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 +1,173 @@
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.token_and_position_embedding import (
20
+ TokenAndPositionEmbedding,
21
+ )
22
+ from keras_hub.src.layers.modeling.transformer_decoder import TransformerDecoder
23
+ from keras_hub.src.models.backbone import Backbone
24
+
25
+
26
+ def opt_kernel_initializer(stddev=0.02):
27
+ return keras.initializers.TruncatedNormal(stddev=stddev)
28
+
29
+
30
+ @keras_hub_export("keras_hub.models.OPTBackbone")
31
+ class OPTBackbone(Backbone):
32
+ """An OPT decoder network.
33
+
34
+ This class implements a Transformer-based decoder model as described in
35
+ ["OPT: Open Pre-trained Transformer Language Models"](https://arxiv.org/abs/2205.01068).
36
+ The default constructor gives a fully customizable, randomly initialized OPT
37
+ model with any number of layers, heads, and embedding dimensions. To load
38
+ preset architectures and weights, use the `from_preset()` constructor.
39
+
40
+ Disclaimer: Pre-trained models are provided on an "as is" basis, without
41
+ warranties or conditions of any kind. The underlying model is provided by a
42
+ third party and subject to a separate license, available
43
+ [here](https://github.com/facebookresearch/fairseq/).
44
+
45
+ Args:
46
+ vocabulary_size: int. The size of the token vocabulary.
47
+ num_layers: int. The number of transformer decoder layers.
48
+ num_heads: int. The number of attention heads for each transformer.
49
+ The hidden size must be divisible by the number of attention heads.
50
+ hidden_dim: int. The hidden size of the transformer decoder layers.
51
+ intermediate_dim: int. The output dimension of the first Dense layer in
52
+ a two-layer feedforward network for each transformer decoder layer.
53
+ dropout: float. Dropout probability for the Transformer decoder.
54
+ max_sequence_length: int. The maximum sequence length that this decoder
55
+ can consume. If `None`, `max_sequence_length` uses the value from
56
+ sequence length. This determines the variable shape for positional
57
+ embeddings.
58
+ dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
59
+ for model computations and weights. Note that some computations,
60
+ such as softmax and layer normalization, will always be done at
61
+ float32 precision regardless of dtype.
62
+
63
+ Examples:
64
+ ```python
65
+ input_data = {
66
+ "token_ids": np.ones(shape=(1, 12), dtype="int32"),
67
+ "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
68
+ }
69
+
70
+ # Pretrained OPT decoder
71
+ model = keras_hub.models.OPTBackbone.from_preset("opt_125m_en")
72
+ model(input_data)
73
+
74
+ # Randomly initialized OPT decoder model with a custom config
75
+ model = keras_hub.models.OPTBackbone(
76
+ vocabulary_size=50265,
77
+ num_layers=4,
78
+ num_heads=4,
79
+ hidden_dim=256,
80
+ intermediate_dim=512,
81
+ max_sequence_length=128,
82
+ )
83
+ model(input_data)
84
+ ```
85
+ """
86
+
87
+ def __init__(
88
+ self,
89
+ vocabulary_size,
90
+ num_layers,
91
+ num_heads,
92
+ hidden_dim,
93
+ intermediate_dim,
94
+ dropout=0.1,
95
+ max_sequence_length=2048,
96
+ dtype=None,
97
+ **kwargs,
98
+ ):
99
+ # === Layers ===
100
+ self.embeddings = TokenAndPositionEmbedding(
101
+ vocabulary_size=vocabulary_size,
102
+ sequence_length=max_sequence_length,
103
+ embedding_dim=hidden_dim,
104
+ embeddings_initializer=opt_kernel_initializer(),
105
+ dtype=dtype,
106
+ name="embeddings",
107
+ )
108
+ self.token_embedding = self.embeddings.token_embedding
109
+ self.transformer_layers = []
110
+ for i in range(num_layers):
111
+ layer = TransformerDecoder(
112
+ intermediate_dim=intermediate_dim,
113
+ num_heads=num_heads,
114
+ dropout=dropout,
115
+ activation="relu",
116
+ layer_norm_epsilon=1e-5,
117
+ normalize_first=True,
118
+ kernel_initializer=opt_kernel_initializer(),
119
+ dtype=dtype,
120
+ name=f"transformer_layer_{i}",
121
+ )
122
+ self.transformer_layers.append(layer)
123
+ self.layer_norm = keras.layers.LayerNormalization(
124
+ axis=-1,
125
+ epsilon=1e-5,
126
+ dtype=dtype,
127
+ name="layer_norm",
128
+ )
129
+
130
+ # === Functional Model ===
131
+ token_id_input = keras.Input(
132
+ shape=(None,), dtype="int32", name="token_ids"
133
+ )
134
+ padding_mask_input = keras.Input(
135
+ shape=(None,), dtype="int32", name="padding_mask"
136
+ )
137
+ x = self.embeddings(token_id_input)
138
+ for transformer_layer in self.transformer_layers:
139
+ x = transformer_layer(x, decoder_padding_mask=padding_mask_input)
140
+ x = self.layer_norm(x)
141
+ super().__init__(
142
+ inputs={
143
+ "token_ids": token_id_input,
144
+ "padding_mask": padding_mask_input,
145
+ },
146
+ outputs=x,
147
+ dtype=dtype,
148
+ **kwargs,
149
+ )
150
+
151
+ # === Config ===
152
+ self.vocabulary_size = vocabulary_size
153
+ self.num_layers = num_layers
154
+ self.num_heads = num_heads
155
+ self.hidden_dim = hidden_dim
156
+ self.intermediate_dim = intermediate_dim
157
+ self.dropout = dropout
158
+ self.max_sequence_length = max_sequence_length
159
+
160
+ def get_config(self):
161
+ config = super().get_config()
162
+ config.update(
163
+ {
164
+ "vocabulary_size": self.vocabulary_size,
165
+ "num_layers": self.num_layers,
166
+ "num_heads": self.num_heads,
167
+ "hidden_dim": self.hidden_dim,
168
+ "intermediate_dim": self.intermediate_dim,
169
+ "dropout": self.dropout,
170
+ "max_sequence_length": self.max_sequence_length,
171
+ }
172
+ )
173
+ return config
@@ -0,0 +1,301 @@
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 import ops
17
+
18
+ from keras_hub.src.api_export import keras_hub_export
19
+ from keras_hub.src.models.causal_lm import CausalLM
20
+ from keras_hub.src.models.opt.opt_backbone import OPTBackbone
21
+ from keras_hub.src.models.opt.opt_causal_lm_preprocessor import (
22
+ OPTCausalLMPreprocessor,
23
+ )
24
+ from keras_hub.src.utils.tensor_utils import any_equal
25
+
26
+
27
+ @keras_hub_export("keras_hub.models.OPTCausalLM")
28
+ class OPTCausalLM(CausalLM):
29
+ """An end-to-end OPT model for causal language modeling.
30
+
31
+ A causal language model (LM) predicts the next token based on previous
32
+ tokens. This task setup can be used to train the model unsupervised on
33
+ plain text input, or to autoregressively generate plain text similar to
34
+ the data used for training. This task can be used for pre-training or
35
+ fine-tuning a GPT-2 model, simply by calling `fit()`.
36
+
37
+ This model has a `generate()` method, which generates text based on a
38
+ prompt. The generation strategy used is controlled by an additional
39
+ `sampler` argument on `compile()`. You can recompile the model with
40
+ different `keras_hub.samplers` objects to control the generation. By
41
+ default, `"top_k"` sampling will be used.
42
+
43
+ This model can optionally be configured with a `preprocessor` layer, in
44
+ which case it will automatically apply preprocessing to string inputs during
45
+ `fit()`, `predict()`, `evaluate()` and `generate()`. This is done by default
46
+ when creating the model with `from_preset()`.
47
+
48
+ Disclaimer: Pre-trained models are provided on an "as is" basis, without
49
+ warranties or conditions of any kind. The underlying model is provided by a
50
+ third party and subject to a separate license, available
51
+ [here](https://github.com/facebookresearch/fairseq/).
52
+
53
+ Args:
54
+ backbone: A `keras_hub.models.OPTBackbone` instance.
55
+ preprocessor: A `keras_hub.models.OPTCausalLMPreprocessor` or `None`.
56
+ If `None`, this model will not apply preprocessing, and inputs
57
+ should be preprocessed before calling the model.
58
+
59
+ Examples:
60
+
61
+ Use `generate()` to do text generation.
62
+ ```python
63
+ opt_lm = keras_hub.models.OPTCausalLM.from_preset("opt_125m_en")
64
+ opt_lm.generate("I want to say", max_length=30)
65
+
66
+ # Generate with batched prompts.
67
+ opt_lm.generate(["This is a", "Where are you"], max_length=30)
68
+ ```
69
+
70
+ Compile the `generate()` function with a custom sampler.
71
+ ```python
72
+ opt_lm = keras_hub.models.OPTCausalLM.from_preset("opt_125m_en")
73
+ opt_lm.compile(sampler="greedy")
74
+ opt_lm.generate("I want to say", max_length=30)
75
+
76
+ opt_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
77
+ opt_lm.generate("I want to say", max_length=30)
78
+ ```
79
+
80
+ Use `generate()` without preprocessing.
81
+ ```python
82
+ # Prompt the model with `5338, 318` (the token ids for `"Who is"`).
83
+ # Use `"padding_mask"` to indicate values that should not be overridden.
84
+ prompt = {
85
+ "token_ids": np.array([[5338, 318, 0, 0, 0]] * 2),
86
+ "padding_mask": np.array([[1, 1, 0, 0, 0]] * 2),
87
+ }
88
+
89
+ opt_lm = keras_hub.models.OPTCausalLM.from_preset(
90
+ "opt_125m_en",
91
+ preprocessor=None,
92
+ )
93
+ opt_lm.generate(prompt)
94
+ ```
95
+
96
+ Call `fit()` on a single batch.
97
+ ```python
98
+ features = ["The quick brown fox jumped.", "I forgot my homework."]
99
+ opt_lm = keras_hub.models.OPTCausalLM.from_preset("opt_125m_en")
100
+ opt_lm.fit(x=features, batch_size=2)
101
+ ```
102
+
103
+ Call `fit()` without preprocessing.
104
+ ```python
105
+ x = {
106
+ "token_ids": np.array([[1, 2, 3, 4, 5]] * 2),
107
+ "padding_mask": np.array([[1, 1, 1, 1, 1]] * 2),
108
+ }
109
+ y = np.array([[2, 3, 4, 5, 0]] * 2)
110
+ sw = np.array([[1, 1, 1, 1, 1]] * 2)
111
+
112
+ opt_lm = keras_hub.models.OPTCausalLM.from_preset(
113
+ "opt_base_en",
114
+ preprocessor=None,
115
+ )
116
+ opt_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)
117
+ ```
118
+
119
+ Custom backbone and vocabulary.
120
+ ```python
121
+ features = ["a quick fox.", "a fox quick."]
122
+ vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
123
+ merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
124
+ merges += ["Ġ f", "o x", "Ġf ox"]
125
+
126
+ tokenizer = keras_hub.models.OPTTokenizer(
127
+ vocabulary=vocab,
128
+ merges=merges,
129
+ )
130
+ preprocessor = keras_hub.models.OPTCausalLMPreprocessor(
131
+ tokenizer=tokenizer,
132
+ sequence_length=128,
133
+ )
134
+ model = keras_hub.models.OPTBackbone(
135
+ vocabulary_size=50265,
136
+ num_layers=4,
137
+ num_heads=4,
138
+ hidden_dim=256,
139
+ intermediate_dim=512,
140
+ max_sequence_length=128,
141
+ )
142
+ opt_lm = keras_hub.models.OPTCausalLM(
143
+ backbone=backbone,
144
+ preprocessor=preprocessor,
145
+ )
146
+ opt_lm.fit(x=features, batch_size=2)
147
+ ```
148
+ """
149
+
150
+ backbone_cls = OPTBackbone
151
+ preprocessor_cls = OPTCausalLMPreprocessor
152
+
153
+ def __init__(
154
+ self,
155
+ backbone,
156
+ preprocessor=None,
157
+ **kwargs,
158
+ ):
159
+ # === Layers ===
160
+ self.backbone = backbone
161
+ self.preprocessor = preprocessor
162
+
163
+ # === Functional Model ===
164
+ inputs = backbone.input
165
+ hidden_states = backbone(inputs)
166
+ outputs = backbone.token_embedding(hidden_states, reverse=True)
167
+ super().__init__(
168
+ inputs=inputs,
169
+ outputs=outputs,
170
+ **kwargs,
171
+ )
172
+
173
+ def call_with_cache(
174
+ self,
175
+ token_ids,
176
+ cache,
177
+ cache_update_index,
178
+ ):
179
+ """Forward pass of `OPTCausalLM` with cache.
180
+
181
+ `call_with_cache` adds an additional forward pass for the model for
182
+ autoregressive inference. Unlike calling the model directly, this method
183
+ allows caching previous key/value Tensors in multi-head attention layer,
184
+ and avoids recomputing the outputs of seen tokens.
185
+
186
+ Args:
187
+ token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
188
+ cache: a dense float Tensor, the cache of key and value.
189
+ cache_update_index: int, or int Tensor. The index of current inputs in the
190
+ whole sequence.
191
+
192
+ Returns:
193
+ A (logits, hidden_states, cache) tuple. Where `logits` is the
194
+ language model logits for the input token_ids, `hidden_states` is
195
+ the final hidden representation of the input tokens, and `cache` is
196
+ the decoding cache.
197
+ """
198
+ x = self.backbone.embeddings(token_ids, start_index=cache_update_index)
199
+ # Each decoder layer has a cache; we update them separately.
200
+ caches = []
201
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
202
+ current_cache = cache[:, i, ...]
203
+ x, next_cache = transformer_layer(
204
+ x,
205
+ self_attention_cache=current_cache,
206
+ self_attention_cache_update_index=cache_update_index,
207
+ )
208
+ caches.append(next_cache)
209
+ cache = ops.stack(caches, axis=1)
210
+ x = self.backbone.layer_norm(x)
211
+ hidden_states = x
212
+ logits = self.backbone.token_embedding(hidden_states, reverse=True)
213
+ return logits, hidden_states, cache
214
+
215
+ def _build_cache(self, token_ids):
216
+ """Build an empty cache for use with `call_with_cache()`."""
217
+ batch_size = ops.shape(token_ids)[0]
218
+ max_length = ops.shape(token_ids)[1]
219
+ num_layers = self.backbone.num_layers
220
+ num_heads = self.backbone.num_heads
221
+ head_dim = self.backbone.hidden_dim // self.backbone.num_heads
222
+ shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
223
+ cache = ops.zeros(shape, dtype=self.compute_dtype)
224
+ # Seed the cache.
225
+ _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
226
+ return hidden_states, cache
227
+
228
+ def generate_step(
229
+ self,
230
+ inputs,
231
+ stop_token_ids=None,
232
+ ):
233
+ """A compilable generation function for a single batch of inputs.
234
+
235
+ This function represents the inner, XLA-compilable, generation function
236
+ for a single batch of inputs. Inputs should have the same structure as
237
+ model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
238
+
239
+ Args:
240
+ inputs: A dictionary with two keys `"token_ids"` and
241
+ `"padding_mask"` and batched tensor values.
242
+ stop_token_ids: Tuple of id's of end token's to stop on. If all
243
+ sequences have produced a new stop token, generation
244
+ will stop.
245
+ """
246
+ token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
247
+ # Create and seed cache with a single forward pass.
248
+ hidden_states, cache = self._build_cache(token_ids)
249
+ # Compute the lengths of all user inputted tokens ids.
250
+ row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
251
+ # Start at the first index that has no user inputted id.
252
+ index = ops.min(row_lengths)
253
+
254
+ def next(prompt, cache, index):
255
+ # The cache index is the index of our previous token.
256
+ cache_update_index = index - 1
257
+ batch_size = ops.shape(prompt)[0]
258
+ prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
259
+ logits, hidden_states, cache = self.call_with_cache(
260
+ prompt,
261
+ cache,
262
+ cache_update_index,
263
+ )
264
+ return (
265
+ ops.squeeze(logits, axis=1),
266
+ ops.squeeze(hidden_states, axis=1),
267
+ cache,
268
+ )
269
+
270
+ token_ids = self.sampler(
271
+ next=next,
272
+ prompt=token_ids,
273
+ cache=cache,
274
+ index=index,
275
+ mask=padding_mask,
276
+ stop_token_ids=stop_token_ids,
277
+ hidden_states=hidden_states,
278
+ model=self,
279
+ )
280
+
281
+ # Compute an output padding mask with the token ids we updated.
282
+ if stop_token_ids is not None:
283
+ # Build a mask of stop token locations not in the original
284
+ # prompt (not in locations where `padding_mask` is True).
285
+ end_locations = any_equal(
286
+ token_ids, stop_token_ids, ops.logical_not(padding_mask)
287
+ )
288
+
289
+ end_locations = ops.cast(end_locations, "int32")
290
+ # Use cumsum to get ones in all locations after end_locations.
291
+ cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
292
+ overflow = cumsum - end_locations
293
+ # Our padding mask is the inverse of these overflow locations.
294
+ padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
295
+ else:
296
+ # Without early stopping, all locations will have been updated.
297
+ padding_mask = ops.ones_like(token_ids, dtype="bool")
298
+ return {
299
+ "token_ids": token_ids,
300
+ "padding_mask": padding_mask,
301
+ }
@@ -0,0 +1,177 @@
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.opt.opt_preprocessor import OPTPreprocessor
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.OPTCausalLMPreprocessor")
27
+ class OPTCausalLMPreprocessor(OPTPreprocessor):
28
+ """OPT Causal LM preprocessor.
29
+
30
+ This preprocessing layer is primarily meant to be used with
31
+ `keras_hub.models.OPTCausalLM`. 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. For use with generation,
34
+ pass `return_labels=False` in which case the output will simply be the
35
+ encoded string features.
36
+
37
+ Args:
38
+ tokenizer: A `keras_hub.models.OPTTokenizer` instance.
39
+ sequence_length: The length of the packed inputs.
40
+ add_start_token: If `True`, the preprocessor will prepend the tokenizer
41
+ start token to each input sequence.
42
+ add_end_token: If `True`, the preprocessor will append the tokenizer
43
+ end token to each input sequence.
44
+
45
+ Call arguments:
46
+ x: A string, `tf.Tensor` or list of python strings.
47
+ y: Label data. Should always be `None` as the layer generates labels.
48
+ sample_weight: Label weights. Should always be `None` as the layer
49
+ generates label weights.
50
+ sequence_length: Pass to override the configured `sequence_length` of
51
+ the layer.
52
+ add_start_token: Pass to override the configured value of
53
+ `add_start_token` on the layer.
54
+ add_end_token: Pass to override the configured value of
55
+ `add_end_token` on the layer.
56
+ return_labels: If `True`, the output `"token_ids"` will be offset by one
57
+ and returned as labels. If `False` only features will be returned.
58
+
59
+ Examples:
60
+ ```python
61
+ # Load the preprocessor from a preset.
62
+ preprocessor = keras_hub.models.OPTCausalLMPreprocessor.from_preset(
63
+ "opt_125m_en"
64
+ )
65
+
66
+ # Tokenize and pack a single sentence.
67
+ sentence = tf.constant("League of legends")
68
+ preprocessor(sentence)
69
+ # Same output.
70
+ preprocessor("League of legends")
71
+
72
+ # Tokenize a batch of sentences.
73
+ sentences = tf.constant(["Taco tuesday", "Fish taco please!"])
74
+ preprocessor(sentences)
75
+ # Same output.
76
+ preprocessor(["Taco tuesday", "Fish taco please!"])
77
+
78
+ # Map a dataset to preprocess a single sentence.
79
+ features = tf.constant(
80
+ [
81
+ "Avatar 2 is amazing!",
82
+ "Well, I am not sure.",
83
+ ]
84
+ )
85
+ labels = tf.constant([1, 0])
86
+ ds = tf.data.Dataset.from_tensor_slices((features, labels))
87
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
88
+
89
+ # Map a dataset to preprocess unlabled sentences.
90
+ ds = tf.data.Dataset.from_tensor_slices(features)
91
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
92
+ ```
93
+ """
94
+
95
+ def call(
96
+ self,
97
+ x,
98
+ y=None,
99
+ sample_weight=None,
100
+ sequence_length=None,
101
+ ):
102
+ if y is not None or sample_weight is not None:
103
+ logging.warning(
104
+ "`GPT2CausalLMPreprocessor` generates `y` and `sample_weight` "
105
+ "based on your input data, but your data already contains `y` "
106
+ "or `sample_weight`. Your `y` and `sample_weight` will be "
107
+ "ignored."
108
+ )
109
+ sequence_length = sequence_length or self.sequence_length
110
+
111
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
112
+ x = self.tokenizer(x)
113
+ # Pad with one extra token to account for the truncation below.
114
+ token_ids, padding_mask = self.packer(
115
+ x,
116
+ sequence_length=sequence_length + 1,
117
+ add_start_value=self.add_start_token,
118
+ add_end_value=self.add_end_token,
119
+ )
120
+ # The last token does not have a next token, so we truncate it out.
121
+ x = {
122
+ "token_ids": token_ids[..., :-1],
123
+ "padding_mask": padding_mask[..., :-1],
124
+ }
125
+ # Target `y` will be the next token.
126
+ y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
127
+ return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
128
+
129
+ def generate_preprocess(
130
+ self,
131
+ x,
132
+ sequence_length=None,
133
+ ):
134
+ """Convert strings to integer token input for generation.
135
+
136
+ Similar to calling the layer for training, this method takes in strings
137
+ or tensor strings, tokenizes and packs the input, and computes a padding
138
+ mask masking all inputs not filled in with a padded value.
139
+
140
+ Unlike calling the layer for training, this method does not compute
141
+ labels and will never append a `tokenizer.end_token_id` to the end of
142
+ the sequence (as generation is expected to continue at the end of the
143
+ inputted prompt).
144
+ """
145
+ if not self.built:
146
+ self.build(None)
147
+
148
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
149
+ x = self.tokenizer(x)
150
+ token_ids, padding_mask = self.packer(
151
+ x, sequence_length=sequence_length, add_end_value=False
152
+ )
153
+ return {
154
+ "token_ids": token_ids,
155
+ "padding_mask": padding_mask,
156
+ }
157
+
158
+ def generate_postprocess(
159
+ self,
160
+ x,
161
+ ):
162
+ """Convert integer token output to strings for generation.
163
+
164
+ This method reverses `generate_preprocess()`, by first removing all
165
+ padding and start/end tokens, and then converting the integer sequence
166
+ back to a string.
167
+ """
168
+ if not self.built:
169
+ self.build(None)
170
+
171
+ token_ids, padding_mask = x["token_ids"], x["padding_mask"]
172
+ ids_to_strip = (
173
+ self.tokenizer.end_token_id,
174
+ self.tokenizer.pad_token_id,
175
+ )
176
+ token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
177
+ return self.tokenizer.detokenize(token_ids)