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 +1,327 @@
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
+ import keras
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.llama.llama_backbone import LlamaBackbone
20
+ from keras_hub.src.models.llama.llama_causal_lm_preprocessor import (
21
+ LlamaCausalLMPreprocessor,
22
+ )
23
+ from keras_hub.src.utils.tensor_utils import any_equal
24
+
25
+
26
+ @keras_hub_export("keras_hub.models.LlamaCausalLM")
27
+ class LlamaCausalLM(CausalLM):
28
+ """An end-to-end Llama 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 LLaMA 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.LlamaBackbone` instance.
44
+ preprocessor: A `keras_hub.models.LlamaCausalLMPreprocessor` 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 = LlamaBackbone
50
+ preprocessor_cls = LlamaCausalLMPreprocessor
51
+
52
+ def __init__(self, backbone, preprocessor=None, **kwargs):
53
+ # === Layers ===
54
+ self.backbone = backbone
55
+ self.preprocessor = preprocessor
56
+
57
+ # === Functional Model ===
58
+ inputs = backbone.inputs
59
+ hidden_states = backbone(inputs)
60
+ outputs = backbone.token_embedding(hidden_states, reverse=True)
61
+ super().__init__(
62
+ inputs=inputs,
63
+ outputs=outputs,
64
+ **kwargs,
65
+ )
66
+
67
+ def call_with_cache(
68
+ self,
69
+ token_ids,
70
+ cache,
71
+ cache_update_index,
72
+ ):
73
+ """Forward pass of `LlamaCausalLM` with cache.
74
+
75
+ `call_with_cache` adds an additional forward pass for the model for
76
+ autoregressive inference. Unlike calling the model directly, this method
77
+ allows caching previous key/value Tensors in multi-head attention layer,
78
+ and avoids recomputing the outputs of seen tokens.
79
+
80
+ Args:
81
+ token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
82
+ cache: a dense float Tensor, the cache of key and value.
83
+ cache_update_index: int, or int Tensor. The index of current inputs
84
+ in the whole sequence.
85
+
86
+ Returns:
87
+ A (logits, hidden_states, cache) tuple. Where `logits` is the
88
+ language model logits for the input token_ids, `hidden_states` is
89
+ the final hidden representation of the input tokens, and `cache` is
90
+ the decoding cache.
91
+ """
92
+ x = self.backbone.token_embedding(token_ids)
93
+ # Each decoder layer has a cache; we update them separately.
94
+ updated_cache = []
95
+ for i in range(self.backbone.num_layers):
96
+ current_cache = cache[:, i, ...]
97
+ x, next_cache = self.backbone.transformer_layers[i](
98
+ x,
99
+ self_attention_cache=current_cache,
100
+ self_attention_cache_update_index=cache_update_index,
101
+ )
102
+ updated_cache.append(next_cache)
103
+ cache = ops.stack(updated_cache, axis=1)
104
+ hidden_states = x = self.backbone.layer_norm(x)
105
+ logits = self.backbone.token_embedding(x, reverse=True)
106
+ return logits, hidden_states, cache
107
+
108
+ def _build_cache(self, token_ids):
109
+ """Build an empty cache for use with `call_with_cache()`."""
110
+ batch_size = ops.shape(token_ids)[0]
111
+ max_length = ops.shape(token_ids)[1]
112
+ num_layers = self.backbone.num_layers
113
+ num_key_value_heads = self.backbone.num_key_value_heads
114
+ head_dim = self.backbone.hidden_dim // self.backbone.num_query_heads
115
+ shape = [
116
+ batch_size,
117
+ num_layers,
118
+ 2,
119
+ max_length,
120
+ num_key_value_heads,
121
+ head_dim,
122
+ ]
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 the end token 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 token 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
+ end_locations = ops.cast(end_locations, "int32")
189
+ # Use cumsum to get ones in all locations after end_locations.
190
+ cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
191
+ overflow = cumsum - end_locations
192
+ # Our padding mask is the inverse of these overflow locations.
193
+ padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
194
+ else:
195
+ # Without early stopping, all locations will have been updated.
196
+ padding_mask = ops.ones_like(token_ids, dtype="bool")
197
+ return {
198
+ "token_ids": token_ids,
199
+ "padding_mask": padding_mask,
200
+ }
201
+
202
+ def score(
203
+ self,
204
+ token_ids,
205
+ padding_mask=None,
206
+ scoring_mode="logits",
207
+ layer_intercept_fn=None,
208
+ target_ids=None,
209
+ ):
210
+ """Score a generation represented by the provided token ids.
211
+
212
+ Args:
213
+ token_ids: A <int>[batch_size, num_tokens] tensor containing tokens
214
+ to score. Typically, this tensor captures the output from a call
215
+ to `LlamaCausalLM.generate()`, i.e., tokens for both the input
216
+ text and the model-generated text.
217
+ padding_mask: A <bool>[batch_size, num_tokens] tensor indicating the
218
+ tokens that should be preserved during generation. This is an
219
+ artifact required by the `LlamaBackbone` and isn't influential
220
+ on the computation of this function. If omitted, this function
221
+ uses `keras.ops.ones()` to create a tensor of the appropriate
222
+ shape.
223
+ scoring_mode: The type of scores to return, either "logits" or
224
+ "loss", both will be per input token.
225
+ layer_intercept_fn: An optional function for augmenting activations
226
+ with additional computation, for example, as part of
227
+ interpretability research. This function will be passed the
228
+ activations as its first parameter and a numeric index
229
+ associated with that backbone layer. _This index _is not_ an
230
+ index into `self.backbone.layers`_. The index -1 accompanies the
231
+ embeddings returned by calling `self.backbone.token_embedding()`
232
+ on `token_ids` in the forward direction. All subsequent indexes
233
+ will be 0-based indices for the activations returned by each of
234
+ the Transformers layers in the backbone. This function must
235
+ return a <float>[batch_size, num_tokens, hidden_dims] tensor
236
+ that can be passed as an input to the next layer in the model.
237
+ target_ids: An <bool>[batch_size, num_tokens] tensor containing the
238
+ predicted tokens against which the loss should be computed. If a
239
+ span of tokens is provided (sequential truthy values along
240
+ axis=1 in the tensor), the loss will be computed as the
241
+ aggregate across those tokens.
242
+
243
+ Raises:
244
+ ValueError: If an unsupported scoring_mode is provided, or if the
245
+ target_ids are not provided when using ScoringMode.LOSS.
246
+
247
+ Returns:
248
+ The per-token scores as a tensor of size
249
+ <float>[batch_size, num_tokens, vocab_size] in "logits" mode, or
250
+ <float>[batch_size, num_tokens] in "loss" mode.
251
+
252
+ Example:
253
+
254
+ Compute gradients between embeddings and loss scores with TensorFlow:
255
+ ```python
256
+ llama_lm = keras_hub.models.LlamaCausalLM.from_preset("llama2_7b_en")
257
+ generations = llama_lm.generate(
258
+ ["This is a", "Where are you"],
259
+ max_length=30
260
+ )
261
+ preprocessed = llama_lm.preprocessor.generate_preprocess(generations)
262
+ generation_ids = preprocessed["token_ids"]
263
+ padding_mask = preprocessed["padding_mask"]
264
+ target_ids = keras.ops.roll(generation_ids, shift=-1, axis=1)
265
+
266
+ embeddings = None
267
+ with tf.GradientTape(watch_accessed_variables=True) as tape:
268
+ def layer_intercept_fn(x, i):
269
+ if i == -1:
270
+ nonlocal embeddings, tape
271
+ embeddings = x
272
+ tape.watch(embeddings)
273
+ return x
274
+
275
+ losses = llama_lm.score(
276
+ token_ids=generation_ids,
277
+ padding_mask=padding_mask,
278
+ scoring_mode="loss",
279
+ layer_intercept_fn=layer_intercept_fn,
280
+ target_ids=target_ids,
281
+ )
282
+
283
+ grads = tape.gradient(losses, embeddings)
284
+ ```
285
+ """
286
+ if scoring_mode not in ("logits", "loss"):
287
+ raise ValueError(
288
+ "Unsupported scoring_mode. Must be one of 'logits' or 'loss'."
289
+ )
290
+
291
+ if scoring_mode == "loss" and target_ids is None:
292
+ raise ValueError(
293
+ "Cannot compute loss without targets. Please provide target "
294
+ "token ids via the target_ids parameter."
295
+ )
296
+
297
+ batch_shape = ops.shape(token_ids)[:2]
298
+ assert len(batch_shape) == 2
299
+
300
+ if padding_mask is None:
301
+ padding_mask = ops.ones(shape=batch_shape)
302
+
303
+ if layer_intercept_fn is None:
304
+
305
+ def default_layer_intercept_fn(x, unused_i):
306
+ return x
307
+
308
+ layer_intercept_fn = default_layer_intercept_fn
309
+
310
+ token_embeddings = self.backbone.token_embedding(token_ids)
311
+ x = layer_intercept_fn(token_embeddings, -1)
312
+
313
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
314
+ x = transformer_layer(x, decoder_padding_mask=padding_mask)
315
+ x = layer_intercept_fn(x, i)
316
+
317
+ x = self.backbone.layer_norm(x)
318
+ logits = self.backbone.token_embedding(x, reverse=True)
319
+
320
+ if scoring_mode == "logits":
321
+ return logits
322
+
323
+ per_token_loss_fn = keras.losses.SparseCategoricalCrossentropy(
324
+ from_logits=True, reduction="none"
325
+ )
326
+ per_token_loss = per_token_loss_fn(target_ids, logits)
327
+ return per_token_loss
@@ -0,0 +1,170 @@
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.llama.llama_preprocessor import LlamaPreprocessor
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.LlamaCausalLMPreprocessor")
27
+ class LlamaCausalLMPreprocessor(LlamaPreprocessor):
28
+ """Llama Causal LM preprocessor.
29
+
30
+ This preprocessing layer is meant for use with
31
+ `keras_hub.models.LlamaCausalLM`. 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.LlamaCausalLM` 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.LlamaTokenizer` 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. Default is `True`.
47
+ add_end_token: If `True`, the preprocessor will append the tokenizer
48
+ end token to each input sequence. Default is `False`.
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.LlamaCausalLMPreprocessor.from_preset(
62
+ "llama_base_en"
63
+ )
64
+
65
+ # Tokenize and pack a single sentence.
66
+ sentence = tf.constant("League of legends")
67
+ preprocessor(sentence)
68
+ # Same output.
69
+ preprocessor("League of legends")
70
+
71
+ # Tokenize a batch of sentences.
72
+ sentences = tf.constant(["Taco tuesday", "Fish taco please!"])
73
+ preprocessor(sentences)
74
+ # Same output.
75
+ preprocessor(["Taco tuesday", "Fish taco please!"])
76
+
77
+ # Map a dataset to preprocess a single sentence.
78
+ features = tf.constant(
79
+ [
80
+ "Avatar 2 is amazing!",
81
+ "Well, I am not sure.",
82
+ ]
83
+ )
84
+ labels = tf.constant([1, 0])
85
+ ds = tf.data.Dataset.from_tensor_slices((features, labels))
86
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
87
+
88
+ # Map a dataset to preprocess unlabled sentences.
89
+ ds = tf.data.Dataset.from_tensor_slices(features)
90
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
91
+ ```
92
+ """
93
+
94
+ def call(
95
+ self,
96
+ x,
97
+ y=None,
98
+ sample_weight=None,
99
+ sequence_length=None,
100
+ ):
101
+ if y is not None or sample_weight is not None:
102
+ logging.warning(
103
+ "`LlamaCausalLMPreprocessor` generates `y` and "
104
+ "`sample_weight` based on your input data, but your data "
105
+ "already contains `y` or `sample_weight`. Your `y` and "
106
+ "`sample_weight` will be ignored."
107
+ )
108
+ sequence_length = sequence_length or self.sequence_length
109
+
110
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
111
+ x = self.tokenizer(x)
112
+ # Pad with one extra token to account for the truncation below.
113
+ token_ids, padding_mask = self.packer(
114
+ x,
115
+ sequence_length=sequence_length + 1,
116
+ add_start_value=self.add_start_token,
117
+ add_end_value=self.add_end_token,
118
+ )
119
+ # The last token does not have a next token, so we truncate it out.
120
+ x = {
121
+ "token_ids": token_ids[..., :-1],
122
+ "padding_mask": padding_mask[..., :-1],
123
+ }
124
+ # Target `y` will be the next token.
125
+ y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
126
+ return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
127
+
128
+ def generate_preprocess(
129
+ self,
130
+ x,
131
+ sequence_length=None,
132
+ ):
133
+ """Convert strings to integer token input for generation.
134
+
135
+ Similar to calling the layer for training, this method takes in strings
136
+ or tensor strings, tokenizes and packs the input, and computes a padding
137
+ mask masking all inputs not filled in with a padded value.
138
+
139
+ Unlike calling the layer for training, this method does not compute
140
+ labels and will never append a `tokenizer.end_token_id` to the end of
141
+ the sequence (as generation is expected to continue at the end of the
142
+ inputted prompt).
143
+ """
144
+ if not self.built:
145
+ self.build(None)
146
+
147
+ x = convert_inputs_to_list_of_tensor_segments(x)[0]
148
+ x = self.tokenizer(x)
149
+ token_ids, padding_mask = self.packer(
150
+ x, sequence_length=sequence_length, add_end_value=False
151
+ )
152
+ return {
153
+ "token_ids": token_ids,
154
+ "padding_mask": padding_mask,
155
+ }
156
+
157
+ def generate_postprocess(
158
+ self,
159
+ x,
160
+ ):
161
+ """Convert integer token output to strings for generation.
162
+
163
+ This method reverses `generate_preprocess()`, by first removing all
164
+ padding and start/end tokens, and then converting the integer sequence
165
+ back to a string.
166
+ """
167
+ token_ids, padding_mask = x["token_ids"], x["padding_mask"]
168
+ ids_to_strip = (self.tokenizer.end_token_id,)
169
+ token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
170
+ return self.tokenizer.detokenize(token_ids)