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 +1,175 @@
1
+ # Copyright 2024 The KerasHub Authors
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # https://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import keras
16
+
17
+ from keras_hub.src.api_export import keras_hub_export
18
+ from keras_hub.src.layers.modeling.reversible_embedding import (
19
+ ReversibleEmbedding,
20
+ )
21
+ from keras_hub.src.models.backbone import Backbone
22
+ from keras_hub.src.models.gpt_neo_x.gpt_neo_x_decoder import GPTNeoXDecoder
23
+ from keras_hub.src.utils.keras_utils import gelu_approximate
24
+
25
+
26
+ def _gpt_neo_x_kernel_initializer(stddev=0.02):
27
+ return keras.initializers.RandomNormal(stddev=stddev)
28
+
29
+
30
+ @keras_hub_export("keras_hub.models.GPTNeoXBackbone")
31
+ class GPTNeoXBackbone(Backbone):
32
+ """GPT-NeoX core network with hyperparameters.
33
+
34
+ This network implements a Transformer-based decoder network,
35
+ Generative Pretrained Transformer-Neo-X (GPTNeoX), as described in
36
+ ["GPT-NeoX-20B: An Open-Source Autoregressive Language Model"](https://arxiv.org/abs/2204.06745).
37
+ It includes the embedding lookups and transformer layers.
38
+
39
+ The default constructor gives a fully customizable, randomly initialized
40
+ GPT-NeoX model with any number of layers, heads, and embedding
41
+ dimensions.
42
+
43
+ Disclaimer: Pre-trained models are provided on an "as is" basis, without
44
+ warranties or conditions of any kind. The underlying model is provided by a
45
+ third party and subject to a separate license, available
46
+ [here](https://github.com/EleutherAI/gpt-neox/).
47
+
48
+ Args:
49
+ vocabulary_size: int. The size of the token vocabulary.
50
+ num_layers: int. The number of transformer layers.
51
+ num_heads: int. The number of attention heads for each transformer.
52
+ The hidden size must be divisible by the number of attention heads.
53
+ hidden_dim: int. The size of the transformer encoding and pooler layers.
54
+ intermediate_dim: int. The output dimension of the first Dense layer in
55
+ a two-layer feedforward network for each transformer.
56
+ dropout: float. Dropout probability for the Transformer encoder.
57
+ layer_norm_epsilon: float. a value added to the denominator for
58
+ numerical stability.
59
+ rotary_max_wavelength: int. The maximum angular wavelength of the
60
+ sine/cosine curves, for rotary embeddings.
61
+ rotary_percentage: float. The percentage by which query, key, value
62
+ matrices are to be rotated
63
+ max_sequence_length: int. The maximum sequence length that this encoder
64
+ can consume. If `None`, `max_sequence_length` uses the value from
65
+ sequence length. This determines the variable shape for positional
66
+ embeddings.
67
+ dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
68
+ for model computations and weights. Note that some computations,
69
+ such as softmax and layer normalization, will always be done at
70
+ float32 precision regardless of dtype.
71
+ """
72
+
73
+ def __init__(
74
+ self,
75
+ vocabulary_size,
76
+ num_layers,
77
+ num_heads,
78
+ hidden_dim,
79
+ intermediate_dim,
80
+ dropout=0.0,
81
+ rotary_percentage=0.25,
82
+ rotary_max_wavelength=10000,
83
+ layer_norm_epsilon=1e-5,
84
+ max_sequence_length=512,
85
+ dtype=None,
86
+ **kwargs,
87
+ ):
88
+ # === Layers ===
89
+ self.token_embedding = ReversibleEmbedding(
90
+ input_dim=vocabulary_size,
91
+ output_dim=hidden_dim,
92
+ embeddings_initializer=_gpt_neo_x_kernel_initializer(stddev=0.01),
93
+ dtype=dtype,
94
+ name="token_embedding",
95
+ )
96
+ self.embeddings_dropout = keras.layers.Dropout(
97
+ dropout,
98
+ dtype=dtype,
99
+ name="embeddings_dropout",
100
+ )
101
+ self.transformer_layers = []
102
+ for i in range(num_layers):
103
+ layer = GPTNeoXDecoder(
104
+ intermediate_dim=intermediate_dim,
105
+ num_heads=num_heads,
106
+ dropout=dropout,
107
+ max_sequence_length=max_sequence_length,
108
+ rotary_percentage=rotary_percentage,
109
+ rotary_max_wavelength=rotary_max_wavelength,
110
+ layer_norm_epsilon=layer_norm_epsilon,
111
+ activation=gelu_approximate,
112
+ kernel_initializer=_gpt_neo_x_kernel_initializer(stddev=0.02),
113
+ dtype=dtype,
114
+ name=f"transformer_layer_{i}",
115
+ )
116
+ self.transformer_layers.append(layer)
117
+ self.layer_norm = keras.layers.LayerNormalization(
118
+ axis=-1,
119
+ epsilon=layer_norm_epsilon,
120
+ dtype=dtype,
121
+ name="layer_norm",
122
+ )
123
+
124
+ # === Functional Model ===
125
+ token_id_input = keras.Input(
126
+ shape=(None,), dtype="int32", name="token_ids"
127
+ )
128
+ padding_mask_input = keras.Input(
129
+ shape=(None,), dtype="int32", name="padding_mask"
130
+ )
131
+ # Embed tokens.
132
+ x = self.token_embedding(token_id_input)
133
+ x = self.embeddings_dropout(x)
134
+ for transformer_layer in self.transformer_layers:
135
+ x = transformer_layer(x, decoder_padding_mask=padding_mask_input)
136
+ sequence_output = self.layer_norm(x)
137
+ super().__init__(
138
+ inputs={
139
+ "token_ids": token_id_input,
140
+ "padding_mask": padding_mask_input,
141
+ },
142
+ outputs=sequence_output,
143
+ dtype=dtype,
144
+ **kwargs,
145
+ )
146
+
147
+ # === Config ===
148
+ self.vocabulary_size = vocabulary_size
149
+ self.num_layers = num_layers
150
+ self.num_heads = num_heads
151
+ self.hidden_dim = hidden_dim
152
+ self.intermediate_dim = intermediate_dim
153
+ self.dropout = dropout
154
+ self.rotary_percentage = rotary_percentage
155
+ self.rotary_max_wavelength = rotary_max_wavelength
156
+ self.max_sequence_length = max_sequence_length
157
+ self.layer_norm_epsilon = layer_norm_epsilon
158
+
159
+ def get_config(self):
160
+ config = super().get_config()
161
+ config.update(
162
+ {
163
+ "vocabulary_size": self.vocabulary_size,
164
+ "num_layers": self.num_layers,
165
+ "num_heads": self.num_heads,
166
+ "hidden_dim": self.hidden_dim,
167
+ "intermediate_dim": self.intermediate_dim,
168
+ "dropout": self.dropout,
169
+ "rotary_percentage": self.rotary_percentage,
170
+ "rotary_max_wavelength": self.rotary_max_wavelength,
171
+ "max_sequence_length": self.max_sequence_length,
172
+ "layer_norm_epsilon": self.layer_norm_epsilon,
173
+ }
174
+ )
175
+ return config
@@ -0,0 +1,201 @@
1
+ # Copyright 2024 The KerasHub Authors
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # https://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ from keras import ops
16
+
17
+ from keras_hub.src.api_export import keras_hub_export
18
+ from keras_hub.src.models.causal_lm import CausalLM
19
+ from keras_hub.src.models.gpt_neo_x.gpt_neo_x_backbone import GPTNeoXBackbone
20
+ from keras_hub.src.models.gpt_neo_x.gpt_neo_x_causal_lm_preprocessor import (
21
+ GPTNeoXCausalLMPreprocessor,
22
+ )
23
+ from keras_hub.src.utils.tensor_utils import any_equal
24
+
25
+
26
+ @keras_hub_export("keras_hub.models.GPTNeoXCausalLM")
27
+ class GPTNeoXCausalLM(CausalLM):
28
+ """An end-to-end GPTNeoX model for causal language modeling.
29
+
30
+ A causal language model (LM) predicts the next token based on previous
31
+ tokens. This task setup can be used to train the model unsupervised on
32
+ plain text input, or to autoregressively generate plain text similar to
33
+ the data used for training. This task can be used for pre-training or
34
+ fine-tuning a GPT-NeoX model, simply by calling `fit()`.
35
+
36
+ This model has a `generate()` method, which generates text based on a
37
+ prompt. The generation strategy used is controlled by an additional
38
+ `sampler` argument on `compile()`. You can recompile the model with
39
+ different `keras_hub.samplers` objects to control the generation. By
40
+ default, `"top_k"` sampling will be used.
41
+
42
+ Args:
43
+ backbone: A `keras_hub.models.GPTNeoXBackbone` instance.
44
+ preprocessor: A `keras_hub.models.GPTNeoXCausalLMPreprocessor` or `None`.
45
+ If `None`, this model will not apply preprocessing, and inputs
46
+ should be preprocessed before calling the model.
47
+ """
48
+
49
+ backbone_cls = GPTNeoXBackbone
50
+ preprocessor_cls = GPTNeoXCausalLMPreprocessor
51
+
52
+ def __init__(
53
+ self,
54
+ backbone,
55
+ preprocessor=None,
56
+ **kwargs,
57
+ ):
58
+ # === Layers ===
59
+ self.backbone = backbone
60
+ self.preprocessor = preprocessor
61
+
62
+ # === Functional Model ===
63
+ inputs = backbone.input
64
+ hidden_states = backbone(inputs)
65
+ outputs = backbone.token_embedding(hidden_states, reverse=True)
66
+ super().__init__(
67
+ inputs=inputs,
68
+ outputs=outputs,
69
+ **kwargs,
70
+ )
71
+
72
+ def call_with_cache(
73
+ self,
74
+ token_ids,
75
+ cache,
76
+ cache_update_index,
77
+ ):
78
+ """Forward pass of `GPTNeoXCausalLM` with cache.
79
+
80
+ `call_with_cache` adds an additional forward pass for the model for
81
+ autoregressive inference. Unlike calling the model directly, this method
82
+ allows caching previous key/value Tensors in multi-head attention layer,
83
+ and avoids recomputing the outputs of seen tokens.
84
+
85
+ Args:
86
+ token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
87
+ cache: a dense float Tensor, the cache of key and value.
88
+ cache_update_index: int, or int Tensor. The index of current inputs
89
+ in the whole sequence.
90
+
91
+ Returns:
92
+ A (logits, hidden_states, cache) tuple. Where `logits` is the
93
+ language model logits for the input token_ids, `hidden_states` is
94
+ the final hidden representation of the input tokens, and `cache` is
95
+ the decoding cache.
96
+ """
97
+ token_embedding = self.backbone.token_embedding(token_ids)
98
+ x = self.backbone.embeddings_dropout(token_embedding)
99
+ # Each decoder layer has a cache; we update them separately.
100
+ caches = []
101
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
102
+ current_cache = cache[:, i, ...]
103
+ x, next_cache = transformer_layer(
104
+ x,
105
+ self_attention_cache=current_cache,
106
+ self_attention_cache_update_index=cache_update_index,
107
+ )
108
+ caches.append(next_cache)
109
+ cache = ops.stack(caches, axis=1)
110
+ x = self.backbone.layer_norm(x)
111
+ hidden_states = x
112
+ logits = self.backbone.token_embedding(hidden_states, reverse=True)
113
+ return logits, hidden_states, cache
114
+
115
+ def _build_cache(self, token_ids):
116
+ """Build an empty cache for use with `call_with_cache()`."""
117
+ batch_size = ops.shape(token_ids)[0]
118
+ max_length = ops.shape(token_ids)[1]
119
+ num_layers = self.backbone.num_layers
120
+ num_heads = self.backbone.num_heads
121
+ head_dim = self.backbone.hidden_dim // self.backbone.num_heads
122
+ shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
123
+ cache = ops.zeros(shape, dtype=self.compute_dtype)
124
+ # Seed the cache.
125
+ _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
126
+ return hidden_states, cache
127
+
128
+ def generate_step(
129
+ self,
130
+ inputs,
131
+ stop_token_ids=None,
132
+ ):
133
+ """A compilable generation function for a single batch of inputs.
134
+
135
+ This function represents the inner, XLA-compilable, generation function
136
+ for a single batch of inputs. Inputs should have the same structure as
137
+ model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
138
+
139
+ Args:
140
+ inputs: A dictionary with two keys `"token_ids"` and
141
+ `"padding_mask"` and batched tensor values.
142
+ stop_token_ids: Tuple of id's of end token's to stop on. If all
143
+ sequences have produced a new stop token, generation
144
+ will stop.
145
+ """
146
+ token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
147
+ # Create and seed cache with a single forward pass.
148
+ hidden_states, cache = self._build_cache(token_ids)
149
+ # Compute the lengths of all user inputted tokens ids.
150
+ row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
151
+ # Start at the first index that has no user inputted id.
152
+ index = ops.min(row_lengths)
153
+
154
+ def next(prompt, cache, index):
155
+ # The cache index is the index of our previous token.
156
+ cache_update_index = index - 1
157
+ batch_size = ops.shape(prompt)[0]
158
+ prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
159
+ logits, hidden_states, cache = self.call_with_cache(
160
+ prompt,
161
+ cache,
162
+ cache_update_index,
163
+ )
164
+ return (
165
+ ops.squeeze(logits, axis=1),
166
+ ops.squeeze(hidden_states, axis=1),
167
+ cache,
168
+ )
169
+
170
+ token_ids = self.sampler(
171
+ next=next,
172
+ prompt=token_ids,
173
+ cache=cache,
174
+ index=index,
175
+ mask=padding_mask,
176
+ stop_token_ids=stop_token_ids,
177
+ hidden_states=hidden_states,
178
+ model=self,
179
+ )
180
+
181
+ # Compute an output padding mask with the token ids we updated.
182
+ if stop_token_ids is not None:
183
+ # Build a mask of stop_tokens locations not in the original
184
+ # prompt (not in locations where `padding_mask` is True).
185
+ end_locations = any_equal(
186
+ token_ids, stop_token_ids, ops.logical_not(padding_mask)
187
+ )
188
+
189
+ end_locations = ops.cast(end_locations, "int32")
190
+ # Use cumsum to get ones in all locations after end_locations.
191
+ cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
192
+ overflow = cumsum - end_locations
193
+ # Our padding mask is the inverse of these overflow locations.
194
+ padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
195
+ else:
196
+ # Without early stopping, all locations will have been updated.
197
+ padding_mask = ops.ones_like(token_ids, dtype="bool")
198
+ return {
199
+ "token_ids": token_ids,
200
+ "padding_mask": padding_mask,
201
+ }
@@ -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
+ 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.gpt_neo_x.gpt_neo_x_preprocessor import (
20
+ GPTNeoXPreprocessor,
21
+ )
22
+ from keras_hub.src.utils.keras_utils import (
23
+ convert_inputs_to_list_of_tensor_segments,
24
+ )
25
+ from keras_hub.src.utils.tensor_utils import strip_to_ragged
26
+
27
+
28
+ @keras_hub_export("keras_hub.models.GPTNeoXCausalLMPreprocessor")
29
+ class GPTNeoXCausalLMPreprocessor(GPTNeoXPreprocessor):
30
+ """GPT-NeoX Causal LM preprocessor.
31
+
32
+ This preprocessing layer is meant for use with
33
+ `keras_hub.models.GPTNeoXCausalLM`. By default, it will take in batches of
34
+ strings, and return outputs in a `(x, y, sample_weight)` format, where the
35
+ `y` label is the next token id in the `x` sequence.
36
+
37
+ For use with generation, the layer also exposes two methods
38
+ `generate_preprocess()` and `generate_postprocess()`. When this preprocessor
39
+ is attached to a `keras_hub.models.GPTNeoXCausalLM` instance, these methods
40
+ will be called implicitly in `generate()`. They can also be called
41
+ standalone (e.g. to precompute preprocessing inputs for generation in a
42
+ separate process).
43
+
44
+ Args:
45
+ tokenizer: A `keras_hub.models.GPTNeoXTokenizer` instance.
46
+ sequence_length: The length of the packed inputs.
47
+ add_start_token: If `True`, the preprocessor will prepend the tokenizer
48
+ start token to each input sequence.
49
+ add_end_token: If `True`, the preprocessor will append the tokenizer
50
+ end token to each input sequence.
51
+
52
+ Call arguments:
53
+ x: A string, `tf.Tensor` or list of python strings.
54
+ y: Label data. Should always be `None` as the layer generates labels.
55
+ sample_weight: Label weights. Should always be `None` as the layer
56
+ generates label weights.
57
+ sequence_length: Pass to override the configured `sequence_length` of
58
+ the layer.
59
+
60
+ """
61
+
62
+ def call(
63
+ self,
64
+ x,
65
+ y=None,
66
+ sample_weight=None,
67
+ sequence_length=None,
68
+ ):
69
+ if y is not None or sample_weight is not None:
70
+ logging.warning(
71
+ "`GPTNeoXCausalLMPreprocessor` generates `y` and `sample_weight` "
72
+ "based on your input data, but your data already contains `y` "
73
+ "or `sample_weight`. Your `y` and `sample_weight` will be "
74
+ "ignored."
75
+ )
76
+ sequence_length = sequence_length or self.sequence_length
77
+
78
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
79
+ x = self.tokenizer(x)
80
+ # Pad with one extra token to account for the truncation below.
81
+ token_ids, padding_mask = self.packer(
82
+ x,
83
+ sequence_length=sequence_length + 1,
84
+ add_start_value=self.add_start_token,
85
+ add_end_value=self.add_end_token,
86
+ )
87
+ # The last token does not have a next token, so we truncate it out.
88
+ x = {
89
+ "token_ids": token_ids[..., :-1],
90
+ "padding_mask": padding_mask[..., :-1],
91
+ }
92
+ # Target `y` will be the next token.
93
+ y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
94
+ return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
95
+
96
+ def generate_preprocess(
97
+ self,
98
+ x,
99
+ sequence_length=None,
100
+ ):
101
+ """Convert strings to integer token input for generation.
102
+
103
+ Similar to calling the layer for training, this method takes in strings
104
+ or tensor strings, tokenizes and packs the input, and computes a padding
105
+ mask masking all inputs not filled in with a padded value.
106
+
107
+ Unlike calling the layer for training, this method does not compute
108
+ labels and will never append a `tokenizer.end_token_id` to the end of
109
+ the sequence (as generation is expected to continue at the end of the
110
+ inputted prompt).
111
+ """
112
+ if not self.built:
113
+ self.build(None)
114
+
115
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
116
+ x = self.tokenizer(x)
117
+ token_ids, padding_mask = self.packer(
118
+ x, sequence_length=sequence_length, add_end_value=False
119
+ )
120
+ return {
121
+ "token_ids": token_ids,
122
+ "padding_mask": padding_mask,
123
+ }
124
+
125
+ def generate_postprocess(
126
+ self,
127
+ x,
128
+ ):
129
+ """Convert integer token output to strings for generation.
130
+
131
+ This method reverses `generate_preprocess()`, by first removing all
132
+ padding and start/end tokens, and then converting the integer sequence
133
+ back to a string.
134
+ """
135
+ if not self.built:
136
+ self.build(None)
137
+
138
+ token_ids, padding_mask = x["token_ids"], x["padding_mask"]
139
+ ids_to_strip = (self.tokenizer.end_token_id,)
140
+ token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
141
+ return self.tokenizer.detokenize(token_ids)