keras-hub-nightly 0.21.0.dev202505230409__py3-none-any.whl → 0.21.0.dev202505240409__py3-none-any.whl

keras_hub/src/models/audio_to_text.py

@@ -0,0 +1,66 @@
+from keras_hub.src.models.seq_2_seq_lm import Seq2SeqLM
+
+
+class AudioToText(Seq2SeqLM):
+    """Base class for audio-to-text models.
+
+    `AudioToText` tasks wrap a `keras_hub.models.Backbone` (capable of
+    processing audio and text features) and a
+    `keras_hub.models.AudioToTextPreprocessor` to create a model for
+    audio-to-text tasks like speech recognition or audio transcription.
+
+    These models typically consist of an encoder that processes audio input
+    and a decoder that generates a textual representation.
+
+    `AudioToText` tasks provide a high-level `generate()` method for
+    auto-regressively generating text from audio input. An optional text
+    prompt can also be provided to the decoder to guide generation. The
+    sampling strategy for generation (e.g., greedy, top-k, top-p) can be
+    controlled via the `sampler` argument in the `compile()` method.
+
+    When calling `fit()`, inputs should consist of audio data and corresponding
+    target text transcriptions. The model is trained to predict the target text
+    token-by-token.
+
+    All `AudioToText` tasks include a `from_preset()` constructor which
+    can be used to load pre-trained configurations and weights for specific
+    audio-to-text models.
+    This constructor can also be called on the base `AudioToText` class,
+    which will automatically select the correct subclass based on the preset.
+
+    Examples:
+    ```python
+    # Load a Moonshine backbone with pre-trained weights.
+    # AudioToText is a base class. You will typically work with a specific
+    # implementation, such as `keras_hub.models.MoonshineAudioToText`.
+    # The following examples demonstrate common usage patterns.
+
+    # Initialize a model from a preset using the specific subclass.
+    audio_to_text = keras_hub.models.MoonshineAudioToText.from_preset(
+        "moonshine_base_en"
+    )
+
+    # Initialize a model from a preset using the base class.
+    audio_to_text_model_base = keras_hub.models.AudioToText.from_preset(
+        "moonshine_base_en"
+    )
+
+    # Generate text from an audio input.
+    audio_input_tensor = keras.random.normal((1, 16000, 1))
+    generated_output = audio_to_text_model.generate(
+        {"audio": audio_input_tensor}
+    )
+
+    # Generate conditioned on the `"The quick brown fox."` as an input sequence.
+    prompted_output = audio_to_text_model.generate(
+        {"audio": audio_input_tensor, "text": "The quick brown fox."}
+    )
+
+    # Use a different sampling strategy for generation.
+    audio_to_text_model.compile(sampler="greedy")
+    greedy_output = audio_to_text_model.generate(
+        {"audio": audio_input_tensor}
+    )
+    """
+
+    # TODO: Fill in once audio to text task model requirements are clearer.

keras_hub/src/models/audio_to_text_preprocessor.py

@@ -0,0 +1,80 @@
+from keras_hub.src.models.seq_2_seq_lm_preprocessor import Seq2SeqLMPreprocessor
+
+
+class AudioToTextPreprocessor(Seq2SeqLMPreprocessor):
+    """Base class for audio-to-text preprocessing layers.
+
+    `AudioToTextPreprocessor` layers wrap an audio feature extractor (specific
+    to the subclass) and a `keras_hub.tokenizer.Tokenizer` to create a
+    preprocessing layer for audio-to-text tasks. It is intended to be
+    paired with a `keras_hub.models.AudioToText` task.
+
+    Subclasses are expected to handle the conversion of raw audio data into
+    numerical features suitable for an encoder, and raw text data into token IDs
+    for a decoder.
+
+    All `AudioToTextPreprocessor` layers take a dictionary as input,
+    typically with keys like `"audio"` (for audio data) and `"text"` (for
+    target transcriptions or decoder prompts).
+
+    This layer will always output a `(x, y, sample_weight)` tuple, where `x`
+    is a dictionary containing processed audio features for the encoder and
+    tokenized text inputs for the decoder. `y` contains the target token IDs
+    (decoder input tokens shifted by one position), and `sample_weight`
+    indicates padding in `y`. The exact keys and structure of features within
+    `x` will depend on the specific subclass and the paired `AudioToText` model.
+
+    An `AudioToTextPreprocessor` includes `generate_preprocess` and
+    `generate_postprocess` methods for use during inference with an
+    `AudioToText` model's `generate()` method.
+
+    All `AudioToTextPreprocessor` tasks include a `from_preset()` constructor
+    which can be used to load a pre-trained configuration, including tokenizer
+    vocabularies and audio feature extraction settings. Calling `from_preset()`
+    on this base class can instantiate the correct subclass registered for the
+    given preset.
+
+    Examples:
+    ```python
+    preprocessor = keras_hub.models.AudioToTextPreprocessor.from_preset(
+        "moonshine_base_en",
+        decoder_sequence_length=10
+    )
+
+    # Process a single audio-text pair.
+    x = {
+        "audio": keras.random.normal((1, 16000, 1)),
+        "text": ["the quick brown fox"]
+    }
+    x, y, sample_weight = preprocessor(x)
+
+    # Process a batch of audio-text pairs.
+    x = {
+        "audio": keras.random.normal((2, 16000, 1)),
+        "text": ["first sentence", "second sentence"]
+    }
+    x, y, sample_weight = preprocessor(x)
+
+    # With a `tf.data.Dataset`.
+    audio_tf = keras.ops.convert_to_tensor(batch_input["audio"])
+    text_tf = batch_input["text"] # List of strings
+    x = {"audio": audio_tf, "text": text_tf}
+    ds = tf.data.Dataset.from_tensor_slices(x)
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    ds = ds.batch(2) # Batching after map
+
+    # Generate preprocess and postprocess.
+    x = preprocessor.generate_preprocess({
+        "audio": keras.random.normal((1, 16000, 1)),
+        "text": ["optional prompt text"]
+    })
+    x = preprocessor.generate_postprocess({
+        "decoder_token_ids": keras.ops.array([[10, 20, 30, 2, 0]]),
+        "decoder_padding_mask": keras.ops.array([
+            [True, True, True, True, False]
+        ])
+    })
+    ```
+    """
+
+    # TODO: Fill in once audio to text task model requirements are clearer.

keras_hub/src/models/moonshine/__init__.py

@@ -0,0 +1,5 @@
+from keras_hub.src.models.moonshine.moonshine_backbone import MoonshineBackbone
+from keras_hub.src.models.moonshine.moonshine_presets import backbone_presets
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, MoonshineBackbone)

keras_hub/src/models/moonshine/moonshine_audio_to_text.py

@@ -1,6 +1,7 @@
 import keras
 
 from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.audio_to_text import AudioToText
 from keras_hub.src.models.moonshine.moonshine_audio_to_text_preprocessor import (  # noqa: E501
     MoonshineAudioToTextPreprocessor,
 )
@@ -9,12 +10,11 @@ from keras_hub.src.models.moonshine.moonshine_backbone import MoonshineBackbone
 from keras_hub.src.models.moonshine.moonshine_backbone import (
     compute_output_lengths,
 )
-from keras_hub.src.models.seq_2_seq_lm import Seq2SeqLM
 from keras_hub.src.utils.tensor_utils import any_equal
 
 
 @keras_hub_export("keras_hub.models.MoonshineAudioToText")
-class MoonshineAudioToText(Seq2SeqLM):
+class MoonshineAudioToText(AudioToText):
     """An end-to-end Moonshine model for audio-to-text tasks.
 
     A Seq2Seq LM designed for audio-to-text tasks, such as speech recognition.

keras_hub/src/models/moonshine/moonshine_audio_to_text_preprocessor.py

@@ -6,16 +6,18 @@ except ImportError:
     tf = None
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.layers.preprocessing.start_end_packer import StartEndPacker
+from keras_hub.src.models.audio_to_text_preprocessor import (
+    AudioToTextPreprocessor,
+)
 from keras_hub.src.models.moonshine.moonshine_backbone import MoonshineBackbone
 from keras_hub.src.models.moonshine.moonshine_tokenizer import (
     MoonshineTokenizer,
 )
-from keras_hub.src.models.seq_2_seq_lm_preprocessor import Seq2SeqLMPreprocessor
 from keras_hub.src.utils.tensor_utils import preprocessing_function
 
 
 @keras_hub_export("keras_hub.models.MoonshineAudioToTextPreprocessor")
-class MoonshineAudioToTextPreprocessor(Seq2SeqLMPreprocessor):
+class MoonshineAudioToTextPreprocessor(AudioToTextPreprocessor):
     """Moonshine Seq2Seq LM preprocessor for audio-to-text tasks.
 
     This preprocessor converts raw audio and text inputs into a format suitable

keras_hub/src/version.py

@@ -1,7 +1,7 @@
 from keras_hub.src.api_export import keras_hub_export
 
 # Unique source of truth for the version number.
-__version__ = "0.21.0.dev202505230409"
+__version__ = "0.21.0.dev202505240409"
 
 
 @keras_hub_export("keras_hub.version")

{keras_hub_nightly-0.21.0.dev202505230409.dist-info → keras_hub_nightly-0.21.0.dev202505240409.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.21.0.dev202505230409
+Version: 0.21.0.dev202505240409
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.21.0.dev202505230409.dist-info → keras_hub_nightly-0.21.0.dev202505240409.dist-info}/RECORD

@@ -5,7 +5,7 @@ keras_hub/models/__init__.py,sha256=itSzodVUeuX6HQnmsSXY0Wv-5Htbu397410R-SFW_4I,
 keras_hub/samplers/__init__.py,sha256=aFQIkiqbZpi8vjrPp2MVII4QUfE-eQjra5fMeHsoy7k,886
 keras_hub/src/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/api_export.py,sha256=9pQZK27JObxWZ96QPLBp1OBsjWigh1iuV6RglPGMRk0,1499
-keras_hub/src/version.py,sha256=SjWdrHYDbNitBzSsMmxG-HvuuqsSB3ICvTQclkoX-Os,222
+keras_hub/src/version.py,sha256=AnU8tBqSqSoLY34F6O-fFt47PXgrGHUbqORa6_sXy6w,222
 keras_hub/src/layers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/layers/modeling/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 keras_hub/src/layers/modeling/alibi_bias.py,sha256=1XBTHI52L_iJDhN_w5ydu_iMhCuTgQAxEPwcLA6BPuk,4411
@@ -41,6 +41,8 @@ keras_hub/src/metrics/rouge_base.py,sha256=Pt2DUznhTTeR-fX1nQ_wSbPtmuTgxQTvrGpu8
 keras_hub/src/metrics/rouge_l.py,sha256=JlZhMBV6wS_6zMd57pkTc6yxHkEJT9fVQMlPZKekQzQ,2729
 keras_hub/src/metrics/rouge_n.py,sha256=JoFtmgjF4Ic263ny6bfD6vMHKreH9le3HnOOxemupRc,3620
 keras_hub/src/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+keras_hub/src/models/audio_to_text.py,sha256=XoOjXtKBX6K1fz-zOXcdVo3FpjuxCMnJZh2LQcYXb_0,2726
+keras_hub/src/models/audio_to_text_preprocessor.py,sha256=GS-WWyJ6aSsPRxi_0bxvxA00h2mT2FEwSdAoQXAUYVI,3249
 keras_hub/src/models/backbone.py,sha256=KS2x3HFWKhEYhroUFT3uZgSkeW_48zPGqUNvxCDDIQQ,11534
 keras_hub/src/models/causal_lm.py,sha256=ReaF-i3SHsCkHh4c28jM72QjMQ8x7yiCwG39FRb-7KE,16786
 keras_hub/src/models/causal_lm_preprocessor.py,sha256=YY7VJZicdmnjDSWi9g4_pEpd5bdJK166GlWcapvokF0,6663
@@ -265,10 +267,10 @@ keras_hub/src/models/mobilenet/mobilenet_image_classifier_preprocessor.py,sha256
 keras_hub/src/models/mobilenet/mobilenet_image_converter.py,sha256=a3Ka0UYYK5wHSOjf2oMHSgofRazTAeUfttklVefq14w,360
 keras_hub/src/models/mobilenet/mobilenet_presets.py,sha256=--nhaM6LmaiCtQlZPDwoQTHW7ciU0igzS4f9ssdD9Lo,1903
 keras_hub/src/models/mobilenet/util.py,sha256=S7j4UacmVIJ3fU8cymyAoK49eHcpWIKTOyUQiEjcbzQ,721
-keras_hub/src/models/moonshine/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+keras_hub/src/models/moonshine/__init__.py,sha256=WK_9Cy1dp5KplNAaTsaJbd-2DGLsiHQsIL5ZnXuCbDQ,275
 keras_hub/src/models/moonshine/moonshine_audio_converter.py,sha256=FnvR7SP44uVOsA3g9azUhQjsVg809eJ5nqoJZQ-DAq0,11854
-keras_hub/src/models/moonshine/moonshine_audio_to_text.py,sha256=295kTM-XfUqb5mYjVSApKzMGPtnRyQdwynqqcPS7a_M,15860
-keras_hub/src/models/moonshine/moonshine_audio_to_text_preprocessor.py,sha256=TZPvyMcPJ4Ojjv7r6ZUeafssIIVtFvPPzdiRHkK9O_A,10002
+keras_hub/src/models/moonshine/moonshine_audio_to_text.py,sha256=dXFtjaxL1jpcIAiiZY1-kcNL-S4RiRJiAC2uR_a3Fyc,15865
+keras_hub/src/models/moonshine/moonshine_audio_to_text_preprocessor.py,sha256=hTw941ww8cJrP5DRrxv2DtZUNLJ9A3cayFhnsG5Ef4g,10016
 keras_hub/src/models/moonshine/moonshine_backbone.py,sha256=XtRUBe_VusXsFRk7-t1JNXM0lxp2UBOJk9v7gfTNDhA,19623
 keras_hub/src/models/moonshine/moonshine_decoder.py,sha256=Exf5Gg1gsCBST53wxOgBetKkhjS8E8QIUIlUwHlOkIY,11816
 keras_hub/src/models/moonshine/moonshine_encoder.py,sha256=NjjMO_FEBlWFSv6Appv8a3V7XovW2afvxxjXwQRgV60,8148
@@ -499,7 +501,7 @@ keras_hub/src/utils/transformers/preset_loader.py,sha256=1nfS5xVsl-JROGXJXltTqV1
 keras_hub/src/utils/transformers/safetensor_utils.py,sha256=CYUHyA4y-B61r7NDnCsFb4t_UmSwZ1k9L-8gzEd6KRg,3339
 keras_hub/tokenizers/__init__.py,sha256=uMjjm0mzUkRb0e4Ac_JK8aJ9cKGUi5UqmzWoWAFJprE,4164
 keras_hub/utils/__init__.py,sha256=jXPqVGBpJr_PpYmqD8aDG-fRMlxH-ulqCR2SZMn288Y,646
-keras_hub_nightly-0.21.0.dev202505230409.dist-info/METADATA,sha256=i-P2LqVLiVN0cIr63OvwxpAmqn2sGBPSUHDqbHiFhcg,7393
-keras_hub_nightly-0.21.0.dev202505230409.dist-info/WHEEL,sha256=zaaOINJESkSfm_4HQVc5ssNzHCPXhJm0kEUakpsEHaU,91
-keras_hub_nightly-0.21.0.dev202505230409.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
-keras_hub_nightly-0.21.0.dev202505230409.dist-info/RECORD,,
+keras_hub_nightly-0.21.0.dev202505240409.dist-info/METADATA,sha256=BJHRD68RtZc8CA6kIFWZxphjYr6g2t62j1FvwLar_LU,7393
+keras_hub_nightly-0.21.0.dev202505240409.dist-info/WHEEL,sha256=zaaOINJESkSfm_4HQVc5ssNzHCPXhJm0kEUakpsEHaU,91
+keras_hub_nightly-0.21.0.dev202505240409.dist-info/top_level.txt,sha256=N4J6piIWBKa38A4uV-CnIopnOEf8mHAbkNXafXm_CuA,10
+keras_hub_nightly-0.21.0.dev202505240409.dist-info/RECORD,,