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 +1,291 @@
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.falcon.falcon_backbone import FalconBackbone
20
+ from keras_hub.src.models.falcon.falcon_causal_lm_preprocessor import (
21
+ FalconCausalLMPreprocessor,
22
+ )
23
+ from keras_hub.src.utils.tensor_utils import any_equal
24
+
25
+
26
+ @keras_hub_export("keras_hub.models.FalconCausalLM")
27
+ class FalconCausalLM(CausalLM):
28
+ """An end-to-end Falcon 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 Falcon 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, `"greedy"` sampling will be used.
41
+
42
+ This model can optionally be configured with a `preprocessor` layer, in
43
+ which case it will automatically apply preprocessing to string inputs during
44
+ `fit()`, `predict()`, `evaluate()` and `generate()`. This is done by default
45
+ when creating the model with `from_preset()`.
46
+
47
+ Args:
48
+ backbone: A `keras_hub.models.FalconBackbone` instance.
49
+ preprocessor: A `keras_hub.models.FalconCausalLMPreprocessor` or `None`.
50
+ If `None`, this model will not apply preprocessing, and inputs
51
+ should be preprocessed before calling the model.
52
+
53
+ Examples:
54
+
55
+ Use `generate()` to do text generation.
56
+ ```python
57
+ falcon_lm = keras_hub.models.FalconCausalLM.from_preset("falcon_refinedweb_1b_en")
58
+ falcon_lm.generate("I want to say", max_length=30)
59
+
60
+ # Generate with batched prompts.
61
+ falcon_lm.generate(["This is a", "Where are you"], max_length=30)
62
+ ```
63
+
64
+ Compile the `generate()` function with a custom sampler.
65
+ ```python
66
+ falcon_lm = keras_hub.models.FalconCausalLM.from_preset("falcon_refinedweb_1b_en")
67
+ falcon_lm.compile(sampler="top_k")
68
+ falcon_lm.generate("I want to say", max_length=30)
69
+
70
+ falcon_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
71
+ falcon_lm.generate("I want to say", max_length=30)
72
+ ```
73
+
74
+ Use `generate()` without preprocessing.
75
+ ```python
76
+ prompt = {
77
+ # Token ids for "<|endoftext|> Keras is".
78
+ "token_ids": np.array([[50256, 17337, 292, 318]] * 2),
79
+ # Use `"padding_mask"` to indicate values that should not be overridden.
80
+ "padding_mask": np.array([[1, 1, 1, 1]] * 2),
81
+ }
82
+
83
+ falcon_lm = keras_hub.models.FalconCausalLM.from_preset(
84
+ "falcon_refinedweb_1b_en",
85
+ preprocessor=None,
86
+ )
87
+ falcon_lm.generate(prompt)
88
+ ```
89
+
90
+ Call `fit()` on a single batch.
91
+ ```python
92
+ features = ["The quick brown fox jumped.", "I forgot my homework."]
93
+ falcon_lm = keras_hub.models.FalconCausalLM.from_preset("falcon_refinedweb_1b_en")
94
+ falcon_lm.fit(x=features, batch_size=2)
95
+ ```
96
+
97
+ Call `fit()` without preprocessing.
98
+ ```python
99
+ x = {
100
+ # Token ids for "<|endoftext|> Keras is deep learning library<|endoftext|>"
101
+ "token_ids": np.array([[50256, 17337, 292, 318, 2769, 4673, 5888, 50256, 0]] * 2),
102
+ "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 0]] * 2),
103
+ }
104
+ y = np.array([[17337, 292, 318, 2769, 4673, 5888, 50256, 0, 0]] * 2)
105
+ sw = np.array([[1, 1, 1, 1, 1, 1, 1, 0, 0]] * 2)
106
+
107
+ falcon_lm = keras_hub.models.FalconCausalLM.from_preset(
108
+ "falcon_refinedweb_1b_en",
109
+ preprocessor=None,
110
+ )
111
+ falcon_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)
112
+ ```
113
+
114
+ Custom backbone and vocabulary.
115
+ ```python
116
+ vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
117
+ merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
118
+ merges += ["Ġ f", "o x", "Ġf ox"]
119
+ tokenizer = keras_hub.models.FalconTokenizer(
120
+ vocabulary=vocab,
121
+ merges=merges,
122
+ )
123
+ preprocessor = keras_hub.models.FalconCausalLMPreprocessor(
124
+ tokenizer=tokenizer,
125
+ sequence_length=128,
126
+ )
127
+ backbone = keras_hub.models.FalconBackbone(
128
+ vocabulary_size=50304,
129
+ num_layers=24,
130
+ num_attention_heads=64,
131
+ hidden_dim=2048,
132
+ intermediate_dim=4*2048,
133
+ )
134
+ falcon_lm = keras_hub.models.FalconCausalLM(
135
+ backbone=backbone,
136
+ preprocessor=preprocessor,
137
+ )
138
+ falcon_lm.fit(x=features, batch_size=2)
139
+ ```
140
+ """
141
+
142
+ backbone_cls = FalconBackbone
143
+ preprocessor_cls = FalconCausalLMPreprocessor
144
+
145
+ def __init__(
146
+ self,
147
+ backbone,
148
+ preprocessor=None,
149
+ **kwargs,
150
+ ):
151
+ # === Layers ===
152
+ self.backbone = backbone
153
+ self.preprocessor = preprocessor
154
+
155
+ # === Functional Model ===
156
+ inputs = backbone.input
157
+ hidden_states = backbone(inputs)
158
+ outputs = backbone.token_embedding(hidden_states, reverse=True)
159
+ super().__init__(
160
+ inputs=inputs,
161
+ outputs=outputs,
162
+ **kwargs,
163
+ )
164
+
165
+ def call_with_cache(
166
+ self,
167
+ token_ids,
168
+ cache,
169
+ cache_update_index,
170
+ ):
171
+ """Forward pass of `FalconCausalLM` with cache.
172
+
173
+ `call_with_cache` adds an additional forward pass for the model for
174
+ autoregressive inference. Unlike calling the model directly, this method
175
+ allows caching previous key/value Tensors in multi-head attention layer,
176
+ and avoids recomputing the outputs of seen tokens.
177
+
178
+ Args:
179
+ token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
180
+ cache: a dense float Tensor, the cache of key and value.
181
+ cache_update_index: int, or int Tensor. The index of current inputs in the
182
+ whole sequence.
183
+
184
+ Returns:
185
+ A (logits, hidden_states, cache) tuple. Where `logits` is the
186
+ language model logits for the input token_ids, `hidden_states` is
187
+ the final hidden representation of the input tokens, and `cache` is
188
+ the decoding cache.
189
+ """
190
+ x = self.backbone.token_embedding(token_ids)
191
+ # Each decoder layer has a cache; we update them separately.
192
+ caches = []
193
+ for i, transformer_layer in enumerate(self.backbone.transformer_layers):
194
+ current_cache = cache[:, i, ...]
195
+ x, next_cache = transformer_layer(
196
+ x,
197
+ attention_cache=current_cache,
198
+ attention_cache_update_index=cache_update_index,
199
+ )
200
+ caches.append(next_cache)
201
+ cache = ops.stack(caches, axis=1)
202
+ hidden_states = x = self.backbone.final_layernorm(x)
203
+ logits = self.backbone.token_embedding(x, reverse=True)
204
+ return logits, hidden_states, cache
205
+
206
+ def _build_cache(self, token_ids):
207
+ """Build an empty cache for use with `call_with_cache()`."""
208
+ batch_size = ops.shape(token_ids)[0]
209
+ max_length = ops.shape(token_ids)[1]
210
+ num_layers = self.backbone.num_layers
211
+ num_heads = self.backbone.num_attention_heads
212
+ head_dim = self.backbone.hidden_dim // self.backbone.num_attention_heads
213
+ shape = [batch_size, num_layers, 2, max_length, num_heads, head_dim]
214
+ cache = ops.zeros(shape, dtype=self.compute_dtype)
215
+ # Seed the cache.
216
+ _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
217
+ return hidden_states, cache
218
+
219
+ def generate_step(
220
+ self,
221
+ inputs,
222
+ stop_token_ids=None,
223
+ ):
224
+ """A compilable generation function for a single batch of inputs.
225
+
226
+ This function represents the inner, XLA-compilable, generation function
227
+ for a single batch of inputs. Inputs should have the same structure as
228
+ model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
229
+
230
+ Args:
231
+ inputs: A dictionary with two keys `"token_ids"` and
232
+ `"padding_mask"` and batched tensor values.
233
+ stop_token_ids: Tuple of id's of end token's to stop on. If all
234
+ sequences have produced a new stop token, generation
235
+ will stop.
236
+ """
237
+ token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
238
+ # Create and seed cache with a single forward pass.
239
+ hidden_states, cache = self._build_cache(token_ids)
240
+ # Compute the lengths of all user inputted tokens ids.
241
+ row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
242
+ # Start at the first index that has no user inputted id.
243
+ index = ops.min(row_lengths)
244
+
245
+ def next(prompt, cache, index):
246
+ # The cache index is the index of our previous token.
247
+ cache_update_index = index - 1
248
+ batch_size = ops.shape(prompt)[0]
249
+ prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
250
+ logits, hidden_states, cache = self.call_with_cache(
251
+ prompt,
252
+ cache,
253
+ cache_update_index,
254
+ )
255
+ return (
256
+ ops.squeeze(logits, axis=1),
257
+ ops.squeeze(hidden_states, axis=1),
258
+ cache,
259
+ )
260
+
261
+ token_ids = self.sampler(
262
+ next=next,
263
+ prompt=token_ids,
264
+ cache=cache,
265
+ index=index,
266
+ mask=padding_mask,
267
+ stop_token_ids=stop_token_ids,
268
+ hidden_states=hidden_states,
269
+ model=self,
270
+ )
271
+
272
+ # Compute an output padding mask with the token ids we updated.
273
+ if stop_token_ids is not None:
274
+ # Build a mask of stop token locations not in the original
275
+ # prompt (not in locations where `padding_mask` is True).
276
+ end_locations = any_equal(
277
+ token_ids, stop_token_ids, ops.logical_not(padding_mask)
278
+ )
279
+ end_locations = ops.cast(end_locations, "int32")
280
+ # Use cumsum to get ones in all locations after end_locations.
281
+ cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
282
+ overflow = cumsum - end_locations
283
+ # Our padding mask is the inverse of these overflow locations.
284
+ padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
285
+ else:
286
+ # Without early stopping, all locations will have been updated.
287
+ padding_mask = ops.ones_like(token_ids, dtype="bool")
288
+ return {
289
+ "token_ids": token_ids,
290
+ "padding_mask": padding_mask,
291
+ }
@@ -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
+ 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.falcon.falcon_preprocessor import FalconPreprocessor
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.FalconCausalLMPreprocessor")
27
+ class FalconCausalLMPreprocessor(FalconPreprocessor):
28
+ """Falcon Causal LM preprocessor.
29
+
30
+ This preprocessing layer is meant for use with
31
+ `keras_hub.models.FalconCausalLM`. 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.FalconCausalLM` 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.FalconTokenizer` instance.
44
+ sequence_length: The length of the packed inputs.
45
+ add_start_token: If `True`, the preprocessor will prepend the tokenizer
46
+ start token to each input sequence.
47
+ add_end_token: If `True`, the preprocessor will append the tokenizer
48
+ end token to each input sequence.
49
+
50
+ Call arguments:
51
+ x: A string, `tf.Tensor` or list of python strings.
52
+ y: Label data. Should always be `None` as the layer generates labels.
53
+ sample_weight: Label weights. Should always be `None` as the layer
54
+ generates label weights.
55
+ sequence_length: Pass to override the configured `sequence_length` of
56
+ the layer.
57
+
58
+ Examples:
59
+ ```python
60
+ # Load the preprocessor from a preset.
61
+ preprocessor = keras_hub.models.FalconCausalLMPreprocessor.from_preset(
62
+ "falcon_refinedweb_1b_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
+ "`FalconCausalLMPreprocessor` generates `y` and `sample_weight` "
104
+ "based on your input data, but your data already contains `y` "
105
+ "or `sample_weight`. Your `y` and `sample_weight` will be "
106
+ "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
+ if not self.built:
168
+ self.build(None)
169
+
170
+ token_ids, padding_mask = x["token_ids"], x["padding_mask"]
171
+ ids_to_strip = (self.tokenizer.end_token_id,)
172
+ token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
173
+ return self.tokenizer.detokenize(token_ids)
@@ -0,0 +1,187 @@
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.preprocessing.start_end_packer import StartEndPacker
20
+ from keras_hub.src.models.falcon.falcon_tokenizer import FalconTokenizer
21
+ from keras_hub.src.models.preprocessor import Preprocessor
22
+ from keras_hub.src.utils.keras_utils import (
23
+ convert_inputs_to_list_of_tensor_segments,
24
+ )
25
+
26
+
27
+ @keras_hub_export("keras_hub.models.FalconPreprocessor")
28
+ class FalconPreprocessor(Preprocessor):
29
+ """Falcon preprocessing layer which tokenizes and packs inputs.
30
+
31
+ This preprocessing layer will do 2 things:
32
+
33
+ - Tokenize the inputs using the `tokenizer`.
34
+ - Construct a dictionary with keys `"token_ids"`, `"padding_mask"`, that can
35
+ be passed directly to a `keras_hub.models.FalconBackbone`.
36
+
37
+ This layer can be used directly with `tf.data.Dataset.map` to preprocess
38
+ string data in the `(x, y, sample_weight)` format used by
39
+ `keras.Model.fit`.
40
+
41
+ The call method of this layer accepts three arguments, `x`, `y`, and
42
+ `sample_weight`. `x` can be a python string or tensor representing a single
43
+ segment, a list of python strings representing a batch of single segments,
44
+ or a list of tensors representing multiple segments to be packed together.
45
+ `y` and `sample_weight` are both optional, can have any format, and will be
46
+ passed through unaltered.
47
+
48
+ `FalconPreprocessor` forces the input to have only one segment, as Falcon is
49
+ mainly used for generation tasks. For tasks having multi-segment inputs
50
+ like "glue/mnli", please use a model designed for classification purposes
51
+ such as BERT or RoBERTa.
52
+
53
+ Args:
54
+ tokenizer: A `keras_hub.models.FalconTokenizer` instance.
55
+ sequence_length: The length of the packed inputs.
56
+ add_start_token: If `True`, the preprocessor will prepend the tokenizer
57
+ start token to each input sequence.
58
+ add_end_token: If `True`, the preprocessor will append the tokenizer
59
+ end token to each input sequence.
60
+
61
+ Call arguments:
62
+ x: A string, `tf.Tensor` or list of python strings.
63
+ y: Any label data. Will be passed through unaltered.
64
+ sample_weight: Any label weight data. Will be passed through unaltered.
65
+ sequence_length: Pass to override the configured `sequence_length` of
66
+ the layer.
67
+
68
+ Examples:
69
+
70
+ Directly calling the layer on data.
71
+ ```python
72
+ preprocessor = keras_hub.models.FalconPreprocessor.from_preset("falcon_rw_1b")
73
+
74
+ # Tokenize and pack a single sentence.
75
+ preprocessor("The quick brown fox jumped.")
76
+
77
+ # Tokenize a batch of single sentences.
78
+ preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
79
+
80
+ # Custom vocabulary.
81
+ features = ["a quick fox.", "a fox quick."]
82
+ vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
83
+ merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
84
+ merges += ["Ġ f", "o x", "Ġf ox"]
85
+ tokenizer = keras_hub.models.FalconTokenizer(
86
+ vocabulary=vocab,
87
+ merges=merges,
88
+ )
89
+ preprocessor = keras_hub.models.FalconPreprocessor(tokenizer=tokenizer)
90
+ preprocessor("The quick brown fox jumped.")
91
+ ```
92
+
93
+ Mapping with `tf.data.Dataset`.
94
+ ```python
95
+ preprocessor = keras_hub.models.FalconPreprocessor.from_preset("falcon_rw_1b")
96
+
97
+ text = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
98
+ label = tf.constant([1, 1])
99
+
100
+ # Map labeled single sentences.
101
+ ds = tf.data.Dataset.from_tensor_slices((text, label))
102
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
103
+
104
+ # Map unlabeled single sentences.
105
+ ds = tf.data.Dataset.from_tensor_slices(text)
106
+ ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
107
+ ```
108
+ """
109
+
110
+ tokenizer_cls = FalconTokenizer
111
+
112
+ def __init__(
113
+ self,
114
+ tokenizer,
115
+ sequence_length=1024,
116
+ add_start_token=True,
117
+ add_end_token=True,
118
+ **kwargs,
119
+ ):
120
+ super().__init__(**kwargs)
121
+ self.tokenizer = tokenizer
122
+ self.packer = None
123
+ self.sequence_length = sequence_length
124
+ self.add_start_token = add_start_token
125
+ self.add_end_token = add_end_token
126
+
127
+ def build(self, input_shape):
128
+ # Defer packer creation to `build()` so that we can be sure tokenizer
129
+ # assets have loaded when restoring a saved model.
130
+ self.packer = StartEndPacker(
131
+ start_value=self.tokenizer.start_token_id,
132
+ end_value=self.tokenizer.end_token_id,
133
+ pad_value=self.tokenizer.pad_token_id,
134
+ sequence_length=self.sequence_length,
135
+ return_padding_mask=True,
136
+ )
137
+ self.built = True
138
+
139
+ def call(
140
+ self,
141
+ x,
142
+ y=None,
143
+ sample_weight=None,
144
+ sequence_length=None,
145
+ ):
146
+ x = convert_inputs_to_list_of_tensor_segments(x)
147
+ if len(x) != 1:
148
+ raise ValueError(
149
+ "Falcon requires each input feature to contain only "
150
+ f"one segment, but received {len(x)}. If you are using Falcon "
151
+ "for a multi-segment classification task, please refer to "
152
+ "classification models like BERT or RoBERTa."
153
+ )
154
+ sequence_length = sequence_length or self.sequence_length
155
+ token_ids, padding_mask = self.packer(
156
+ self.tokenizer(x[0]),
157
+ sequence_length=sequence_length,
158
+ add_start_value=self.add_start_token,
159
+ add_end_value=self.add_end_token,
160
+ )
161
+ x = {
162
+ "token_ids": token_ids,
163
+ "padding_mask": padding_mask,
164
+ }
165
+ return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
166
+
167
+ def get_config(self):
168
+ config = super().get_config()
169
+ config.update(
170
+ {
171
+ "sequence_length": self.sequence_length,
172
+ "add_start_token": self.add_start_token,
173
+ "add_end_token": self.add_end_token,
174
+ }
175
+ )
176
+ return config
177
+
178
+ @property
179
+ def sequence_length(self):
180
+ """The padded length of model input sequences."""
181
+ return self._sequence_length
182
+
183
+ @sequence_length.setter
184
+ def sequence_length(self, value):
185
+ self._sequence_length = value
186
+ if self.packer is not None:
187
+ self.packer.sequence_length = value
@@ -0,0 +1,30 @@
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
+ """Falcon model preset configurations."""
15
+
16
+ backbone_presets = {
17
+ "falcon_refinedweb_1b_en": {
18
+ "metadata": {
19
+ "description": (
20
+ "24-layer Falcon model (Falcon with 1B parameters), trained on "
21
+ "350B tokens of RefinedWeb dataset."
22
+ ),
23
+ "params": 1311625216,
24
+ "official_name": "Falcon",
25
+ "path": "falcon",
26
+ "model_card": "https://huggingface.co/tiiuae/falcon-rw-1b",
27
+ },
28
+ "kaggle_handle": "kaggle://keras/falcon/keras/falcon_refinedweb_1b_en/1",
29
+ },
30
+ }