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 +1,153 @@
|
|
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
|
+
"""Whisper Cached Multi-Head Attention layer."""
|
15
|
+
|
16
|
+
import collections
|
17
|
+
import string
|
18
|
+
|
19
|
+
import keras
|
20
|
+
|
21
|
+
from keras_hub.src.layers.modeling.cached_multi_head_attention import (
|
22
|
+
CachedMultiHeadAttention,
|
23
|
+
)
|
24
|
+
|
25
|
+
|
26
|
+
def _index_to_einsum_variable(i):
|
27
|
+
"""Converts an index to a einsum variable name.
|
28
|
+
|
29
|
+
We simply map indices to lowercase characters, e.g. 0 -> 'a', 1 -> 'b'.
|
30
|
+
"""
|
31
|
+
return string.ascii_lowercase[i]
|
32
|
+
|
33
|
+
|
34
|
+
def _build_proj_equation(free_dims, bound_dims, output_dims):
|
35
|
+
"""Builds an einsum equation for projections inside multi-head attention."""
|
36
|
+
input_str = ""
|
37
|
+
kernel_str = ""
|
38
|
+
output_str = ""
|
39
|
+
bias_axes = ""
|
40
|
+
letter_offset = 0
|
41
|
+
for i in range(free_dims):
|
42
|
+
char = _index_to_einsum_variable(i + letter_offset)
|
43
|
+
input_str += char
|
44
|
+
output_str += char
|
45
|
+
|
46
|
+
letter_offset += free_dims
|
47
|
+
for i in range(bound_dims):
|
48
|
+
char = _index_to_einsum_variable(i + letter_offset)
|
49
|
+
input_str += char
|
50
|
+
kernel_str += char
|
51
|
+
|
52
|
+
letter_offset += bound_dims
|
53
|
+
for i in range(output_dims):
|
54
|
+
char = _index_to_einsum_variable(i + letter_offset)
|
55
|
+
kernel_str += char
|
56
|
+
output_str += char
|
57
|
+
bias_axes += char
|
58
|
+
equation = f"{input_str},{kernel_str}->{output_str}"
|
59
|
+
|
60
|
+
return equation, bias_axes, len(output_str)
|
61
|
+
|
62
|
+
|
63
|
+
def _get_output_shape(output_rank, known_last_dims):
|
64
|
+
return [None] * (output_rank - len(known_last_dims)) + list(known_last_dims)
|
65
|
+
|
66
|
+
|
67
|
+
@keras.saving.register_keras_serializable(package="keras_hub")
|
68
|
+
class WhisperCachedMultiHeadAttention(CachedMultiHeadAttention):
|
69
|
+
"""Whisper Cached Multi-Head Attention layer.
|
70
|
+
|
71
|
+
Inherits from `keras_hub.layers.CachedMultiHeadAttention`, and overrides the
|
72
|
+
`build` method so that Q, V projection layers have bias
|
73
|
+
whereas K projection layer does not.
|
74
|
+
"""
|
75
|
+
|
76
|
+
def build(
|
77
|
+
self,
|
78
|
+
query_shape,
|
79
|
+
value_shape,
|
80
|
+
key_shape=None,
|
81
|
+
):
|
82
|
+
key_shape = value_shape if key_shape is None else key_shape
|
83
|
+
query_rank = len(query_shape)
|
84
|
+
value_rank = len(value_shape)
|
85
|
+
key_rank = len(key_shape)
|
86
|
+
einsum_equation, bias_axes, output_rank = _build_proj_equation(
|
87
|
+
query_rank - 1, bound_dims=1, output_dims=2
|
88
|
+
)
|
89
|
+
self._query_dense = keras.layers.EinsumDense(
|
90
|
+
einsum_equation,
|
91
|
+
output_shape=_get_output_shape(
|
92
|
+
output_rank - 1, [self._num_heads, self._key_dim]
|
93
|
+
),
|
94
|
+
bias_axes=bias_axes if self._use_bias else None,
|
95
|
+
name="query",
|
96
|
+
**self._get_common_kwargs_for_sublayer(),
|
97
|
+
)
|
98
|
+
self._query_dense.build(query_shape)
|
99
|
+
einsum_equation, bias_axes, output_rank = _build_proj_equation(
|
100
|
+
key_rank - 1, bound_dims=1, output_dims=2
|
101
|
+
)
|
102
|
+
self._key_dense = keras.layers.EinsumDense(
|
103
|
+
einsum_equation,
|
104
|
+
output_shape=_get_output_shape(
|
105
|
+
output_rank - 1, [self._num_heads, self._key_dim]
|
106
|
+
),
|
107
|
+
bias_axes=None,
|
108
|
+
name="key",
|
109
|
+
**self._get_common_kwargs_for_sublayer(),
|
110
|
+
)
|
111
|
+
self._key_dense.build(key_shape)
|
112
|
+
einsum_equation, bias_axes, output_rank = _build_proj_equation(
|
113
|
+
value_rank - 1, bound_dims=1, output_dims=2
|
114
|
+
)
|
115
|
+
self._value_dense = keras.layers.EinsumDense(
|
116
|
+
einsum_equation,
|
117
|
+
output_shape=_get_output_shape(
|
118
|
+
output_rank - 1, [self._num_heads, self._value_dim]
|
119
|
+
),
|
120
|
+
bias_axes=bias_axes if self._use_bias else None,
|
121
|
+
name="value",
|
122
|
+
**self._get_common_kwargs_for_sublayer(),
|
123
|
+
)
|
124
|
+
self._value_dense.build(value_shape)
|
125
|
+
|
126
|
+
# Builds the attention computations for multi-head dot product
|
127
|
+
# attention. These computations could be wrapped into the keras
|
128
|
+
# attention layer once it supports multi-head einsum computations.
|
129
|
+
self._build_attention(output_rank)
|
130
|
+
|
131
|
+
if self._output_shape:
|
132
|
+
if not isinstance(self._output_shape, collections.abc.Sized):
|
133
|
+
output_shape = [self._output_shape]
|
134
|
+
else:
|
135
|
+
output_shape = self._output_shape
|
136
|
+
else:
|
137
|
+
output_shape = [query_shape[-1]]
|
138
|
+
einsum_equation, bias_axes, output_rank = _build_proj_equation(
|
139
|
+
query_rank - 1, bound_dims=2, output_dims=len(output_shape)
|
140
|
+
)
|
141
|
+
self._output_dense = keras.layers.EinsumDense(
|
142
|
+
einsum_equation,
|
143
|
+
output_shape=_get_output_shape(output_rank - 1, output_shape),
|
144
|
+
bias_axes=bias_axes if self._use_bias else None,
|
145
|
+
name="attention_output",
|
146
|
+
**self._get_common_kwargs_for_sublayer(),
|
147
|
+
)
|
148
|
+
output_dense_input_shape = list(
|
149
|
+
self._query_dense.compute_output_shape(query_shape)
|
150
|
+
)
|
151
|
+
output_dense_input_shape[-1] = self._value_dim
|
152
|
+
self._output_dense.build(tuple(output_dense_input_shape))
|
153
|
+
self.built = True
|
@@ -0,0 +1,141 @@
|
|
1
|
+
# Copyright 2024 The KerasHub Authors
|
2
|
+
#
|
3
|
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4
|
+
# you may not use this file except in compliance with the License.
|
5
|
+
# You may obtain a copy of the License at
|
6
|
+
#
|
7
|
+
# https://www.apache.org/licenses/LICENSE-2.0
|
8
|
+
#
|
9
|
+
# Unless required by applicable law or agreed to in writing, software
|
10
|
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11
|
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12
|
+
# See the License for the specific language governing permissions and
|
13
|
+
# limitations under the License.
|
14
|
+
"""Whisper decoder block."""
|
15
|
+
|
16
|
+
|
17
|
+
import keras
|
18
|
+
|
19
|
+
from keras_hub.src.layers.modeling.transformer_decoder import TransformerDecoder
|
20
|
+
from keras_hub.src.models.whisper.whisper_cached_multi_head_attention import (
|
21
|
+
WhisperCachedMultiHeadAttention,
|
22
|
+
)
|
23
|
+
from keras_hub.src.utils.keras_utils import clone_initializer
|
24
|
+
|
25
|
+
|
26
|
+
@keras.saving.register_keras_serializable(package="keras_hub")
|
27
|
+
class WhisperDecoder(TransformerDecoder):
|
28
|
+
"""Whisper decoder.
|
29
|
+
|
30
|
+
Inherits from `keras_hub.layers.TransformerDecoder`, and overrides the
|
31
|
+
`build` method to use the
|
32
|
+
`keras_hub.models.whisper.whisper_multi_head_attention.WhisperMultiHeadAttention`
|
33
|
+
layer instead of `keras.layers.MultiHeadAttention` and
|
34
|
+
`keras_hub.models.whisper.whisper_cached_multi_head_attention.WhisperCachedMultiHeadAttention`
|
35
|
+
instead of `keras_hub.layers.cached_multi_head_attention.CachedMultiHeadAttention`.
|
36
|
+
"""
|
37
|
+
|
38
|
+
def build(
|
39
|
+
self,
|
40
|
+
decoder_sequence_shape,
|
41
|
+
encoder_sequence_shape,
|
42
|
+
):
|
43
|
+
self._decoder_sequence_shape = decoder_sequence_shape
|
44
|
+
self._encoder_sequence_shape = encoder_sequence_shape
|
45
|
+
# Infer the dimension of our hidden feature size from the build shape.
|
46
|
+
hidden_dim = decoder_sequence_shape[-1]
|
47
|
+
# Attention head size is `hidden_dim` over the number of heads.
|
48
|
+
head_dim = int(hidden_dim // self.num_heads)
|
49
|
+
if head_dim == 0:
|
50
|
+
raise ValueError(
|
51
|
+
"Attention `head_dim` computed cannot be zero. "
|
52
|
+
f"The `hidden_dim` value of {hidden_dim} has to be equal to "
|
53
|
+
f"or greater than `num_heads` value of {self.num_heads}."
|
54
|
+
)
|
55
|
+
|
56
|
+
# Self attention layers.
|
57
|
+
self._self_attention_layer = WhisperCachedMultiHeadAttention(
|
58
|
+
num_heads=self.num_heads,
|
59
|
+
key_dim=head_dim,
|
60
|
+
dropout=self.dropout,
|
61
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
62
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
63
|
+
dtype=self.dtype_policy,
|
64
|
+
name="self_attention",
|
65
|
+
)
|
66
|
+
|
67
|
+
self._self_attention_layer.build(
|
68
|
+
query_shape=decoder_sequence_shape,
|
69
|
+
value_shape=decoder_sequence_shape,
|
70
|
+
)
|
71
|
+
self._self_attention_layer_norm = keras.layers.LayerNormalization(
|
72
|
+
epsilon=self.layer_norm_epsilon,
|
73
|
+
dtype=self.dtype_policy,
|
74
|
+
name="self_attention_layer_norm",
|
75
|
+
)
|
76
|
+
self._self_attention_layer_norm.build(decoder_sequence_shape)
|
77
|
+
self._self_attention_dropout = keras.layers.Dropout(
|
78
|
+
rate=self.dropout,
|
79
|
+
dtype=self.dtype_policy,
|
80
|
+
name="self_attention_dropout",
|
81
|
+
)
|
82
|
+
|
83
|
+
self._cross_attention_layer = WhisperCachedMultiHeadAttention(
|
84
|
+
num_heads=self.num_heads,
|
85
|
+
key_dim=head_dim,
|
86
|
+
value_dim=head_dim,
|
87
|
+
dropout=self.dropout,
|
88
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
89
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
90
|
+
dtype=self.dtype_policy,
|
91
|
+
name="cross_attention",
|
92
|
+
)
|
93
|
+
self._cross_attention_layer.build(
|
94
|
+
query_shape=decoder_sequence_shape,
|
95
|
+
value_shape=encoder_sequence_shape,
|
96
|
+
)
|
97
|
+
self._cross_attention_layer_norm = keras.layers.LayerNormalization(
|
98
|
+
epsilon=self.layer_norm_epsilon,
|
99
|
+
dtype=self.dtype_policy,
|
100
|
+
name="cross_attention_layer_norm",
|
101
|
+
)
|
102
|
+
self._cross_attention_layer_norm.build(decoder_sequence_shape)
|
103
|
+
self._cross_attention_dropout = keras.layers.Dropout(
|
104
|
+
rate=self.dropout,
|
105
|
+
dtype=self.dtype_policy,
|
106
|
+
name="cross_attention_dropout",
|
107
|
+
)
|
108
|
+
|
109
|
+
# Feedforward layers.
|
110
|
+
self._feedforward_intermediate_dense = keras.layers.Dense(
|
111
|
+
self.intermediate_dim,
|
112
|
+
activation=self.activation,
|
113
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
114
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
115
|
+
dtype=self.dtype_policy,
|
116
|
+
name="feedforward_intermediate_dense",
|
117
|
+
)
|
118
|
+
self._feedforward_intermediate_dense.build(decoder_sequence_shape)
|
119
|
+
self._feedforward_output_dense = keras.layers.Dense(
|
120
|
+
hidden_dim,
|
121
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
122
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
123
|
+
dtype=self.dtype_policy,
|
124
|
+
name="feedforward_output_dense",
|
125
|
+
)
|
126
|
+
intermediate_shape = list(decoder_sequence_shape)
|
127
|
+
intermediate_shape[-1] = self.intermediate_dim
|
128
|
+
self._feedforward_output_dense.build(tuple(intermediate_shape))
|
129
|
+
self._feedforward_layer_norm = keras.layers.LayerNormalization(
|
130
|
+
epsilon=self.layer_norm_epsilon,
|
131
|
+
dtype=self.dtype_policy,
|
132
|
+
name="feedforward_layer_norm",
|
133
|
+
)
|
134
|
+
self._feedforward_layer_norm.build(decoder_sequence_shape)
|
135
|
+
self._feedforward_dropout = keras.layers.Dropout(
|
136
|
+
rate=self.dropout,
|
137
|
+
dtype=self.dtype_policy,
|
138
|
+
name="feedforward_dropout",
|
139
|
+
)
|
140
|
+
# Create layers based on input shape.
|
141
|
+
self.built = True
|
@@ -0,0 +1,106 @@
|
|
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
|
+
"""Whisper encoder block."""
|
15
|
+
|
16
|
+
|
17
|
+
import keras
|
18
|
+
|
19
|
+
from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
|
20
|
+
from keras_hub.src.models.whisper.whisper_cached_multi_head_attention import (
|
21
|
+
WhisperCachedMultiHeadAttention,
|
22
|
+
)
|
23
|
+
from keras_hub.src.utils.keras_utils import clone_initializer
|
24
|
+
|
25
|
+
|
26
|
+
@keras.saving.register_keras_serializable(package="keras_hub")
|
27
|
+
class WhisperEncoder(TransformerEncoder):
|
28
|
+
"""Whisper encoder.
|
29
|
+
|
30
|
+
Inherits from `keras_hub.layers.TransformerEncoder`, and overrides the
|
31
|
+
`_build` method to use the
|
32
|
+
`keras_hub.models.whisper.whisper_multi_head_attention.WhisperCachedMultiHeadAttention`
|
33
|
+
layer instead of `keras.layers.MultiHeadAttention`.
|
34
|
+
"""
|
35
|
+
|
36
|
+
def build(self, inputs_shape):
|
37
|
+
# Infer the dimension of our hidden feature size from the build shape.
|
38
|
+
hidden_dim = inputs_shape[-1]
|
39
|
+
# Attention head size is `hidden_dim` over the number of heads.
|
40
|
+
key_dim = int(hidden_dim // self.num_heads)
|
41
|
+
if key_dim == 0:
|
42
|
+
raise ValueError(
|
43
|
+
"Attention `key_dim` computed cannot be zero. "
|
44
|
+
f"The `hidden_dim` value of {hidden_dim} has to be equal to "
|
45
|
+
f"or greater than `num_heads` value of {self.num_heads}."
|
46
|
+
)
|
47
|
+
|
48
|
+
# Self attention layers.
|
49
|
+
self._self_attention_layer = WhisperCachedMultiHeadAttention(
|
50
|
+
num_heads=self.num_heads,
|
51
|
+
key_dim=key_dim,
|
52
|
+
dropout=self.dropout,
|
53
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
54
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
55
|
+
dtype=self.dtype_policy,
|
56
|
+
name="self_attention_layer",
|
57
|
+
)
|
58
|
+
self._self_attention_layer.build(
|
59
|
+
query_shape=inputs_shape,
|
60
|
+
value_shape=inputs_shape,
|
61
|
+
)
|
62
|
+
|
63
|
+
self._self_attention_layer_norm = keras.layers.LayerNormalization(
|
64
|
+
epsilon=self.layer_norm_epsilon,
|
65
|
+
dtype=self.dtype_policy,
|
66
|
+
name="self_attention_layer_norm",
|
67
|
+
)
|
68
|
+
self._self_attention_layer_norm.build(inputs_shape)
|
69
|
+
self._self_attention_dropout = keras.layers.Dropout(
|
70
|
+
rate=self.dropout,
|
71
|
+
dtype=self.dtype_policy,
|
72
|
+
name="self_attention_dropout",
|
73
|
+
)
|
74
|
+
|
75
|
+
# Feedforward layers.
|
76
|
+
self._feedforward_layer_norm = keras.layers.LayerNormalization(
|
77
|
+
epsilon=self.layer_norm_epsilon,
|
78
|
+
dtype=self.dtype_policy,
|
79
|
+
name="feedforward_layer_norm",
|
80
|
+
)
|
81
|
+
self._feedforward_layer_norm.build(inputs_shape)
|
82
|
+
self._feedforward_intermediate_dense = keras.layers.Dense(
|
83
|
+
self.intermediate_dim,
|
84
|
+
activation=self.activation,
|
85
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
86
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
87
|
+
dtype=self.dtype_policy,
|
88
|
+
name="feedforward_intermediate_dense",
|
89
|
+
)
|
90
|
+
self._feedforward_intermediate_dense.build(inputs_shape)
|
91
|
+
self._feedforward_output_dense = keras.layers.Dense(
|
92
|
+
hidden_dim,
|
93
|
+
kernel_initializer=clone_initializer(self.kernel_initializer),
|
94
|
+
bias_initializer=clone_initializer(self.bias_initializer),
|
95
|
+
dtype=self.dtype_policy,
|
96
|
+
name="feedforward_output_dense",
|
97
|
+
)
|
98
|
+
intermediate_shape = list(inputs_shape)
|
99
|
+
intermediate_shape[-1] = self.intermediate_dim
|
100
|
+
self._feedforward_output_dense.build(tuple(intermediate_shape))
|
101
|
+
self._feedforward_dropout = keras.layers.Dropout(
|
102
|
+
rate=self.dropout,
|
103
|
+
dtype=self.dtype_policy,
|
104
|
+
name="feedforward_dropout",
|
105
|
+
)
|
106
|
+
self.built = True
|