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.
- keras_hub/__init__.py +52 -0
- keras_hub/api/__init__.py +27 -0
- keras_hub/api/layers/__init__.py +47 -0
- keras_hub/api/metrics/__init__.py +24 -0
- keras_hub/api/models/__init__.py +249 -0
- keras_hub/api/samplers/__init__.py +29 -0
- keras_hub/api/tokenizers/__init__.py +35 -0
- keras_hub/src/__init__.py +13 -0
- keras_hub/src/api_export.py +53 -0
- keras_hub/src/layers/__init__.py +13 -0
- keras_hub/src/layers/modeling/__init__.py +13 -0
- keras_hub/src/layers/modeling/alibi_bias.py +143 -0
- keras_hub/src/layers/modeling/cached_multi_head_attention.py +137 -0
- keras_hub/src/layers/modeling/f_net_encoder.py +200 -0
- keras_hub/src/layers/modeling/masked_lm_head.py +239 -0
- keras_hub/src/layers/modeling/position_embedding.py +123 -0
- keras_hub/src/layers/modeling/reversible_embedding.py +311 -0
- keras_hub/src/layers/modeling/rotary_embedding.py +169 -0
- keras_hub/src/layers/modeling/sine_position_encoding.py +108 -0
- keras_hub/src/layers/modeling/token_and_position_embedding.py +150 -0
- keras_hub/src/layers/modeling/transformer_decoder.py +496 -0
- keras_hub/src/layers/modeling/transformer_encoder.py +262 -0
- keras_hub/src/layers/modeling/transformer_layer_utils.py +106 -0
- keras_hub/src/layers/preprocessing/__init__.py +13 -0
- keras_hub/src/layers/preprocessing/masked_lm_mask_generator.py +220 -0
- keras_hub/src/layers/preprocessing/multi_segment_packer.py +319 -0
- keras_hub/src/layers/preprocessing/preprocessing_layer.py +62 -0
- keras_hub/src/layers/preprocessing/random_deletion.py +271 -0
- keras_hub/src/layers/preprocessing/random_swap.py +267 -0
- keras_hub/src/layers/preprocessing/start_end_packer.py +219 -0
- keras_hub/src/metrics/__init__.py +13 -0
- keras_hub/src/metrics/bleu.py +394 -0
- keras_hub/src/metrics/edit_distance.py +197 -0
- keras_hub/src/metrics/perplexity.py +181 -0
- keras_hub/src/metrics/rouge_base.py +204 -0
- keras_hub/src/metrics/rouge_l.py +97 -0
- keras_hub/src/metrics/rouge_n.py +125 -0
- keras_hub/src/models/__init__.py +13 -0
- keras_hub/src/models/albert/__init__.py +20 -0
- keras_hub/src/models/albert/albert_backbone.py +267 -0
- keras_hub/src/models/albert/albert_classifier.py +202 -0
- keras_hub/src/models/albert/albert_masked_lm.py +129 -0
- keras_hub/src/models/albert/albert_masked_lm_preprocessor.py +194 -0
- keras_hub/src/models/albert/albert_preprocessor.py +206 -0
- keras_hub/src/models/albert/albert_presets.py +70 -0
- keras_hub/src/models/albert/albert_tokenizer.py +119 -0
- keras_hub/src/models/backbone.py +311 -0
- keras_hub/src/models/bart/__init__.py +20 -0
- keras_hub/src/models/bart/bart_backbone.py +261 -0
- keras_hub/src/models/bart/bart_preprocessor.py +276 -0
- keras_hub/src/models/bart/bart_presets.py +74 -0
- keras_hub/src/models/bart/bart_seq_2_seq_lm.py +490 -0
- keras_hub/src/models/bart/bart_seq_2_seq_lm_preprocessor.py +262 -0
- keras_hub/src/models/bart/bart_tokenizer.py +124 -0
- keras_hub/src/models/bert/__init__.py +23 -0
- keras_hub/src/models/bert/bert_backbone.py +227 -0
- keras_hub/src/models/bert/bert_classifier.py +183 -0
- keras_hub/src/models/bert/bert_masked_lm.py +131 -0
- keras_hub/src/models/bert/bert_masked_lm_preprocessor.py +198 -0
- keras_hub/src/models/bert/bert_preprocessor.py +184 -0
- keras_hub/src/models/bert/bert_presets.py +147 -0
- keras_hub/src/models/bert/bert_tokenizer.py +112 -0
- keras_hub/src/models/bloom/__init__.py +20 -0
- keras_hub/src/models/bloom/bloom_attention.py +186 -0
- keras_hub/src/models/bloom/bloom_backbone.py +173 -0
- keras_hub/src/models/bloom/bloom_causal_lm.py +298 -0
- keras_hub/src/models/bloom/bloom_causal_lm_preprocessor.py +176 -0
- keras_hub/src/models/bloom/bloom_decoder.py +206 -0
- keras_hub/src/models/bloom/bloom_preprocessor.py +185 -0
- keras_hub/src/models/bloom/bloom_presets.py +121 -0
- keras_hub/src/models/bloom/bloom_tokenizer.py +116 -0
- keras_hub/src/models/causal_lm.py +383 -0
- keras_hub/src/models/classifier.py +109 -0
- keras_hub/src/models/csp_darknet/__init__.py +13 -0
- keras_hub/src/models/csp_darknet/csp_darknet_backbone.py +410 -0
- keras_hub/src/models/csp_darknet/csp_darknet_image_classifier.py +133 -0
- keras_hub/src/models/deberta_v3/__init__.py +24 -0
- keras_hub/src/models/deberta_v3/deberta_v3_backbone.py +210 -0
- keras_hub/src/models/deberta_v3/deberta_v3_classifier.py +228 -0
- keras_hub/src/models/deberta_v3/deberta_v3_masked_lm.py +135 -0
- keras_hub/src/models/deberta_v3/deberta_v3_masked_lm_preprocessor.py +191 -0
- keras_hub/src/models/deberta_v3/deberta_v3_preprocessor.py +206 -0
- keras_hub/src/models/deberta_v3/deberta_v3_presets.py +82 -0
- keras_hub/src/models/deberta_v3/deberta_v3_tokenizer.py +155 -0
- keras_hub/src/models/deberta_v3/disentangled_attention_encoder.py +227 -0
- keras_hub/src/models/deberta_v3/disentangled_self_attention.py +412 -0
- keras_hub/src/models/deberta_v3/relative_embedding.py +94 -0
- keras_hub/src/models/densenet/__init__.py +13 -0
- keras_hub/src/models/densenet/densenet_backbone.py +210 -0
- keras_hub/src/models/densenet/densenet_image_classifier.py +131 -0
- keras_hub/src/models/distil_bert/__init__.py +26 -0
- keras_hub/src/models/distil_bert/distil_bert_backbone.py +187 -0
- keras_hub/src/models/distil_bert/distil_bert_classifier.py +208 -0
- keras_hub/src/models/distil_bert/distil_bert_masked_lm.py +137 -0
- keras_hub/src/models/distil_bert/distil_bert_masked_lm_preprocessor.py +194 -0
- keras_hub/src/models/distil_bert/distil_bert_preprocessor.py +175 -0
- keras_hub/src/models/distil_bert/distil_bert_presets.py +57 -0
- keras_hub/src/models/distil_bert/distil_bert_tokenizer.py +114 -0
- keras_hub/src/models/electra/__init__.py +20 -0
- keras_hub/src/models/electra/electra_backbone.py +247 -0
- keras_hub/src/models/electra/electra_preprocessor.py +154 -0
- keras_hub/src/models/electra/electra_presets.py +95 -0
- keras_hub/src/models/electra/electra_tokenizer.py +104 -0
- keras_hub/src/models/f_net/__init__.py +20 -0
- keras_hub/src/models/f_net/f_net_backbone.py +236 -0
- keras_hub/src/models/f_net/f_net_classifier.py +154 -0
- keras_hub/src/models/f_net/f_net_masked_lm.py +132 -0
- keras_hub/src/models/f_net/f_net_masked_lm_preprocessor.py +196 -0
- keras_hub/src/models/f_net/f_net_preprocessor.py +177 -0
- keras_hub/src/models/f_net/f_net_presets.py +43 -0
- keras_hub/src/models/f_net/f_net_tokenizer.py +95 -0
- keras_hub/src/models/falcon/__init__.py +20 -0
- keras_hub/src/models/falcon/falcon_attention.py +156 -0
- keras_hub/src/models/falcon/falcon_backbone.py +164 -0
- keras_hub/src/models/falcon/falcon_causal_lm.py +291 -0
- keras_hub/src/models/falcon/falcon_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/falcon/falcon_preprocessor.py +187 -0
- keras_hub/src/models/falcon/falcon_presets.py +30 -0
- keras_hub/src/models/falcon/falcon_tokenizer.py +110 -0
- keras_hub/src/models/falcon/falcon_transformer_decoder.py +255 -0
- keras_hub/src/models/feature_pyramid_backbone.py +73 -0
- keras_hub/src/models/gemma/__init__.py +20 -0
- keras_hub/src/models/gemma/gemma_attention.py +250 -0
- keras_hub/src/models/gemma/gemma_backbone.py +316 -0
- keras_hub/src/models/gemma/gemma_causal_lm.py +448 -0
- keras_hub/src/models/gemma/gemma_causal_lm_preprocessor.py +167 -0
- keras_hub/src/models/gemma/gemma_decoder_block.py +241 -0
- keras_hub/src/models/gemma/gemma_preprocessor.py +191 -0
- keras_hub/src/models/gemma/gemma_presets.py +248 -0
- keras_hub/src/models/gemma/gemma_tokenizer.py +103 -0
- keras_hub/src/models/gemma/rms_normalization.py +40 -0
- keras_hub/src/models/gpt2/__init__.py +20 -0
- keras_hub/src/models/gpt2/gpt2_backbone.py +199 -0
- keras_hub/src/models/gpt2/gpt2_causal_lm.py +437 -0
- keras_hub/src/models/gpt2/gpt2_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/gpt2/gpt2_preprocessor.py +187 -0
- keras_hub/src/models/gpt2/gpt2_presets.py +82 -0
- keras_hub/src/models/gpt2/gpt2_tokenizer.py +110 -0
- keras_hub/src/models/gpt_neo_x/__init__.py +13 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py +251 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_backbone.py +175 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm.py +201 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_causal_lm_preprocessor.py +141 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_decoder.py +258 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_preprocessor.py +145 -0
- keras_hub/src/models/gpt_neo_x/gpt_neo_x_tokenizer.py +88 -0
- keras_hub/src/models/image_classifier.py +90 -0
- keras_hub/src/models/llama/__init__.py +20 -0
- keras_hub/src/models/llama/llama_attention.py +225 -0
- keras_hub/src/models/llama/llama_backbone.py +188 -0
- keras_hub/src/models/llama/llama_causal_lm.py +327 -0
- keras_hub/src/models/llama/llama_causal_lm_preprocessor.py +170 -0
- keras_hub/src/models/llama/llama_decoder.py +246 -0
- keras_hub/src/models/llama/llama_layernorm.py +48 -0
- keras_hub/src/models/llama/llama_preprocessor.py +189 -0
- keras_hub/src/models/llama/llama_presets.py +80 -0
- keras_hub/src/models/llama/llama_tokenizer.py +84 -0
- keras_hub/src/models/llama3/__init__.py +20 -0
- keras_hub/src/models/llama3/llama3_backbone.py +84 -0
- keras_hub/src/models/llama3/llama3_causal_lm.py +46 -0
- keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/llama3/llama3_preprocessor.py +21 -0
- keras_hub/src/models/llama3/llama3_presets.py +69 -0
- keras_hub/src/models/llama3/llama3_tokenizer.py +63 -0
- keras_hub/src/models/masked_lm.py +101 -0
- keras_hub/src/models/mistral/__init__.py +20 -0
- keras_hub/src/models/mistral/mistral_attention.py +238 -0
- keras_hub/src/models/mistral/mistral_backbone.py +203 -0
- keras_hub/src/models/mistral/mistral_causal_lm.py +328 -0
- keras_hub/src/models/mistral/mistral_causal_lm_preprocessor.py +175 -0
- keras_hub/src/models/mistral/mistral_layer_norm.py +48 -0
- keras_hub/src/models/mistral/mistral_preprocessor.py +190 -0
- keras_hub/src/models/mistral/mistral_presets.py +48 -0
- keras_hub/src/models/mistral/mistral_tokenizer.py +82 -0
- keras_hub/src/models/mistral/mistral_transformer_decoder.py +265 -0
- keras_hub/src/models/mix_transformer/__init__.py +13 -0
- keras_hub/src/models/mix_transformer/mix_transformer_backbone.py +181 -0
- keras_hub/src/models/mix_transformer/mix_transformer_classifier.py +133 -0
- keras_hub/src/models/mix_transformer/mix_transformer_layers.py +300 -0
- keras_hub/src/models/opt/__init__.py +20 -0
- keras_hub/src/models/opt/opt_backbone.py +173 -0
- keras_hub/src/models/opt/opt_causal_lm.py +301 -0
- keras_hub/src/models/opt/opt_causal_lm_preprocessor.py +177 -0
- keras_hub/src/models/opt/opt_preprocessor.py +188 -0
- keras_hub/src/models/opt/opt_presets.py +72 -0
- keras_hub/src/models/opt/opt_tokenizer.py +116 -0
- keras_hub/src/models/pali_gemma/__init__.py +23 -0
- keras_hub/src/models/pali_gemma/pali_gemma_backbone.py +277 -0
- keras_hub/src/models/pali_gemma/pali_gemma_causal_lm.py +313 -0
- keras_hub/src/models/pali_gemma/pali_gemma_causal_lm_preprocessor.py +147 -0
- keras_hub/src/models/pali_gemma/pali_gemma_decoder_block.py +160 -0
- keras_hub/src/models/pali_gemma/pali_gemma_presets.py +78 -0
- keras_hub/src/models/pali_gemma/pali_gemma_tokenizer.py +79 -0
- keras_hub/src/models/pali_gemma/pali_gemma_vit.py +566 -0
- keras_hub/src/models/phi3/__init__.py +20 -0
- keras_hub/src/models/phi3/phi3_attention.py +260 -0
- keras_hub/src/models/phi3/phi3_backbone.py +224 -0
- keras_hub/src/models/phi3/phi3_causal_lm.py +218 -0
- keras_hub/src/models/phi3/phi3_causal_lm_preprocessor.py +173 -0
- keras_hub/src/models/phi3/phi3_decoder.py +260 -0
- keras_hub/src/models/phi3/phi3_layernorm.py +48 -0
- keras_hub/src/models/phi3/phi3_preprocessor.py +190 -0
- keras_hub/src/models/phi3/phi3_presets.py +50 -0
- keras_hub/src/models/phi3/phi3_rotary_embedding.py +137 -0
- keras_hub/src/models/phi3/phi3_tokenizer.py +94 -0
- keras_hub/src/models/preprocessor.py +207 -0
- keras_hub/src/models/resnet/__init__.py +13 -0
- keras_hub/src/models/resnet/resnet_backbone.py +612 -0
- keras_hub/src/models/resnet/resnet_image_classifier.py +136 -0
- keras_hub/src/models/roberta/__init__.py +20 -0
- keras_hub/src/models/roberta/roberta_backbone.py +184 -0
- keras_hub/src/models/roberta/roberta_classifier.py +209 -0
- keras_hub/src/models/roberta/roberta_masked_lm.py +136 -0
- keras_hub/src/models/roberta/roberta_masked_lm_preprocessor.py +198 -0
- keras_hub/src/models/roberta/roberta_preprocessor.py +192 -0
- keras_hub/src/models/roberta/roberta_presets.py +43 -0
- keras_hub/src/models/roberta/roberta_tokenizer.py +132 -0
- keras_hub/src/models/seq_2_seq_lm.py +54 -0
- keras_hub/src/models/t5/__init__.py +20 -0
- keras_hub/src/models/t5/t5_backbone.py +261 -0
- keras_hub/src/models/t5/t5_layer_norm.py +35 -0
- keras_hub/src/models/t5/t5_multi_head_attention.py +324 -0
- keras_hub/src/models/t5/t5_presets.py +95 -0
- keras_hub/src/models/t5/t5_tokenizer.py +100 -0
- keras_hub/src/models/t5/t5_transformer_layer.py +178 -0
- keras_hub/src/models/task.py +419 -0
- keras_hub/src/models/vgg/__init__.py +13 -0
- keras_hub/src/models/vgg/vgg_backbone.py +158 -0
- keras_hub/src/models/vgg/vgg_image_classifier.py +124 -0
- keras_hub/src/models/vit_det/__init__.py +13 -0
- keras_hub/src/models/vit_det/vit_det_backbone.py +204 -0
- keras_hub/src/models/vit_det/vit_layers.py +565 -0
- keras_hub/src/models/whisper/__init__.py +20 -0
- keras_hub/src/models/whisper/whisper_audio_feature_extractor.py +260 -0
- keras_hub/src/models/whisper/whisper_backbone.py +305 -0
- keras_hub/src/models/whisper/whisper_cached_multi_head_attention.py +153 -0
- keras_hub/src/models/whisper/whisper_decoder.py +141 -0
- keras_hub/src/models/whisper/whisper_encoder.py +106 -0
- keras_hub/src/models/whisper/whisper_preprocessor.py +326 -0
- keras_hub/src/models/whisper/whisper_presets.py +148 -0
- keras_hub/src/models/whisper/whisper_tokenizer.py +163 -0
- keras_hub/src/models/xlm_roberta/__init__.py +26 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_backbone.py +81 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_classifier.py +225 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm.py +141 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_masked_lm_preprocessor.py +195 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_preprocessor.py +205 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_presets.py +43 -0
- keras_hub/src/models/xlm_roberta/xlm_roberta_tokenizer.py +191 -0
- keras_hub/src/models/xlnet/__init__.py +13 -0
- keras_hub/src/models/xlnet/relative_attention.py +459 -0
- keras_hub/src/models/xlnet/xlnet_backbone.py +222 -0
- keras_hub/src/models/xlnet/xlnet_content_and_query_embedding.py +133 -0
- keras_hub/src/models/xlnet/xlnet_encoder.py +378 -0
- keras_hub/src/samplers/__init__.py +13 -0
- keras_hub/src/samplers/beam_sampler.py +207 -0
- keras_hub/src/samplers/contrastive_sampler.py +231 -0
- keras_hub/src/samplers/greedy_sampler.py +50 -0
- keras_hub/src/samplers/random_sampler.py +77 -0
- keras_hub/src/samplers/sampler.py +237 -0
- keras_hub/src/samplers/serialization.py +97 -0
- keras_hub/src/samplers/top_k_sampler.py +92 -0
- keras_hub/src/samplers/top_p_sampler.py +113 -0
- keras_hub/src/tests/__init__.py +13 -0
- keras_hub/src/tests/test_case.py +608 -0
- keras_hub/src/tokenizers/__init__.py +13 -0
- keras_hub/src/tokenizers/byte_pair_tokenizer.py +638 -0
- keras_hub/src/tokenizers/byte_tokenizer.py +299 -0
- keras_hub/src/tokenizers/sentence_piece_tokenizer.py +267 -0
- keras_hub/src/tokenizers/sentence_piece_tokenizer_trainer.py +150 -0
- keras_hub/src/tokenizers/tokenizer.py +235 -0
- keras_hub/src/tokenizers/unicode_codepoint_tokenizer.py +355 -0
- keras_hub/src/tokenizers/word_piece_tokenizer.py +544 -0
- keras_hub/src/tokenizers/word_piece_tokenizer_trainer.py +176 -0
- keras_hub/src/utils/__init__.py +13 -0
- keras_hub/src/utils/keras_utils.py +130 -0
- keras_hub/src/utils/pipeline_model.py +293 -0
- keras_hub/src/utils/preset_utils.py +621 -0
- keras_hub/src/utils/python_utils.py +21 -0
- keras_hub/src/utils/tensor_utils.py +206 -0
- keras_hub/src/utils/timm/__init__.py +13 -0
- keras_hub/src/utils/timm/convert.py +37 -0
- keras_hub/src/utils/timm/convert_resnet.py +171 -0
- keras_hub/src/utils/transformers/__init__.py +13 -0
- keras_hub/src/utils/transformers/convert.py +101 -0
- keras_hub/src/utils/transformers/convert_bert.py +173 -0
- keras_hub/src/utils/transformers/convert_distilbert.py +184 -0
- keras_hub/src/utils/transformers/convert_gemma.py +187 -0
- keras_hub/src/utils/transformers/convert_gpt2.py +186 -0
- keras_hub/src/utils/transformers/convert_llama3.py +136 -0
- keras_hub/src/utils/transformers/convert_pali_gemma.py +303 -0
- keras_hub/src/utils/transformers/safetensor_utils.py +97 -0
- keras_hub/src/version_utils.py +23 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/METADATA +34 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/RECORD +297 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/WHEEL +5 -0
- keras_hub_nightly-0.15.0.dev20240823171555.dist-info/top_level.txt +1 -0
@@ -0,0 +1,188 @@
|
|
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.opt.opt_tokenizer import OPTTokenizer
|
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.OPTPreprocessor")
|
28
|
+
class OPTPreprocessor(Preprocessor):
|
29
|
+
"""OPT preprocessing layer which tokenizes and packs inputs.
|
30
|
+
|
31
|
+
This preprocessing layer will do 2 things:
|
32
|
+
|
33
|
+
- Tokenize the input using the `tokenizer`.
|
34
|
+
- Construct a dictionary with keys `"token_ids"`, `"padding_mask"`, that can
|
35
|
+
be passed directly to a `keras_hub.models.OPTBackbone`.
|
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
|
+
`OPTPreprocessor` forces the input to have only one segment, as OPT 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.OPTTokenizer` instance.
|
55
|
+
sequence_length: The length of the packed inputs.
|
56
|
+
add_start_token: If `True`, the preprocessor will append 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.OPTPreprocessor.from_preset("opt_125m_en")
|
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.OPTTokenizer(
|
86
|
+
vocabulary=vocab,
|
87
|
+
merges=merges,
|
88
|
+
)
|
89
|
+
preprocessor = keras_hub.models.OPTPreprocessor(tokenizer=tokenizer)
|
90
|
+
preprocessor("The quick brown fox jumped.")
|
91
|
+
```
|
92
|
+
|
93
|
+
Mapping with `tf.data.Dataset`.
|
94
|
+
```python
|
95
|
+
preprocessor = keras_hub.models.OPTPreprocessor.from_preset("opt_125m_en")
|
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 = OPTTokenizer
|
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
|
+
|
122
|
+
self.tokenizer = tokenizer
|
123
|
+
self.packer = None
|
124
|
+
self.sequence_length = sequence_length
|
125
|
+
self.add_start_token = add_start_token
|
126
|
+
self.add_end_token = add_end_token
|
127
|
+
|
128
|
+
def build(self, input_shape):
|
129
|
+
# Defer packer creation to `build()` so that we can be sure tokenizer
|
130
|
+
# assets have loaded when restoring a saved model.
|
131
|
+
self.packer = StartEndPacker(
|
132
|
+
start_value=self.tokenizer.start_token_id,
|
133
|
+
end_value=self.tokenizer.end_token_id,
|
134
|
+
pad_value=self.tokenizer.pad_token_id,
|
135
|
+
sequence_length=self.sequence_length,
|
136
|
+
return_padding_mask=True,
|
137
|
+
)
|
138
|
+
self.built = True
|
139
|
+
|
140
|
+
def get_config(self):
|
141
|
+
config = super().get_config()
|
142
|
+
config.update(
|
143
|
+
{
|
144
|
+
"sequence_length": self.sequence_length,
|
145
|
+
"add_start_token": self.add_start_token,
|
146
|
+
"add_end_token": self.add_end_token,
|
147
|
+
}
|
148
|
+
)
|
149
|
+
return config
|
150
|
+
|
151
|
+
def call(
|
152
|
+
self,
|
153
|
+
x,
|
154
|
+
y=None,
|
155
|
+
sample_weight=None,
|
156
|
+
sequence_length=None,
|
157
|
+
):
|
158
|
+
x = convert_inputs_to_list_of_tensor_segments(x)
|
159
|
+
if len(x) != 1:
|
160
|
+
raise ValueError(
|
161
|
+
"OPT requires each input feature to contain only "
|
162
|
+
f"one segment, but received {len(x)}. If you are using OPT "
|
163
|
+
"for a multi-segment classification task, please refer to "
|
164
|
+
"classification models like BERT or RoBERTa."
|
165
|
+
)
|
166
|
+
sequence_length = sequence_length or self.sequence_length
|
167
|
+
token_ids, padding_mask = self.packer(
|
168
|
+
self.tokenizer(x[0]),
|
169
|
+
sequence_length=sequence_length,
|
170
|
+
add_start_value=self.add_start_token,
|
171
|
+
add_end_value=self.add_end_token,
|
172
|
+
)
|
173
|
+
x = {
|
174
|
+
"token_ids": token_ids,
|
175
|
+
"padding_mask": padding_mask,
|
176
|
+
}
|
177
|
+
return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
|
178
|
+
|
179
|
+
@property
|
180
|
+
def sequence_length(self):
|
181
|
+
"""The padded length of model input sequences."""
|
182
|
+
return self._sequence_length
|
183
|
+
|
184
|
+
@sequence_length.setter
|
185
|
+
def sequence_length(self, value):
|
186
|
+
self._sequence_length = value
|
187
|
+
if self.packer is not None:
|
188
|
+
self.packer.sequence_length = value
|
@@ -0,0 +1,72 @@
|
|
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
|
+
"""OPT model preset configurations."""
|
15
|
+
|
16
|
+
# Metadata for loading pretrained model weights.
|
17
|
+
backbone_presets = {
|
18
|
+
"opt_125m_en": {
|
19
|
+
"metadata": {
|
20
|
+
"description": (
|
21
|
+
"12-layer OPT model where case in maintained. Trained on "
|
22
|
+
"BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
|
23
|
+
),
|
24
|
+
"params": 125237760,
|
25
|
+
"official_name": "OPT",
|
26
|
+
"path": "opt",
|
27
|
+
"model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
|
28
|
+
},
|
29
|
+
"kaggle_handle": "kaggle://keras/opt/keras/opt_125m_en/2",
|
30
|
+
},
|
31
|
+
# We skip the 350m checkpoint because it does not match the structure of
|
32
|
+
# other checkpoints.
|
33
|
+
"opt_1.3b_en": {
|
34
|
+
"metadata": {
|
35
|
+
"description": (
|
36
|
+
"24-layer OPT model where case in maintained. Trained on "
|
37
|
+
"BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
|
38
|
+
),
|
39
|
+
"params": 1315753984,
|
40
|
+
"official_name": "OPT",
|
41
|
+
"path": "opt",
|
42
|
+
"model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
|
43
|
+
},
|
44
|
+
"kaggle_handle": "kaggle://keras/opt/keras/opt_1.3b_en/2",
|
45
|
+
},
|
46
|
+
"opt_2.7b_en": {
|
47
|
+
"metadata": {
|
48
|
+
"description": (
|
49
|
+
"32-layer OPT model where case in maintained. Trained on "
|
50
|
+
"BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
|
51
|
+
),
|
52
|
+
"params": 2700000000,
|
53
|
+
"official_name": "OPT",
|
54
|
+
"path": "opt",
|
55
|
+
"model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
|
56
|
+
},
|
57
|
+
"kaggle_handle": "kaggle://keras/opt/keras/opt_2.7b_en/2",
|
58
|
+
},
|
59
|
+
"opt_6.7b_en": {
|
60
|
+
"metadata": {
|
61
|
+
"description": (
|
62
|
+
"32-layer OPT model where case in maintained. Trained on "
|
63
|
+
"BookCorpus, CommonCrawl, Pile, and PushShift.io corpora."
|
64
|
+
),
|
65
|
+
"params": 6700000000,
|
66
|
+
"official_name": "OPT",
|
67
|
+
"path": "opt",
|
68
|
+
"model_card": "https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/model_card.md",
|
69
|
+
},
|
70
|
+
"kaggle_handle": "kaggle://keras/opt/keras/opt_6.7b_en/2",
|
71
|
+
},
|
72
|
+
}
|
@@ -0,0 +1,116 @@
|
|
1
|
+
# Copyright 2024 The KerasHub Authors
|
2
|
+
#
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4
|
+
# you may not use this file except in compliance with the License.
|
5
|
+
# You may obtain a copy of the License at
|
6
|
+
#
|
7
|
+
# https://www.apache.org/licenses/LICENSE-2.0
|
8
|
+
#
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
+
# See the License for the specific language governing permissions and
|
13
|
+
# limitations under the License.
|
14
|
+
|
15
|
+
|
16
|
+
from keras_hub.src.api_export import keras_hub_export
|
17
|
+
from keras_hub.src.tokenizers.byte_pair_tokenizer import BytePairTokenizer
|
18
|
+
|
19
|
+
|
20
|
+
@keras_hub_export("keras_hub.models.OPTTokenizer")
|
21
|
+
class OPTTokenizer(BytePairTokenizer):
|
22
|
+
"""An OPT tokenizer using Byte-Pair Encoding subword segmentation.
|
23
|
+
|
24
|
+
This tokenizer class will tokenize raw strings into integer sequences and
|
25
|
+
is based on `keras_hub.tokenizers.BytePairTokenizer`. Unlike the
|
26
|
+
underlying tokenizer, it will check for all special tokens needed by OPT
|
27
|
+
models and provides a `from_preset()` method to automatically download
|
28
|
+
a matching vocabulary for a OPT preset.
|
29
|
+
|
30
|
+
This tokenizer does not provide truncation or padding of inputs.
|
31
|
+
|
32
|
+
If input is a batch of strings (rank > 0), the layer will output a
|
33
|
+
`tf.RaggedTensor` where the last dimension of the output is ragged.
|
34
|
+
If input is a scalar string (rank == 0), the layer will output a dense
|
35
|
+
`tf.Tensor` with static shape `[None]`.
|
36
|
+
|
37
|
+
Args:
|
38
|
+
vocabulary: string or dict, maps token to integer ids. If it is a
|
39
|
+
string, it should be the file path to a json file.
|
40
|
+
merges: string or list, contains the merge rule. If it is a string,
|
41
|
+
it should be the file path to merge rules. The merge rule file
|
42
|
+
should have one merge rule per line. Every merge rule contains
|
43
|
+
merge entities separated by a space.
|
44
|
+
|
45
|
+
Examples:
|
46
|
+
```python
|
47
|
+
# Unbatched input.
|
48
|
+
tokenizer = keras_hub.models.OPTTokenizer.from_preset(
|
49
|
+
"opt_125m_en",
|
50
|
+
)
|
51
|
+
tokenizer("The quick brown fox jumped.")
|
52
|
+
|
53
|
+
# Batched input.
|
54
|
+
tokenizer(["The quick brown fox jumped.", "The fox slept."])
|
55
|
+
|
56
|
+
# Detokenization.
|
57
|
+
tokenizer.detokenize(tokenizer("The quick brown fox jumped."))
|
58
|
+
|
59
|
+
# Custom vocabulary.
|
60
|
+
vocab = {"<pad>": 1, "</s>": 2, "Ġquick": 4, "Ġfox": 5}
|
61
|
+
merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
|
62
|
+
merges += ["Ġ f", "o x", "Ġf ox"]
|
63
|
+
tokenizer = keras_hub.models.OPTTokenizer(vocabulary=vocab, merges=merges)
|
64
|
+
tokenizer("The quick brown fox jumped.")
|
65
|
+
```
|
66
|
+
"""
|
67
|
+
|
68
|
+
def __init__(
|
69
|
+
self,
|
70
|
+
vocabulary=None,
|
71
|
+
merges=None,
|
72
|
+
**kwargs,
|
73
|
+
):
|
74
|
+
self.start_token = "</s>"
|
75
|
+
self.pad_token = "<pad>"
|
76
|
+
self.end_token = "</s>"
|
77
|
+
|
78
|
+
super().__init__(
|
79
|
+
vocabulary=vocabulary,
|
80
|
+
merges=merges,
|
81
|
+
unsplittable_tokens=[
|
82
|
+
self.start_token,
|
83
|
+
self.pad_token,
|
84
|
+
self.end_token,
|
85
|
+
],
|
86
|
+
**kwargs,
|
87
|
+
)
|
88
|
+
|
89
|
+
def set_vocabulary_and_merges(self, vocabulary, merges):
|
90
|
+
super().set_vocabulary_and_merges(vocabulary, merges)
|
91
|
+
|
92
|
+
if vocabulary is not None:
|
93
|
+
# Check for necessary special tokens.
|
94
|
+
for token in [self.start_token, self.pad_token, self.end_token]:
|
95
|
+
if token not in self.vocabulary:
|
96
|
+
raise ValueError(
|
97
|
+
f"Cannot find token `'{token}'` in the provided "
|
98
|
+
f"`vocabulary`. Please provide `'{token}'` in your "
|
99
|
+
"`vocabulary` or use a pretrained `vocabulary` name."
|
100
|
+
)
|
101
|
+
|
102
|
+
self.start_token_id = self.token_to_id(self.start_token)
|
103
|
+
self.pad_token_id = self.token_to_id(self.pad_token)
|
104
|
+
self.end_token_id = self.token_to_id(self.end_token)
|
105
|
+
else:
|
106
|
+
self.start_token_id = None
|
107
|
+
self.pad_token_id = None
|
108
|
+
self.end_token_id = None
|
109
|
+
|
110
|
+
def get_config(self):
|
111
|
+
config = super().get_config()
|
112
|
+
# In the constructor, we pass the list of special tokens to the
|
113
|
+
# `unsplittable_tokens` arg of the superclass' constructor. Hence, we
|
114
|
+
# delete it from the config here.
|
115
|
+
del config["unsplittable_tokens"]
|
116
|
+
return config
|
@@ -0,0 +1,23 @@
|
|
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
|
+
from keras_hub.src.models.pali_gemma.pali_gemma_backbone import (
|
15
|
+
PaliGemmaBackbone,
|
16
|
+
)
|
17
|
+
from keras_hub.src.models.pali_gemma.pali_gemma_presets import backbone_presets
|
18
|
+
from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
|
19
|
+
PaliGemmaTokenizer,
|
20
|
+
)
|
21
|
+
from keras_hub.src.utils.preset_utils import register_presets
|
22
|
+
|
23
|
+
register_presets(backbone_presets, (PaliGemmaBackbone, PaliGemmaTokenizer))
|
@@ -0,0 +1,277 @@
|
|
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.layers.modeling.reversible_embedding import (
|
19
|
+
ReversibleEmbedding,
|
20
|
+
)
|
21
|
+
from keras_hub.src.models.backbone import Backbone
|
22
|
+
from keras_hub.src.models.gemma.rms_normalization import RMSNormalization
|
23
|
+
from keras_hub.src.models.pali_gemma.pali_gemma_decoder_block import (
|
24
|
+
PaliGemmaDecoderBlock,
|
25
|
+
)
|
26
|
+
from keras_hub.src.models.pali_gemma.pali_gemma_vit import PaliGemmaVit
|
27
|
+
|
28
|
+
|
29
|
+
@keras_hub_export("keras_hub.models.PaliGemmaBackbone")
|
30
|
+
class PaliGemmaBackbone(Backbone):
|
31
|
+
"""PaliGemma core network with hyperparameters.
|
32
|
+
|
33
|
+
This backbone implements the mixed-modality PaliGemma architecture. It
|
34
|
+
contains a Visual Transformer network, as well as text token embedding
|
35
|
+
layer, followed by a backend-agnostic concatenation operation to
|
36
|
+
construct a sequence of representations of mixed type embeddings (visual
|
37
|
+
and textual). Then, the concatenated sequence is passed through a series
|
38
|
+
of Mixed Modality Decoder Blocks. The returned value from calling this model
|
39
|
+
represents probabilistic values for output tokens.
|
40
|
+
|
41
|
+
For a higher-level object for text-generation,
|
42
|
+
see `keras_hub.models.PaliGemmaCausalLM`.
|
43
|
+
|
44
|
+
The default constructor gives a fully customizable, randomly initialized
|
45
|
+
PaliGemma model with any number of vit layers, heads, embedding
|
46
|
+
dimensions, and equivalent configuration for Paligemma Decoder layers. To
|
47
|
+
load preset architectures and weights, use the `from_preset` constructor.
|
48
|
+
|
49
|
+
Args:
|
50
|
+
vocabulary_size: int. The size of the token vocabulary.
|
51
|
+
image_size: int. The resolution of the image in both width and height.
|
52
|
+
Note: input images must be square.
|
53
|
+
num_layers: int. The number of transformer mixed decoder layers.
|
54
|
+
num_query_heads: int. The number of heads for the query projections in
|
55
|
+
the mixed decoder attention layer.
|
56
|
+
num_key_value_heads: int. The number of heads for the key and value
|
57
|
+
projections in the mixed decoder attention layers.
|
58
|
+
hidden_dim: int. The size of the transformer hidden state at the end
|
59
|
+
of each mixed transformer layer.
|
60
|
+
intermediate_dim: int. The output dimension of the first Dense layer in
|
61
|
+
a two-layer feedforward network for each transformer decoder block.
|
62
|
+
head_dim: int. The size of each attention head in the mixed decoder.
|
63
|
+
vit_patch_size: int. The size of each square patch in the input image.
|
64
|
+
vit_num_heads: int. The number of attention heads for the vision(image)
|
65
|
+
transformer encoder.
|
66
|
+
vit_hidden_dim: int. The size of the transformer hidden state at the end
|
67
|
+
of each vision transformer layer.
|
68
|
+
vit_num_layers: int. The number of vision transformer layers.
|
69
|
+
vit_intermediate_dim: int. The output dimension of the first Dense layer
|
70
|
+
in a two-layer feedforward network for vision transformer.
|
71
|
+
vit_pooling: string. The encoded vision embeddings are pooled using the
|
72
|
+
specified polling setting. The accepted values are `"map"`, `"gap"`,
|
73
|
+
`"0"` or `"none"`. Defaults to `"none"`.
|
74
|
+
vit_classifier_activation: activation function. The activation that
|
75
|
+
is used for final output classification in the vision transformer.
|
76
|
+
vit_name: string. The name used for vision transformer layers.
|
77
|
+
include_rescaling: bool. If true, the image input will be rescaled from
|
78
|
+
the range `[0, 255]`, to the range `[0, 1]`.
|
79
|
+
layer_norm_epsilon: float. The epsilon value user for every layer norm
|
80
|
+
in all transformer blocks.
|
81
|
+
dropout: float. Dropout probability for the Transformer decoder blocks.
|
82
|
+
dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
|
83
|
+
for the models computations and weights. Note that some
|
84
|
+
computations, such as softmax and layer normalization will always
|
85
|
+
be done a float32 precision regardless of dtype.
|
86
|
+
|
87
|
+
Example:
|
88
|
+
```python
|
89
|
+
input_data = {
|
90
|
+
"token_ids": np.ones(shape=(1, 12), dtype="int32"),
|
91
|
+
"images": np.random.uniform(size=(1, 224, 224, 3)),
|
92
|
+
"padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
|
93
|
+
}
|
94
|
+
|
95
|
+
# Pretrained PaliGemma decoder.
|
96
|
+
model = keras_hub.models.PaliGemmaBackbone.from_preset("pali_gemma_mix_224")
|
97
|
+
model(input_data)
|
98
|
+
|
99
|
+
# Randomly initialized PaliGemma decoder with custom config.
|
100
|
+
model = keras_hub.models.PaliGemmaBackbone(
|
101
|
+
vocabulary_size=50257,
|
102
|
+
images_size=224,
|
103
|
+
num_layers=12,
|
104
|
+
num_query_heads=12,
|
105
|
+
num_key_value_heads=1,
|
106
|
+
hidden_dim=768,
|
107
|
+
intermediate_dim=3072,
|
108
|
+
head_dim=64,
|
109
|
+
vit_patch_size=14,
|
110
|
+
vit_num_heads=8,
|
111
|
+
vit_hidden_dim=768,
|
112
|
+
vit_intermediate_dim=3072,
|
113
|
+
vit_num_layers=2,
|
114
|
+
)
|
115
|
+
model(input_data)
|
116
|
+
```
|
117
|
+
"""
|
118
|
+
|
119
|
+
def __init__(
|
120
|
+
self,
|
121
|
+
vocabulary_size,
|
122
|
+
image_size,
|
123
|
+
num_layers,
|
124
|
+
num_query_heads,
|
125
|
+
num_key_value_heads,
|
126
|
+
hidden_dim,
|
127
|
+
intermediate_dim,
|
128
|
+
head_dim,
|
129
|
+
vit_patch_size,
|
130
|
+
vit_num_heads,
|
131
|
+
vit_hidden_dim,
|
132
|
+
vit_num_layers,
|
133
|
+
vit_intermediate_dim=None, # TODO remove default
|
134
|
+
vit_pooling=None,
|
135
|
+
vit_classifier_activation=None,
|
136
|
+
vit_name=None,
|
137
|
+
include_rescaling=True,
|
138
|
+
layer_norm_epsilon=1e-6,
|
139
|
+
dropout=0,
|
140
|
+
dtype=None,
|
141
|
+
**kwargs,
|
142
|
+
):
|
143
|
+
# === Layers ===
|
144
|
+
self.token_embedding = ReversibleEmbedding(
|
145
|
+
input_dim=vocabulary_size,
|
146
|
+
output_dim=hidden_dim,
|
147
|
+
tie_weights=True,
|
148
|
+
embeddings_initializer=keras.initializers.VarianceScaling(
|
149
|
+
scale=1.0,
|
150
|
+
mode="fan_in",
|
151
|
+
distribution="untruncated_normal",
|
152
|
+
seed=None,
|
153
|
+
),
|
154
|
+
dtype=dtype,
|
155
|
+
name="token_embedding",
|
156
|
+
)
|
157
|
+
# TODO Remove this. Work around for previous serialization bug.
|
158
|
+
vit_intermediate_dim = vit_intermediate_dim or 4304
|
159
|
+
self.vit_encoder = PaliGemmaVit(
|
160
|
+
image_size=image_size,
|
161
|
+
include_rescaling=include_rescaling,
|
162
|
+
patch_size=vit_patch_size,
|
163
|
+
num_heads=vit_num_heads,
|
164
|
+
hidden_dim=vit_hidden_dim,
|
165
|
+
num_layers=vit_num_layers,
|
166
|
+
intermediate_dim=vit_intermediate_dim,
|
167
|
+
pooling=vit_pooling,
|
168
|
+
num_classes=hidden_dim,
|
169
|
+
classifier_activation=vit_classifier_activation,
|
170
|
+
dtype=dtype,
|
171
|
+
name=vit_name,
|
172
|
+
)
|
173
|
+
self.transformer_layers = []
|
174
|
+
for i in range(num_layers):
|
175
|
+
layer = PaliGemmaDecoderBlock(
|
176
|
+
hidden_dim=hidden_dim,
|
177
|
+
intermediate_dim=intermediate_dim,
|
178
|
+
num_query_heads=num_query_heads,
|
179
|
+
head_dim=head_dim,
|
180
|
+
num_key_value_heads=num_key_value_heads,
|
181
|
+
dropout=dropout,
|
182
|
+
dtype=dtype,
|
183
|
+
name=f"decoder_block_{i}",
|
184
|
+
)
|
185
|
+
self.transformer_layers.append(layer)
|
186
|
+
self.layer_norm = RMSNormalization(
|
187
|
+
epsilon=layer_norm_epsilon,
|
188
|
+
dtype=dtype,
|
189
|
+
name="final_normalization",
|
190
|
+
)
|
191
|
+
|
192
|
+
# === Functional Model ===
|
193
|
+
image_input = self.vit_encoder.inputs[0]
|
194
|
+
token_id_input = keras.Input(
|
195
|
+
shape=(None,), dtype="int32", name="token_ids"
|
196
|
+
)
|
197
|
+
padding_mask_input = keras.Input(
|
198
|
+
shape=(None,), dtype="int32", name="padding_mask"
|
199
|
+
)
|
200
|
+
response_mask_input = keras.Input(
|
201
|
+
shape=(None,), dtype="int32", name="response_mask"
|
202
|
+
)
|
203
|
+
img_embeddings = self.vit_encoder(image_input)
|
204
|
+
text_embeddings = self.token_embedding(token_id_input)
|
205
|
+
text_embeddings = text_embeddings * ops.cast(
|
206
|
+
ops.sqrt(hidden_dim), text_embeddings.dtype
|
207
|
+
)
|
208
|
+
x = ops.concatenate((img_embeddings, text_embeddings), axis=1)
|
209
|
+
for transformer_layer in self.transformer_layers:
|
210
|
+
x = transformer_layer(
|
211
|
+
x,
|
212
|
+
padding_mask=padding_mask_input,
|
213
|
+
response_mask=response_mask_input,
|
214
|
+
)
|
215
|
+
sequence_output = self.layer_norm(x)
|
216
|
+
super().__init__(
|
217
|
+
inputs={
|
218
|
+
"images": image_input,
|
219
|
+
"token_ids": token_id_input,
|
220
|
+
"padding_mask": padding_mask_input,
|
221
|
+
"response_mask": response_mask_input,
|
222
|
+
},
|
223
|
+
outputs=sequence_output,
|
224
|
+
dtype=dtype,
|
225
|
+
**kwargs,
|
226
|
+
)
|
227
|
+
|
228
|
+
# === Config ===
|
229
|
+
self.vocabulary_size = vocabulary_size
|
230
|
+
self.image_size = image_size
|
231
|
+
self.include_rescaling = include_rescaling
|
232
|
+
self.num_layers = num_layers
|
233
|
+
self.num_query_heads = num_query_heads
|
234
|
+
self.num_key_value_heads = num_key_value_heads
|
235
|
+
self.hidden_dim = hidden_dim
|
236
|
+
self.intermediate_dim = intermediate_dim
|
237
|
+
self.head_dim = head_dim
|
238
|
+
self.layer_norm_epsilon = layer_norm_epsilon
|
239
|
+
self.dropout = dropout
|
240
|
+
# VIT Params
|
241
|
+
self.vit_patch_size = vit_patch_size
|
242
|
+
self.vit_num_heads = vit_num_heads
|
243
|
+
self.vit_hidden_dim = vit_hidden_dim
|
244
|
+
self.vit_num_layers = vit_num_layers
|
245
|
+
self.vit_intermediate_dim = vit_intermediate_dim
|
246
|
+
self.vit_pooling = vit_pooling
|
247
|
+
self.vit_classifier_activation = vit_classifier_activation
|
248
|
+
self.vit_name = vit_name
|
249
|
+
# Keep the image_sequence_length as a backbone property for easy access.
|
250
|
+
self.image_sequence_length = self.vit_encoder.image_sequence_length
|
251
|
+
|
252
|
+
def get_config(self):
|
253
|
+
config = super().get_config()
|
254
|
+
config.update(
|
255
|
+
{
|
256
|
+
"vocabulary_size": self.vocabulary_size,
|
257
|
+
"image_size": self.image_size,
|
258
|
+
"include_rescaling": self.include_rescaling,
|
259
|
+
"num_layers": self.num_layers,
|
260
|
+
"num_query_heads": self.num_query_heads,
|
261
|
+
"num_key_value_heads": self.num_key_value_heads,
|
262
|
+
"hidden_dim": self.hidden_dim,
|
263
|
+
"intermediate_dim": self.intermediate_dim,
|
264
|
+
"head_dim": self.head_dim,
|
265
|
+
"layer_norm_epsilon": self.layer_norm_epsilon,
|
266
|
+
"dropout": self.dropout,
|
267
|
+
"vit_patch_size": self.vit_patch_size,
|
268
|
+
"vit_num_heads": self.vit_num_heads,
|
269
|
+
"vit_hidden_dim": self.vit_hidden_dim,
|
270
|
+
"vit_num_layers": self.vit_num_layers,
|
271
|
+
"vit_intermediate_dim": self.vit_intermediate_dim,
|
272
|
+
"vit_pooling": self.vit_pooling,
|
273
|
+
"vit_classifier_activation": self.vit_classifier_activation,
|
274
|
+
"vit_name": self.vit_name,
|
275
|
+
}
|
276
|
+
)
|
277
|
+
return config
|