
outputs = self.masked_lm_head(
|
135
|
+
backbone_outputs, inputs["mask_positions"]
|
136
|
+
)
|
137
|
+
super().__init__(
|
138
|
+
inputs=inputs,
|
139
|
+
outputs=outputs,
|
140
|
+
**kwargs,
|
141
|
+
)
|
@@ -0,0 +1,195 @@
|
|
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 absl import logging
|
17
|
+
|
18
|
+
from keras_hub.src.api_export import keras_hub_export
|
19
|
+
from keras_hub.src.layers.preprocessing.masked_lm_mask_generator import (
|
20
|
+
MaskedLMMaskGenerator,
|
21
|
+
)
|
22
|
+
from keras_hub.src.models.xlm_roberta.xlm_roberta_preprocessor import (
|
23
|
+
XLMRobertaPreprocessor,
|
24
|
+
)
|
25
|
+
|
26
|
+
|
27
|
+
@keras_hub_export("keras_hub.models.XLMRobertaMaskedLMPreprocessor")
|
28
|
+
class XLMRobertaMaskedLMPreprocessor(XLMRobertaPreprocessor):
|
29
|
+
"""XLM-RoBERTa preprocessing for the masked language modeling task.
|
30
|
+
|
31
|
+
This preprocessing layer will prepare inputs for a masked language modeling
|
32
|
+
task. It is primarily intended for use with the
|
33
|
+
`keras_hub.models.XLMRobertaMaskedLM` task model. Preprocessing will occur in
|
34
|
+
multiple steps.
|
35
|
+
|
36
|
+
1. Tokenize any number of input segments using the `tokenizer`.
|
37
|
+
2. Pack the inputs together with the appropriate `"<s>"`, `"</s>"` and
|
38
|
+
`"<pad>"` tokens, i.e., adding a single `"<s>"` at the start of the
|
39
|
+
entire sequence, `"</s></s>"` between each segment,
|
40
|
+
and a `"</s>"` at the end of the entire sequence.
|
41
|
+
3. Randomly select non-special tokens to mask, controlled by
|
42
|
+
`mask_selection_rate`.
|
43
|
+
4. Construct a `(x, y, sample_weight)` tuple suitable for training with a
|
44
|
+
`keras_hub.models.XLMRobertaMaskedLM` task model.
|
45
|
+
|
46
|
+
Args:
|
47
|
+
tokenizer: A `keras_hub.models.XLMRobertaTokenizer` instance.
|
48
|
+
sequence_length: int. The length of the packed inputs.
|
49
|
+
truncate: string. The algorithm to truncate a list of batched segments
|
50
|
+
to fit within `sequence_length`. The value can be either
|
51
|
+
`round_robin` or `waterfall`:
|
52
|
+
- `"round_robin"`: Available space is assigned one token at a
|
53
|
+
time in a round-robin fashion to the inputs that still need
|
54
|
+
some, until the limit is reached.
|
55
|
+
- `"waterfall"`: The allocation of the budget is done using a
|
56
|
+
"waterfall" algorithm that allocates quota in a
|
57
|
+
left-to-right manner and fills up the buckets until we run
|
58
|
+
out of budget. It supports an arbitrary number of segments.
|
59
|
+
mask_selection_rate: float. The probability an input token will be
|
60
|
+
dynamically masked.
|
61
|
+
mask_selection_length: int. The maximum number of masked tokens
|
62
|
+
in a given sample.
|
63
|
+
mask_token_rate: float. The probability the a selected token will be
|
64
|
+
replaced with the mask token.
|
65
|
+
random_token_rate: float. The probability the a selected token will be
|
66
|
+
replaced with a random token from the vocabulary. A selected token
|
67
|
+
will be left as is with probability
|
68
|
+
`1 - mask_token_rate - random_token_rate`.
|
69
|
+
|
70
|
+
Call arguments:
|
71
|
+
x: A tensor of single string sequences, or a tuple of multiple
|
72
|
+
tensor sequences to be packed together. Inputs may be batched or
|
73
|
+
unbatched. For single sequences, raw python inputs will be converted
|
74
|
+
to tensors. For multiple sequences, pass tensors directly.
|
75
|
+
y: Label data. Should always be `None` as the layer generates labels.
|
76
|
+
sample_weight: Label weights. Should always be `None` as the layer
|
77
|
+
generates label weights.
|
78
|
+
|
79
|
+
Examples:
|
80
|
+
|
81
|
+
Directly calling the layer on data.
|
82
|
+
```python
|
83
|
+
# Load the preprocessor from a preset.
|
84
|
+
preprocessor = keras_hub.models.XLMRobertaMaskedLMPreprocessor.from_preset(
|
85
|
+
"xlm_roberta_base_multi"
|
86
|
+
)
|
87
|
+
|
88
|
+
# Tokenize and mask a single sentence.
|
89
|
+
preprocessor("The quick brown fox jumped.")
|
90
|
+
# Tokenize and mask a batch of single sentences.
|
91
|
+
preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
|
92
|
+
# Tokenize and mask sentence pairs.
|
93
|
+
# In this case, always convert input to tensors before calling the layer.
|
94
|
+
first = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
|
95
|
+
second = tf.constant(["The fox tripped.", "Oh look, a whale."])
|
96
|
+
preprocessor((first, second))
|
97
|
+
```
|
98
|
+
|
99
|
+
Mapping with `tf.data.Dataset`.
|
100
|
+
```python
|
101
|
+
preprocessor = keras_hub.models.XLMRobertaMaskedLMPreprocessor.from_preset(
|
102
|
+
"xlm_roberta_base_multi"
|
103
|
+
)
|
104
|
+
first = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
|
105
|
+
second = tf.constant(["The fox tripped.", "Oh look, a whale."])
|
106
|
+
|
107
|
+
# Map single sentences.
|
108
|
+
ds = tf.data.Dataset.from_tensor_slices(first)
|
109
|
+
ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
|
110
|
+
|
111
|
+
# Map sentence pairs.
|
112
|
+
ds = tf.data.Dataset.from_tensor_slices((first, second))
|
113
|
+
# Watch out for tf.data's default unpacking of tuples here!
|
114
|
+
# Best to invoke the `preprocessor` directly in this case.
|
115
|
+
ds = ds.map(
|
116
|
+
lambda first, second: preprocessor(x=(first, second)),
|
117
|
+
num_parallel_calls=tf.data.AUTOTUNE,
|
118
|
+
)
|
119
|
+
```
|
120
|
+
```
|
121
|
+
"""
|
122
|
+
|
123
|
+
def __init__(
|
124
|
+
self,
|
125
|
+
tokenizer,
|
126
|
+
sequence_length=512,
|
127
|
+
truncate="round_robin",
|
128
|
+
mask_selection_rate=0.15,
|
129
|
+
mask_selection_length=96,
|
130
|
+
mask_token_rate=0.8,
|
131
|
+
random_token_rate=0.1,
|
132
|
+
**kwargs,
|
133
|
+
):
|
134
|
+
super().__init__(
|
135
|
+
tokenizer,
|
136
|
+
sequence_length=sequence_length,
|
137
|
+
truncate=truncate,
|
138
|
+
**kwargs,
|
139
|
+
)
|
140
|
+
self.mask_selection_rate = mask_selection_rate
|
141
|
+
self.mask_selection_length = mask_selection_length
|
142
|
+
self.mask_token_rate = mask_token_rate
|
143
|
+
self.random_token_rate = random_token_rate
|
144
|
+
self.masker = None
|
145
|
+
|
146
|
+
def build(self, input_shape):
|
147
|
+
super().build(input_shape)
|
148
|
+
# Defer masker creation to `build()` so that we can be sure tokenizer
|
149
|
+
# assets have loaded when restoring a saved model.
|
150
|
+
self.masker = MaskedLMMaskGenerator(
|
151
|
+
mask_selection_rate=self.mask_selection_rate,
|
152
|
+
mask_selection_length=self.mask_selection_length,
|
153
|
+
mask_token_rate=self.mask_token_rate,
|
154
|
+
random_token_rate=self.random_token_rate,
|
155
|
+
vocabulary_size=self.tokenizer.vocabulary_size(),
|
156
|
+
mask_token_id=self.tokenizer.mask_token_id,
|
157
|
+
unselectable_token_ids=[
|
158
|
+
self.tokenizer.start_token_id,
|
159
|
+
self.tokenizer.end_token_id,
|
160
|
+
self.tokenizer.pad_token_id,
|
161
|
+
],
|
162
|
+
)
|
163
|
+
|
164
|
+
def get_config(self):
|
165
|
+
config = super().get_config()
|
166
|
+
config.update(
|
167
|
+
{
|
168
|
+
"mask_selection_rate": self.mask_selection_rate,
|
169
|
+
"mask_selection_length": self.mask_selection_length,
|
170
|
+
"mask_token_rate": self.mask_token_rate,
|
171
|
+
"random_token_rate": self.random_token_rate,
|
172
|
+
}
|
173
|
+
)
|
174
|
+
return config
|
175
|
+
|
176
|
+
def call(self, x, y=None, sample_weight=None):
|
177
|
+
if y is not None or sample_weight is not None:
|
178
|
+
logging.warning(
|
179
|
+
f"{self.__class__.__name__} generates `y` and `sample_weight` "
|
180
|
+
"based on your input data, but your data already contains `y` "
|
181
|
+
"or `sample_weight`. Your `y` and `sample_weight` will be "
|
182
|
+
"ignored."
|
183
|
+
)
|
184
|
+
|
185
|
+
x = super().call(x)
|
186
|
+
token_ids, padding_mask = x["token_ids"], x["padding_mask"]
|
187
|
+
masker_outputs = self.masker(token_ids)
|
188
|
+
x = {
|
189
|
+
"token_ids": masker_outputs["token_ids"],
|
190
|
+
"padding_mask": padding_mask,
|
191
|
+
"mask_positions": masker_outputs["mask_positions"],
|
192
|
+
}
|
193
|
+
y = masker_outputs["mask_ids"]
|
194
|
+
sample_weight = masker_outputs["mask_weights"]
|
195
|
+
return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
|