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.
Files changed (297) hide show
  1. keras_hub/__init__.py +52 -0
  2. keras_hub/api/__init__.py +27 -0
  3. keras_hub/api/layers/__init__.py +47 -0
  4. keras_hub/api/metrics/__init__.py +24 -0
  5. keras_hub/api/models/__init__.py +249 -0
  6. keras_hub/api/samplers/__init__.py +29 -0
  7. keras_hub/api/tokenizers/__init__.py +35 -0
  8. keras_hub/src/__init__.py +13 -0
  9. keras_hub/src/api_export.py +53 -0
  10. keras_hub/src/layers/__init__.py +13 -0
  11. keras_hub/src/layers/modeling/__init__.py +13 -0
  12. keras_hub/src/layers/modeling/alibi_bias.py +143 -0
  13. keras_hub/src/layers/modeling/cached_multi_head_attention.py +137 -0
  14. keras_hub/src/layers/modeling/f_net_encoder.py +200 -0
  15. keras_hub/src/layers/modeling/masked_lm_head.py +239 -0
  16. keras_hub/src/layers/modeling/position_embedding.py +123 -0
  17. keras_hub/src/layers/modeling/reversible_embedding.py +311 -0
  18. keras_hub/src/layers/modeling/rotary_embedding.py +169 -0
  19. keras_hub/src/layers/modeling/sine_position_encoding.py +108 -0
  20. keras_hub/src/layers/modeling/token_and_position_embedding.py +150 -0
  21. keras_hub/src/layers/modeling/transformer_decoder.py +496 -0
  22. keras_hub/src/layers/modeling/transformer_encoder.py +262 -0
  23. keras_hub/src/layers/modeling/transformer_layer_utils.py +106 -0
  24. keras_hub/src/layers/preprocessing/__init__.py +13 -0
  25. keras_hub/src/layers/preprocessing/masked_lm_mask_generator.py +220 -0
  26. keras_hub/src/layers/preprocessing/multi_segment_packer.py +319 -0
  27. keras_hub/src/layers/preprocessing/preprocessing_layer.py +62 -0
  28. keras_hub/src/layers/preprocessing/random_deletion.py +271 -0
  29. keras_hub/src/layers/preprocessing/random_swap.py +267 -0
  30. keras_hub/src/layers/preprocessing/start_end_packer.py +219 -0
  31. keras_hub/src/metrics/__init__.py +13 -0
  32. keras_hub/src/metrics/bleu.py +394 -0
  33. keras_hub/src/metrics/edit_distance.py +197 -0
  34. keras_hub/src/metrics/perplexity.py +181 -0
  35. keras_hub/src/metrics/rouge_base.py +204 -0
  36. keras_hub/src/metrics/rouge_l.py +97 -0
  37. keras_hub/src/metrics/rouge_n.py +125 -0
  38. keras_hub/src/models/__init__.py +13 -0
  39. keras_hub/src/models/albert/__init__.py +20 -0
  40. keras_hub/src/models/albert/albert_backbone.py +267 -0
  41. keras_hub/src/models/albert/albert_classifier.py +202 -0
  42. keras_hub/src/models/albert/albert_masked_lm.py +129 -0
  43. keras_hub/src/models/albert/albert_masked_lm_preprocessor.py +194 -0
  44. keras_hub/src/models/albert/albert_preprocessor.py +206 -0
  45. keras_hub/src/models/albert/albert_presets.py +70 -0
  46. keras_hub/src/models/albert/albert_tokenizer.py +119 -0
  47. keras_hub/src/models/backbone.py +311 -0
  48. keras_hub/src/models/bart/__init__.py +20 -0
  49. keras_hub/src/models/bart/bart_backbone.py +261 -0
  50. keras_hub/src/models/bart/bart_preprocessor.py +276 -0
  51. keras_hub/src/models/bart/bart_presets.py +74 -0
  52. keras_hub/src/models/bart/bart_seq_2_seq_lm.py +490 -0
  53. keras_hub/src/models/bart/bart_seq_2_seq_lm_preprocessor.py +262 -0
  54. keras_hub/src/models/bart/bart_tokenizer.py +124 -0
  55. keras_hub/src/models/bert/__init__.py +23 -0
  56. keras_hub/src/models/bert/bert_backbone.py +227 -0
  57. keras_hub/src/models/bert/bert_classifier.py +183 -0
  58. keras_hub/src/models/bert/bert_masked_lm.py +131 -0
  59. keras_hub/src/models/bert/bert_masked_lm_preprocessor.py +198 -0
  60. keras_hub/src/models/bert/bert_preprocessor.py +184 -0
  61. keras_hub/src/models/bert/bert_presets.py +147 -0
  62. keras_hub/src/models/bert/bert_tokenizer.py +112 -0
  63. keras_hub/src/models/bloom/__init__.py +20 -0
  64. keras_hub/src/models/bloom/bloom_attention.py +186 -0
  65. keras_hub/src/models/bloom/bloom_backbone.py +173 -0
  66. keras_hub/src/models/bloom/bloom_causal_lm.py +298 -0
  67. keras_hub/src/models/bloom/bloom_causal_lm_preprocessor.py +176 -0
  68. keras_hub/src/models/bloom/bloom_decoder.py +206 -0
  69. keras_hub/src/models/bloom/bloom_preprocessor.py +185 -0
  70. keras_hub/src/models/bloom/bloom_presets.py +121 -0
  71. keras_hub/src/models/bloom/bloom_tokenizer.py +116 -0
  72. keras_hub/src/models/causal_lm.py +383 -0
  73. keras_hub/src/models/classifier.py +109 -0
  74. keras_hub/src/models/csp_darknet/__init__.py +13 -0
  75. keras_hub/src/models/csp_darknet/csp_darknet_backbone.py +410 -0
  76. keras_hub/src/models/csp_darknet/csp_darknet_image_classifier.py +133 -0
  77. keras_hub/src/models/deberta_v3/__init__.py +24 -0
  78. keras_hub/src/models/deberta_v3/deberta_v3_backbone.py +210 -0
  79. keras_hub/src/models/deberta_v3/deberta_v3_classifier.py +228 -0
  80. keras_hub/src/models/deberta_v3/deberta_v3_masked_lm.py +135 -0
  81. keras_hub/src/models/deberta_v3/deberta_v3_masked_lm_preprocessor.py +191 -0
  82. keras_hub/src/models/deberta_v3/deberta_v3_preprocessor.py +206 -0
  83. keras_hub/src/models/deberta_v3/deberta_v3_presets.py +82 -0
  84. keras_hub/src/models/deberta_v3/deberta_v3_tokenizer.py +155 -0
  85. keras_hub/src/models/deberta_v3/disentangled_attention_encoder.py +227 -0
  86. keras_hub/src/models/deberta_v3/disentangled_self_attention.py +412 -0
  87. keras_hub/src/models/deberta_v3/relative_embedding.py +94 -0
  88. keras_hub/src/models/densenet/__init__.py +13 -0
  89. keras_hub/src/models/densenet/densenet_backbone.py +210 -0
  90. keras_hub/src/models/densenet/densenet_image_classifier.py +131 -0
  91. keras_hub/src/models/distil_bert/__init__.py +26 -0
  92. keras_hub/src/models/distil_bert/distil_bert_backbone.py +187 -0
  93. keras_hub/src/models/distil_bert/distil_bert_classifier.py +208 -0
  94. keras_hub/src/models/distil_bert/distil_bert_masked_lm.py +137 -0
  95. keras_hub/src/models/distil_bert/distil_bert_masked_lm_preprocessor.py +194 -0
  96. keras_hub/src/models/distil_bert/distil_bert_preprocessor.py +175 -0
  97. keras_hub/src/models/distil_bert/distil_bert_presets.py +57 -0
  98. keras_hub/src/models/distil_bert/distil_bert_tokenizer.py +114 -0
  99. keras_hub/src/models/electra/__init__.py +20 -0
  100. keras_hub/src/models/electra/electra_backbone.py +247 -0
  101. keras_hub/src/models/electra/electra_preprocessor.py +154 -0
  102. keras_hub/src/models/electra/electra_presets.py +95 -0
  103. keras_hub/src/models/electra/electra_tokenizer.py +104 -0
  104. keras_hub/src/models/f_net/__init__.py +20 -0
  105. keras_hub/src/models/f_net/f_net_backbone.py +236 -0
  106. keras_hub/src/models/f_net/f_net_classifier.py +154 -0
  107. keras_hub/src/models/f_net/f_net_masked_lm.py +132 -0
  108. keras_hub/src/models/f_net/f_net_masked_lm_preprocessor.py +196 -0
  109. keras_hub/src/models/f_net/f_net_preprocessor.py +177 -0
  110. keras_hub/src/models/f_net/f_net_presets.py +43 -0
  111. keras_hub/src/models/f_net/f_net_tokenizer.py +95 -0
  112. keras_hub/src/models/falcon/__init__.py +20 -0
  113. keras_hub/src/models/falcon/falcon_attention.py +156 -0
  114. keras_hub/src/models/falcon/falcon_backbone.py +164 -0
  115. keras_hub/src/models/falcon/falcon_causal_lm.py +291 -0
  116. keras_hub/src/models/falcon/falcon_causal_lm_preprocessor.py +173 -0
  117. keras_hub/src/models/falcon/falcon_preprocessor.py +187 -0
  118. keras_hub/src/models/falcon/falcon_presets.py +30 -0
  119. keras_hub/src/models/falcon/falcon_tokenizer.py +110 -0
  120. keras_hub/src/models/falcon/falcon_transformer_decoder.py +255 -0
  121. keras_hub/src/models/feature_pyramid_backbone.py +73 -0
  122. keras_hub/src/models/gemma/__init__.py +20 -0
  123. keras_hub/src/models/gemma/gemma_attention.py +250 -0
  124. keras_hub/src/models/gemma/gemma_backbone.py +316 -0
  125. keras_hub/src/models/gemma/gemma_causal_lm.py +448 -0
  126. keras_hub/src/models/gemma/gemma_causal_lm_preprocessor.py +167 -0
  127. keras_hub/src/models/gemma/gemma_decoder_block.py +241 -0
  128. keras_hub/src/models/gemma/gemma_preprocessor.py +191 -0
  129. keras_hub/src/models/gemma/gemma_presets.py +248 -0
  130. keras_hub/src/models/gemma/gemma_tokenizer.py +103 -0
  131. keras_hub/src/models/gemma/rms_normalization.py +40 -0
  132. keras_hub/src/models/gpt2/__init__.py +20 -0
  133. keras_hub/src/models/gpt2/gpt2_backbone.py +199 -0
  134. keras_hub/src/models/gpt2/gpt2_causal_lm.py +437 -0
  135. keras_hub/src/models/gpt2/gpt2_causal_lm_preprocessor.py +173 -0
  136. keras_hub/src/models/gpt2/gpt2_preprocessor.py +187 -0
  137. keras_hub/src/models/gpt2/gpt2_presets.py +82 -0
  138. keras_hub/src/models/gpt2/gpt2_tokenizer.py +110 -0
  139. keras_hub/src/models/gpt_neo_x/__init__.py +13 -0
  140. keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py +251 -0
  141. keras_hub/src/models/gpt_neo_x/gpt_neo_x_backbone.py +175 -0
  142. keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm.py +201 -0
  143. keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm_preprocessor.py +141 -0
  144. keras_hub/src/models/gpt_neo_x/gpt_neo_x_decoder.py +258 -0
  145. keras_hub/src/models/gpt_neo_x/gpt_neo_x_preprocessor.py +145 -0
  146. keras_hub/src/models/gpt_neo_x/gpt_neo_x_tokenizer.py +88 -0
  147. keras_hub/src/models/image_classifier.py +90 -0
  148. keras_hub/src/models/llama/__init__.py +20 -0
  149. keras_hub/src/models/llama/llama_attention.py +225 -0
  150. keras_hub/src/models/llama/llama_backbone.py +188 -0
  151. keras_hub/src/models/llama/llama_causal_lm.py +327 -0
  152. keras_hub/src/models/llama/llama_causal_lm_preprocessor.py +170 -0
  153. keras_hub/src/models/llama/llama_decoder.py +246 -0
  154. keras_hub/src/models/llama/llama_layernorm.py +48 -0
  155. keras_hub/src/models/llama/llama_preprocessor.py +189 -0
  156. keras_hub/src/models/llama/llama_presets.py +80 -0
  157. keras_hub/src/models/llama/llama_tokenizer.py +84 -0
  158. keras_hub/src/models/llama3/__init__.py +20 -0
  159. keras_hub/src/models/llama3/llama3_backbone.py +84 -0
  160. keras_hub/src/models/llama3/llama3_causal_lm.py +46 -0
  161. keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.py +173 -0
  162. keras_hub/src/models/llama3/llama3_preprocessor.py +21 -0
  163. keras_hub/src/models/llama3/llama3_presets.py +69 -0
  164. keras_hub/src/models/llama3/llama3_tokenizer.py +63 -0
  165. keras_hub/src/models/masked_lm.py +101 -0
  166. keras_hub/src/models/mistral/__init__.py +20 -0
  167. keras_hub/src/models/mistral/mistral_attention.py +238 -0
  168. keras_hub/src/models/mistral/mistral_backbone.py +203 -0
  169. keras_hub/src/models/mistral/mistral_causal_lm.py +328 -0
  170. keras_hub/src/models/mistral/mistral_causal_lm_preprocessor.py +175 -0
  171. keras_hub/src/models/mistral/mistral_layer_norm.py +48 -0
  172. keras_hub/src/models/mistral/mistral_preprocessor.py +190 -0
  173. keras_hub/src/models/mistral/mistral_presets.py +48 -0
  174. keras_hub/src/models/mistral/mistral_tokenizer.py +82 -0
  175. keras_hub/src/models/mistral/mistral_transformer_decoder.py +265 -0
  176. keras_hub/src/models/mix_transformer/__init__.py +13 -0
  177. keras_hub/src/models/mix_transformer/mix_transformer_backbone.py +181 -0
  178. keras_hub/src/models/mix_transformer/mix_transformer_classifier.py +133 -0
  179. keras_hub/src/models/mix_transformer/mix_transformer_layers.py +300 -0
  180. keras_hub/src/models/opt/__init__.py +20 -0
  181. keras_hub/src/models/opt/opt_backbone.py +173 -0
  182. keras_hub/src/models/opt/opt_causal_lm.py +301 -0
  183. keras_hub/src/models/opt/opt_causal_lm_preprocessor.py +177 -0
  184. keras_hub/src/models/opt/opt_preprocessor.py +188 -0
  185. keras_hub/src/models/opt/opt_presets.py +72 -0
  186. keras_hub/src/models/opt/opt_tokenizer.py +116 -0
  187. keras_hub/src/models/pali_gemma/__init__.py +23 -0
  188. keras_hub/src/models/pali_gemma/pali_gemma_backbone.py +277 -0
  189. keras_hub/src/models/pali_gemma/pali_gemma_causal_lm.py +313 -0
  190. keras_hub/src/models/pali_gemma/pali_gemma_causal_lm_preprocessor.py +147 -0
  191. keras_hub/src/models/pali_gemma/pali_gemma_decoder_block.py +160 -0
  192. keras_hub/src/models/pali_gemma/pali_gemma_presets.py +78 -0
  193. keras_hub/src/models/pali_gemma/pali_gemma_tokenizer.py +79 -0
  194. keras_hub/src/models/pali_gemma/pali_gemma_vit.py +566 -0
  195. keras_hub/src/models/phi3/__init__.py +20 -0
  196. keras_hub/src/models/phi3/phi3_attention.py +260 -0
  197. keras_hub/src/models/phi3/phi3_backbone.py +224 -0
  198. keras_hub/src/models/phi3/phi3_causal_lm.py +218 -0
  199. keras_hub/src/models/phi3/phi3_causal_lm_preprocessor.py +173 -0
  200. keras_hub/src/models/phi3/phi3_decoder.py +260 -0
  201. keras_hub/src/models/phi3/phi3_layernorm.py +48 -0
  202. keras_hub/src/models/phi3/phi3_preprocessor.py +190 -0
  203. keras_hub/src/models/phi3/phi3_presets.py +50 -0
  204. keras_hub/src/models/phi3/phi3_rotary_embedding.py +137 -0
  205. keras_hub/src/models/phi3/phi3_tokenizer.py +94 -0
  206. keras_hub/src/models/preprocessor.py +207 -0
  207. keras_hub/src/models/resnet/__init__.py +13 -0
  208. keras_hub/src/models/resnet/resnet_backbone.py +612 -0
  209. keras_hub/src/models/resnet/resnet_image_classifier.py +136 -0
  210. keras_hub/src/models/roberta/__init__.py +20 -0
  211. keras_hub/src/models/roberta/roberta_backbone.py +184 -0
  212. keras_hub/src/models/roberta/roberta_classifier.py +209 -0
  213. keras_hub/src/models/roberta/roberta_masked_lm.py +136 -0
  214. keras_hub/src/models/roberta/roberta_masked_lm_preprocessor.py +198 -0
  215. keras_hub/src/models/roberta/roberta_preprocessor.py +192 -0
  216. keras_hub/src/models/roberta/roberta_presets.py +43 -0
  217. keras_hub/src/models/roberta/roberta_tokenizer.py +132 -0
  218. keras_hub/src/models/seq_2_seq_lm.py +54 -0
  219. keras_hub/src/models/t5/__init__.py +20 -0
  220. keras_hub/src/models/t5/t5_backbone.py +261 -0
  221. keras_hub/src/models/t5/t5_layer_norm.py +35 -0
  222. keras_hub/src/models/t5/t5_multi_head_attention.py +324 -0
  223. keras_hub/src/models/t5/t5_presets.py +95 -0
  224. keras_hub/src/models/t5/t5_tokenizer.py +100 -0
  225. keras_hub/src/models/t5/t5_transformer_layer.py +178 -0
  226. keras_hub/src/models/task.py +419 -0
  227. keras_hub/src/models/vgg/__init__.py +13 -0
  228. keras_hub/src/models/vgg/vgg_backbone.py +158 -0
  229. keras_hub/src/models/vgg/vgg_image_classifier.py +124 -0
  230. keras_hub/src/models/vit_det/__init__.py +13 -0
  231. keras_hub/src/models/vit_det/vit_det_backbone.py +204 -0
  232. keras_hub/src/models/vit_det/vit_layers.py +565 -0
  233. keras_hub/src/models/whisper/__init__.py +20 -0
  234. keras_hub/src/models/whisper/whisper_audio_feature_extractor.py +260 -0
  235. keras_hub/src/models/whisper/whisper_backbone.py +305 -0
  236. keras_hub/src/models/whisper/whisper_cached_multi_head_attention.py +153 -0
  237. keras_hub/src/models/whisper/whisper_decoder.py +141 -0
  238. keras_hub/src/models/whisper/whisper_encoder.py +106 -0
  239. keras_hub/src/models/whisper/whisper_preprocessor.py +326 -0
  240. keras_hub/src/models/whisper/whisper_presets.py +148 -0
  241. keras_hub/src/models/whisper/whisper_tokenizer.py +163 -0
  242. keras_hub/src/models/xlm_roberta/__init__.py +26 -0
  243. keras_hub/src/models/xlm_roberta/xlm_roberta_backbone.py +81 -0
  244. keras_hub/src/models/xlm_roberta/xlm_roberta_classifier.py +225 -0
  245. keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm.py +141 -0
  246. keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm_preprocessor.py +195 -0
  247. keras_hub/src/models/xlm_roberta/xlm_roberta_preprocessor.py +205 -0
  248. keras_hub/src/models/xlm_roberta/xlm_roberta_presets.py +43 -0
  249. keras_hub/src/models/xlm_roberta/xlm_roberta_tokenizer.py +191 -0
  250. keras_hub/src/models/xlnet/__init__.py +13 -0
  251. keras_hub/src/models/xlnet/relative_attention.py +459 -0
  252. keras_hub/src/models/xlnet/xlnet_backbone.py +222 -0
  253. keras_hub/src/models/xlnet/xlnet_content_and_query_embedding.py +133 -0
  254. keras_hub/src/models/xlnet/xlnet_encoder.py +378 -0
  255. keras_hub/src/samplers/__init__.py +13 -0
  256. keras_hub/src/samplers/beam_sampler.py +207 -0
  257. keras_hub/src/samplers/contrastive_sampler.py +231 -0
  258. keras_hub/src/samplers/greedy_sampler.py +50 -0
  259. keras_hub/src/samplers/random_sampler.py +77 -0
  260. keras_hub/src/samplers/sampler.py +237 -0
  261. keras_hub/src/samplers/serialization.py +97 -0
  262. keras_hub/src/samplers/top_k_sampler.py +92 -0
  263. keras_hub/src/samplers/top_p_sampler.py +113 -0
  264. keras_hub/src/tests/__init__.py +13 -0
  265. keras_hub/src/tests/test_case.py +608 -0
  266. keras_hub/src/tokenizers/__init__.py +13 -0
  267. keras_hub/src/tokenizers/byte_pair_tokenizer.py +638 -0
  268. keras_hub/src/tokenizers/byte_tokenizer.py +299 -0
  269. keras_hub/src/tokenizers/sentence_piece_tokenizer.py +267 -0
  270. keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py +150 -0
  271. keras_hub/src/tokenizers/tokenizer.py +235 -0
  272. keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py +355 -0
  273. keras_hub/src/tokenizers/word_piece_tokenizer.py +544 -0
  274. keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py +176 -0
  275. keras_hub/src/utils/__init__.py +13 -0
  276. keras_hub/src/utils/keras_utils.py +130 -0
  277. keras_hub/src/utils/pipeline_model.py +293 -0
  278. keras_hub/src/utils/preset_utils.py +621 -0
  279. keras_hub/src/utils/python_utils.py +21 -0
  280. keras_hub/src/utils/tensor_utils.py +206 -0
  281. keras_hub/src/utils/timm/__init__.py +13 -0
  282. keras_hub/src/utils/timm/convert.py +37 -0
  283. keras_hub/src/utils/timm/convert_resnet.py +171 -0
  284. keras_hub/src/utils/transformers/__init__.py +13 -0
  285. keras_hub/src/utils/transformers/convert.py +101 -0
  286. keras_hub/src/utils/transformers/convert_bert.py +173 -0
  287. keras_hub/src/utils/transformers/convert_distilbert.py +184 -0
  288. keras_hub/src/utils/transformers/convert_gemma.py +187 -0
  289. keras_hub/src/utils/transformers/convert_gpt2.py +186 -0
  290. keras_hub/src/utils/transformers/convert_llama3.py +136 -0
  291. keras_hub/src/utils/transformers/convert_pali_gemma.py +303 -0
  292. keras_hub/src/utils/transformers/safetensor_utils.py +97 -0
  293. keras_hub/src/version_utils.py +23 -0
  294. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/METADATA +34 -0
  295. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/RECORD +297 -0
  296. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/WHEEL +5 -0
  297. keras_hub_nightly-0.15.0.dev20240823171555.dist-info/top_level.txt +1 -0
@@ -0,0 +1,448 @@
1
+ # Copyright 2024 The KerasHub Authors
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # https://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+
16
+ import keras
17
+ from keras import ops
18
+
19
+ from keras_hub.src.api_export import keras_hub_export
20
+ from keras_hub.src.models.causal_lm import CausalLM
21
+ from keras_hub.src.models.gemma.gemma_backbone import GemmaBackbone
22
+ from keras_hub.src.models.gemma.gemma_causal_lm_preprocessor import (
23
+ GemmaCausalLMPreprocessor,
24
+ )
25
+ from keras_hub.src.utils.tensor_utils import any_equal
26
+
27
+
28
+ @keras_hub_export("keras_hub.models.GemmaCausalLM")
29
+ class GemmaCausalLM(CausalLM):
30
+ """An end-to-end Gemma model for causal language modeling.
31
+
32
+ A causal language model (LM) predicts the next token based on previous
33
+ tokens. This task setup can be used to train the model unsupervised on
34
+ plain text input, or to autoregressively generate plain text similar to
35
+ the data used for training. This task can be used for pre-training or
36
+ fine-tuning a Gemma model, simply by calling `fit()`.
37
+
38
+ This model has a `generate()` method, which generates text based on a
39
+ prompt. The generation strategy used is controlled by an additional
40
+ `sampler` argument on `compile()`. You can recompile the model with
41
+ different `keras_hub.samplers` objects to control the generation. By
42
+ default, `"greedy"` sampling will be used.
43
+
44
+ This model can optionally be configured with a `preprocessor` layer, in
45
+ which case it will automatically apply preprocessing to string inputs during
46
+ `fit()`, `predict()`, `evaluate()` and `generate()`. This is done by default
47
+ when creating the model with `from_preset()`.
48
+
49
+ Args:
50
+ backbone: A `keras_hub.models.GemmaBackbone` instance.
51
+ preprocessor: A `keras_hub.models.GemmaCausalLMPreprocessor` or `None`.
52
+ If `None`, this model will not apply preprocessing, and inputs
53
+ should be preprocessed before calling the model.
54
+
55
+ Examples:
56
+
57
+ Use `generate()` to do text generation.
58
+ ```python
59
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset("gemma_2b_en")
60
+ gemma_lm.generate("I want to say", max_length=30)
61
+
62
+ # Generate with batched prompts.
63
+ gemma_lm.generate(["This is a", "Where are you"], max_length=30)
64
+ ```
65
+
66
+ Compile the `generate()` function with a custom sampler.
67
+ ```python
68
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset("gemma_2b_en")
69
+ gemma_lm.compile(sampler="top_k")
70
+ gemma_lm.generate("I want to say", max_length=30)
71
+
72
+ gemma_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
73
+ gemma_lm.generate("I want to say", max_length=30)
74
+ ```
75
+
76
+ Use `generate()` without preprocessing.
77
+ ```python
78
+ prompt = {
79
+ # Token ids for "<bos> Keras is".
80
+ "token_ids": np.array([[2, 214064, 603, 0, 0, 0, 0]] * 2),
81
+ # Use `"padding_mask"` to indicate values that should not be overridden.
82
+ "padding_mask": np.array([[1, 1, 1, 0, 0, 0, 0]] * 2),
83
+ }
84
+
85
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset(
86
+ "gemma_2b_en",
87
+ preprocessor=None,
88
+ )
89
+ gemma_lm.generate(prompt)
90
+ ```
91
+
92
+ Call `fit()` on a single batch.
93
+ ```python
94
+ features = ["The quick brown fox jumped.", "I forgot my homework."]
95
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset("gemma_2b_en")
96
+ gemma_lm.fit(x=features, batch_size=2)
97
+ ```
98
+
99
+ Call `fit()` with LoRA fine-tuning enabled.
100
+ ```python
101
+ features = ["The quick brown fox jumped.", "I forgot my homework."]
102
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset("gemma_2b_en")
103
+ gemma.backbone.enable_lora(rank=4)
104
+ gemma_lm.fit(x=features, batch_size=2)
105
+ ```
106
+
107
+ Call `fit()` without preprocessing.
108
+ ```python
109
+ x = {
110
+ # Token ids for "<bos> Keras is deep learning library<eos>"
111
+ "token_ids": np.array([[2, 214064, 603, 5271, 6044, 9581, 1, 0]] * 2),
112
+ "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 0]] * 2),
113
+ }
114
+ y = np.array([[214064, 603, 5271, 6044, 9581, 3, 0, 0]] * 2)
115
+ sw = np.array([[1, 1, 1, 1, 1, 1, 0, 0]] * 2)
116
+
117
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset(
118
+ "gemma_2b_en",
119
+ preprocessor=None,
120
+ )
121
+ gemma_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)
122
+ ```
123
+
124
+ Custom backbone and vocabulary.
125
+ ```python
126
+ tokenizer = keras_hub.models.GemmaTokenizer(
127
+ proto="proto.spm",
128
+ )
129
+ preprocessor = keras_hub.models.GemmaCausalLMPreprocessor(
130
+ tokenizer=tokenizer,
131
+ sequence_length=128,
132
+ )
133
+ backbone = keras_hub.models.GemmaBackbone(
134
+ vocabulary_size=30552,
135
+ num_layers=4,
136
+ num_heads=4,
137
+ hidden_dim=256,
138
+ intermediate_dim=512,
139
+ max_sequence_length=128,
140
+ )
141
+ gemma_lm = keras_hub.models.GemmaCausalLM(
142
+ backbone=backbone,
143
+ preprocessor=preprocessor,
144
+ )
145
+ gemma_lm.fit(x=features, batch_size=2)
146
+ ```
147
+ """
148
+
149
+ backbone_cls = GemmaBackbone
150
+ preprocessor_cls = GemmaCausalLMPreprocessor
151
+
152
+ def __init__(
153
+ self,
154
+ backbone,
155
+ preprocessor=None,
156
+ **kwargs,
157
+ ):
158
+ # === Layers ===
159
+ self.backbone = backbone
160
+ self.preprocessor = preprocessor
161
+
162
+ # === Functional Model ===
163
+ inputs = backbone.input
164
+ hidden_states = backbone(inputs)
165
+ outputs = backbone.token_embedding(hidden_states, reverse=True)
166
+ super().__init__(
167
+ inputs=inputs,
168
+ outputs=outputs,
169
+ **kwargs,
170
+ )
171
+
172
+ def compile(
173
+ self,
174
+ optimizer="auto",
175
+ loss="auto",
176
+ *,
177
+ weighted_metrics="auto",
178
+ sampler="greedy",
179
+ **kwargs,
180
+ ):
181
+ super().compile(
182
+ optimizer=optimizer,
183
+ loss=loss,
184
+ weighted_metrics=weighted_metrics,
185
+ sampler=sampler,
186
+ **kwargs,
187
+ )
188
+
189
+ def call_with_cache(
190
+ self,
191
+ token_ids,
192
+ cache,
193
+ cache_update_index,
194
+ ):
195
+ """Forward pass of `GemmaCausalLM` with cache.
196
+
197
+ `call_with_cache` adds an additional forward pass for the model for
198
+ autoregressive inference. Unlike calling the model directly, this method
199
+ allows caching previous key/value Tensors in multi-head attention layer,
200
+ and avoids recomputing the outputs of seen tokens.
201
+
202
+ Args:
203
+ token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
204
+ cache: a dense float Tensor, the cache of key and value.
205
+ cache_update_index: int, or int Tensor. The index of current inputs in the
206
+ whole sequence.
207
+
208
+ Returns:
209
+ A (logits, hidden_states, cache) tuple. Where `logits` is the
210
+ language model logits for the input token_ids, `hidden_states` is
211
+ the final hidden representation of the input tokens, and `cache` is
212
+ the decoding cache.
213
+ """
214
+ x = self.backbone.token_embedding(token_ids)
215
+ x = x * ops.cast(ops.sqrt(self.backbone.hidden_dim), x.dtype)
216
+ # Each decoder layer has a cache; we update them separately.
217
+ caches = []
218
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
219
+ current_cache = cache[:, i, ...]
220
+ x, next_cache = transformer_layer(
221
+ x,
222
+ cache=current_cache,
223
+ cache_update_index=cache_update_index,
224
+ )
225
+ caches.append(next_cache)
226
+
227
+ cache = ops.stack(caches, axis=1)
228
+ hidden_states = x = self.backbone.layer_norm(x)
229
+ logits = self.backbone.token_embedding(x, reverse=True)
230
+ return logits, hidden_states, cache
231
+
232
+ def _build_cache(self, token_ids):
233
+ """Build an empty cache for use with `call_with_cache()`."""
234
+ batch_size = ops.shape(token_ids)[0]
235
+ max_length = ops.shape(token_ids)[1]
236
+ num_layers = self.backbone.num_layers
237
+ num_heads = self.backbone.num_key_value_heads
238
+ head_dim = self.backbone.head_dim
239
+ shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
240
+ cache = ops.zeros(shape, dtype=self.compute_dtype)
241
+ # Seed the cache.
242
+ _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
243
+ return hidden_states, cache
244
+
245
+ def generate_step(
246
+ self,
247
+ inputs,
248
+ stop_token_ids=None,
249
+ ):
250
+ """A compilable generation function for a single batch of inputs.
251
+
252
+ This function represents the inner, XLA-compilable, generation function
253
+ for a single batch of inputs. Inputs should have the same structure as
254
+ model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
255
+
256
+ Args:
257
+ inputs: A dictionary with two keys `"token_ids"` and
258
+ `"padding_mask"` and batched tensor values.
259
+ stop_token_ids: Tuple of id's of end token's to stop on. If all
260
+ sequences have produced a new stop token, generation
261
+ will stop.
262
+ """
263
+ token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
264
+ # Create and seed cache with a single forward pass.
265
+ hidden_states, cache = self._build_cache(token_ids)
266
+ # Compute the lengths of all user inputted tokens ids.
267
+ row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
268
+ # Start at the first index that has no user inputted id.
269
+ index = ops.min(row_lengths)
270
+
271
+ def next(prompt, cache, index):
272
+ # The cache index is the index of our previous token.
273
+ cache_update_index = index - 1
274
+ batch_size = ops.shape(prompt)[0]
275
+ prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
276
+ logits, hidden_states, cache = self.call_with_cache(
277
+ prompt,
278
+ cache,
279
+ cache_update_index,
280
+ )
281
+ return (
282
+ ops.squeeze(logits, axis=1),
283
+ ops.squeeze(hidden_states, axis=1),
284
+ cache,
285
+ )
286
+
287
+ token_ids = self.sampler(
288
+ next=next,
289
+ prompt=token_ids,
290
+ cache=cache,
291
+ index=index,
292
+ mask=padding_mask,
293
+ stop_token_ids=stop_token_ids,
294
+ hidden_states=hidden_states,
295
+ model=self,
296
+ )
297
+
298
+ # Compute an output padding mask with the token ids we updated.
299
+ if stop_token_ids is not None:
300
+ # Build a mask of `stop_token_ids` locations not in the original
301
+ # prompt (not in locations where `padding_mask` is True).
302
+ end_locations = any_equal(
303
+ token_ids, stop_token_ids, ops.logical_not(padding_mask)
304
+ )
305
+
306
+ end_locations = ops.cast(end_locations, "int32")
307
+ # Use cumsum to get ones in all locations after end_locations.
308
+ cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
309
+ overflow = cumsum - end_locations
310
+ # Our padding mask is the inverse of these overflow locations.
311
+ padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
312
+ else:
313
+ # Without early stopping, all locations will have been updated.
314
+ padding_mask = ops.ones_like(token_ids, dtype="bool")
315
+ return {
316
+ "token_ids": token_ids,
317
+ "padding_mask": padding_mask,
318
+ }
319
+
320
+ def score(
321
+ self,
322
+ token_ids,
323
+ padding_mask=None,
324
+ scoring_mode="logits",
325
+ layer_intercept_fn=None,
326
+ target_ids=None,
327
+ ):
328
+ """Score a generation represented by the provided token ids.
329
+
330
+ Args:
331
+ token_ids: A <int>[batch_size, num_tokens] tensor containing tokens
332
+ to score. Typically, this tensor captures the output from a call
333
+ to `GemmaCausalLM.generate()`, i.e., tokens for both the input
334
+ text and the model-generated text.
335
+ padding_mask: A <bool>[batch_size, num_tokens] tensor indicating the
336
+ tokens that should be preserved during generation. This is an
337
+ artifact required by the GemmaBackbone and isn't influential on
338
+ the computation of this function. If omitted, this function uses
339
+ `keras.ops.ones()` to create a tensor of the appropriate shape.
340
+ scoring_mode: The type of scores to return, either "logits" or
341
+ "loss", both will be per input token.
342
+ layer_intercept_fn: An optional function for augmenting activations
343
+ with additional computation, for example, as part of
344
+ interpretability research. This function will be passed the
345
+ activations as its first parameter and a numeric index
346
+ associated with that backbone layer. _This index _is not_ an
347
+ index into `self.backbone.layers`_. The index -1 accompanies the
348
+ embeddings returned by calling `self.backbone.token_embedding()`
349
+ on `token_ids` in the forward direction. All subsequent indexes
350
+ will be 0-based indices for the activations returned by each of
351
+ the Transformers layers in the backbone. This function must
352
+ return a <float>[batch_size, num_tokens, hidden_dims] tensor
353
+ that can be passed as an input to the next layer in the model.
354
+ target_ids: An <bool>[batch_size, num_tokens] tensor containing the
355
+ predicted tokens against which the loss should be computed. If a
356
+ span of tokens is provided (sequential truthy values along
357
+ axis=1 in the tensor), the loss will be computed as the
358
+ aggregate across those tokens.
359
+
360
+ Raises:
361
+ ValueError: If an unsupported scoring_mode is provided, or if the
362
+ target_ids are not provided when using ScoringMode.LOSS.
363
+
364
+ Returns:
365
+ The per-token scores as a tensor of size
366
+ <float>[batch_size, num_tokens, vocab_size] in "logits" mode, or
367
+ <float>[batch_size, num_tokens] in "loss" mode.
368
+
369
+ Example:
370
+
371
+ Compute gradients between embeddings and loss scores with TensorFlow:
372
+ ```python
373
+ gemma_lm = keras_hub.models.GemmaCausalLM.from_preset(
374
+ "gemma_2b_en"
375
+ )
376
+ generations = gemma_lm.generate(
377
+ ["This is a", "Where are you"],
378
+ max_length=30
379
+ )
380
+ preprocessed = gemma_lm.preprocessor.generate_preprocess(generations)
381
+ generation_ids = preprocessed["token_ids"]
382
+ padding_mask = preprocessed["padding_mask"]
383
+ target_ids = keras.ops.roll(generation_ids, shift=-1, axis=1)
384
+
385
+ embeddings = None
386
+ with tf.GradientTape(watch_accessed_variables=True) as tape:
387
+ def layer_intercept_fn(x, i):
388
+ if i == -1:
389
+ nonlocal embeddings, tape
390
+ embeddings = x
391
+ tape.watch(embeddings)
392
+ return x
393
+
394
+ losses = gemma_lm.score(
395
+ token_ids=generation_ids,
396
+ padding_mask=padding_mask,
397
+ scoring_mode="loss",
398
+ layer_intercept_fn=layer_intercept_fn,
399
+ target_ids=target_ids,
400
+ )
401
+
402
+ grads = tape.gradient(losses, embeddings)
403
+ ```
404
+ """
405
+ if scoring_mode not in ("logits", "loss"):
406
+ raise ValueError(
407
+ "Unsupported scoring_mode. Must be one of 'logits' or 'loss'."
408
+ )
409
+
410
+ if scoring_mode == "loss" and target_ids is None:
411
+ raise ValueError(
412
+ "Cannot compute loss without targets. Please provide target "
413
+ "token ids via the target_ids parameter."
414
+ )
415
+
416
+ batch_shape = ops.shape(token_ids)[:2]
417
+ assert len(batch_shape) == 2
418
+
419
+ if padding_mask is None:
420
+ padding_mask = ops.ones(shape=batch_shape)
421
+
422
+ if layer_intercept_fn is None:
423
+
424
+ def default_layer_intercept_fn(x, unused_i):
425
+ return x
426
+
427
+ layer_intercept_fn = default_layer_intercept_fn
428
+
429
+ token_embeddings = self.backbone.token_embedding(token_ids)
430
+ x = layer_intercept_fn(token_embeddings, -1)
431
+
432
+ x = token_embeddings * ops.cast(
433
+ ops.sqrt(self.backbone.hidden_dim), dtype=self.compute_dtype
434
+ )
435
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
436
+ x = transformer_layer(x, padding_mask=padding_mask)
437
+ x = layer_intercept_fn(x, i)
438
+ x = self.backbone.layer_norm(x)
439
+ logits = self.backbone.token_embedding(x, reverse=True)
440
+
441
+ if scoring_mode == "logits":
442
+ return logits
443
+
444
+ per_token_loss_fn = keras.losses.SparseCategoricalCrossentropy(
445
+ from_logits=True, reduction="none"
446
+ )
447
+ per_token_loss = per_token_loss_fn(target_ids, logits)
448
+ return per_token_loss
@@ -0,0 +1,167 @@
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.models.gemma.gemma_preprocessor import GemmaPreprocessor
20
+ from keras_hub.src.utils.keras_utils import (
21
+ convert_inputs_to_list_of_tensor_segments,
22
+ )
23
+ from keras_hub.src.utils.tensor_utils import strip_to_ragged
24
+
25
+
26
+ @keras_hub_export("keras_hub.models.GemmaCausalLMPreprocessor")
27
+ class GemmaCausalLMPreprocessor(GemmaPreprocessor):
28
+ """Gemma Causal LM preprocessor.
29
+
30
+ This preprocessing layer is meant for use with
31
+ `keras_hub.models.GemmaCausalLM`. By default, it will take in batches of
32
+ strings, and return outputs in a `(x, y, sample_weight)` format, where the
33
+ `y` label is the next token id in the `x` sequence.
34
+
35
+ For use with generation, the layer also exposes two methods
36
+ `generate_preprocess()` and `generate_postprocess()`. When this preprocessor
37
+ is attached to a `keras_hub.models.GemmaCausalLM` instance, these methods
38
+ will be called implicitly in `generate()`. They can also be called
39
+ standalone (e.g. to precompute preprocessing inputs for generation in a
40
+ separate process).
41
+
42
+ Args:
43
+ tokenizer: A `keras_hub.models.GemmaTokenizer` instance.
44
+ sequence_length: The length of the packed inputs.
45
+ add_start_token: If `True`, the preprocessor will prepend the tokenizer
46
+ start token to each input sequence.
47
+ add_end_token: If `True`, the preprocessor will append the tokenizer
48
+ end token to each input sequence.
49
+
50
+ Call arguments:
51
+ x: A string, `tf.Tensor` or list of python strings.
52
+ y: Label data. Should always be `None` as the layer generates labels.
53
+ sample_weight: Label weights. Should always be `None` as the layer
54
+ generates label weights.
55
+ sequence_length: Pass to override the configured `sequence_length` of
56
+ the layer.
57
+
58
+ Examples:
59
+ ```python
60
+ # Load the preprocessor from a preset.
61
+ preprocessor = keras_hub.models.GemmaCausalLMPreprocessor.from_preset(
62
+ "gemma_2b_en"
63
+ )
64
+
65
+ # Tokenize and pack a single sentence.
66
+ preprocessor("The quick brown fox jumped.")
67
+
68
+ # Tokenize a batch of sentences.
69
+ preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
70
+
71
+ # Apply tokenization to a `tf.data.Dataset`.
72
+ features = tf.constant(["The quick brown fox.", "Call me Ishmael."])
73
+ ds = tf.data.Dataset.from_tensor_slices(features)
74
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
75
+
76
+ # Prepare tokens for generation (no end token).
77
+ preprocessor.generate_preprocess(["The quick brown fox jumped."])
78
+
79
+ # Map generation outputs back to strings.
80
+ preprocessor.generate_postprocess({
81
+ 'token_ids': np.array([[2, 714, 4320, 8426, 25341, 32292, 235265, 0]]),
82
+ 'padding_mask': np.array([[ 1, 1, 1, 1, 1, 1, 1, 0]]),
83
+ })
84
+ ```
85
+ """
86
+
87
+ def call(
88
+ self,
89
+ x,
90
+ y=None,
91
+ sample_weight=None,
92
+ sequence_length=None,
93
+ ):
94
+ if y is not None or sample_weight is not None:
95
+ logging.warning(
96
+ "`GemmaCausalLMPreprocessor` generates `y` and `sample_weight` "
97
+ "based on your input data, but your data already contains `y` "
98
+ "or `sample_weight`. Your `y` and `sample_weight` will be "
99
+ "ignored."
100
+ )
101
+ sequence_length = sequence_length or self.sequence_length
102
+
103
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
104
+ x = self.tokenizer(x)
105
+ # Pad with one extra token to account for the truncation below.
106
+ token_ids, padding_mask = self.packer(
107
+ x,
108
+ sequence_length=sequence_length + 1,
109
+ add_start_value=self.add_start_token,
110
+ add_end_value=self.add_end_token,
111
+ )
112
+ # The last token does not have a next token, so we truncate it out.
113
+ x = {
114
+ "token_ids": token_ids[..., :-1],
115
+ "padding_mask": padding_mask[..., :-1],
116
+ }
117
+ # Target `y` will be the next token.
118
+ y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
119
+ return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
120
+
121
+ def generate_preprocess(
122
+ self,
123
+ x,
124
+ sequence_length=None,
125
+ ):
126
+ """Convert strings to integer token input for generation.
127
+
128
+ Similar to calling the layer for training, this method takes in strings
129
+ or tensor strings, tokenizes and packs the input, and computes a padding
130
+ mask masking all inputs not filled in with a padded value.
131
+
132
+ Unlike calling the layer for training, this method does not compute
133
+ labels and will never append a `tokenizer.end_token_id` to the end of
134
+ the sequence (as generation is expected to continue at the end of the
135
+ inputted prompt).
136
+ """
137
+ if not self.built:
138
+ self.build(None)
139
+
140
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
141
+ x = self.tokenizer(x)
142
+ token_ids, padding_mask = self.packer(
143
+ x, sequence_length=sequence_length, add_end_value=False
144
+ )
145
+ return {
146
+ "token_ids": token_ids,
147
+ "padding_mask": padding_mask,
148
+ }
149
+
150
+ def generate_postprocess(self, x):
151
+ """Convert integer token output to strings for generation.
152
+
153
+ This method reverses `generate_preprocess()`, by first removing all
154
+ padding and start/end tokens, and then converting the integer sequence
155
+ back to a string.
156
+ """
157
+ if not self.built:
158
+ self.build(None)
159
+
160
+ token_ids, padding_mask = x["token_ids"], x["padding_mask"]
161
+ ids_to_strip = (
162
+ self.tokenizer.start_token_id,
163
+ self.tokenizer.end_token_id,
164
+ self.tokenizer.pad_token_id,
165
+ )
166
+ token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
167
+ return self.tokenizer.detokenize(token_ids)