keras-hub-nightly 0.26.0.dev202601130438__tar.gz → 0.26.0.dev202601170438__tar.gz

{keras_hub_nightly-0.26.0.dev202601130438 → keras_hub_nightly-0.26.0.dev202601170438}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: keras-hub-nightly
-Version: 0.26.0.dev202601130438
+Version: 0.26.0.dev202601170438
 Summary: Pretrained models for Keras.
 Author-email: Keras team <keras-users@googlegroups.com>
 License-Expression: Apache-2.0

{keras_hub_nightly-0.26.0.dev202601130438 → keras_hub_nightly-0.26.0.dev202601170438}/keras_hub/api/models/__init__.py

@@ -211,6 +211,12 @@ from keras_hub.src.models.distil_bert.distil_bert_text_classifier_preprocessor i
 from keras_hub.src.models.distil_bert.distil_bert_tokenizer import (
     DistilBertTokenizer as DistilBertTokenizer,
 )
+from keras_hub.src.models.edrec.edrec_backbone import (
+    EdRecBackbone as EdRecBackbone,
+)
+from keras_hub.src.models.edrec.edrec_seq2seq_lm import (
+    EdRecSeq2SeqLM as EdRecSeq2SeqLM,
+)
 from keras_hub.src.models.efficientnet.efficientnet_backbone import (
     EfficientNetBackbone as EfficientNetBackbone,
 )

keras_hub_nightly-0.26.0.dev202601170438/keras_hub/src/models/edrec/edrec_backbone.py

@@ -0,0 +1,147 @@
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.edrec.edrec_layers import EdRecDecoderBlock
+from keras_hub.src.models.edrec.edrec_layers import EdRecEncoderBlock
+
+
+@keras_hub_export("keras_hub.models.EdRecBackbone")
+class EdRecBackbone(Backbone):
+    """EdRec Backbone model.
+
+    Args:
+        vocab_size: int, size of the vocabulary.
+        num_layers_enc: int, number of encoder layers.
+        num_layers_dec: int, number of decoder layers.
+        hidden_dim: int, hidden dimension (d_model).
+        intermediate_dim: int, intermediate dimension (d_ff).
+        num_heads: int, number of attention heads.
+        dropout: float, dropout rate.
+        epsilon: float, epsilon for simple RMSNorm.
+    """
+
+    def __init__(
+        self,
+        vocab_size,
+        num_layers_enc,
+        num_layers_dec,
+        hidden_dim,
+        intermediate_dim,
+        num_heads,
+        dropout=0.0,
+        epsilon=1e-6,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.embedding = keras.layers.Embedding(
+            input_dim=vocab_size,
+            output_dim=hidden_dim,
+            dtype=dtype,
+            name="embedding",
+        )
+        self.encoder_layers = []
+        for i in range(num_layers_enc):
+            self.encoder_layers.append(
+                EdRecEncoderBlock(
+                    hidden_dim=hidden_dim,
+                    num_heads=num_heads,
+                    intermediate_dim=intermediate_dim,
+                    dropout_rate=dropout,
+                    epsilon=epsilon,
+                    dtype=dtype,
+                    name=f"encoder_layer_{i}",
+                )
+            )
+        self.decoder_layers = []
+        for i in range(num_layers_dec):
+            self.decoder_layers.append(
+                EdRecDecoderBlock(
+                    hidden_dim=hidden_dim,
+                    num_heads=num_heads,
+                    intermediate_dim=intermediate_dim,
+                    dropout_rate=dropout,
+                    epsilon=epsilon,
+                    dtype=dtype,
+                    name=f"decoder_layer_{i}",
+                )
+            )
+
+        # === Functional Model ===
+        encoder_token_ids = keras.Input(
+            shape=(None,), dtype="int32", name="encoder_token_ids"
+        )
+        decoder_token_ids = keras.Input(
+            shape=(None,), dtype="int32", name="decoder_token_ids"
+        )
+        encoder_padding_mask = keras.Input(
+            shape=(None,), dtype="bool", name="encoder_padding_mask"
+        )
+        decoder_padding_mask = keras.Input(
+            shape=(None,), dtype="bool", name="decoder_padding_mask"
+        )
+
+        # Encoder
+        x_enc = self.embedding(encoder_token_ids)
+
+        for layer in self.encoder_layers:
+            x_enc = layer(
+                x_enc,
+                padding_mask=encoder_padding_mask,
+            )
+
+        # Decoder
+        x_dec = self.embedding(decoder_token_ids)
+        for layer in self.decoder_layers:
+            x_dec, _, _ = layer(
+                x_dec,
+                encoder_outputs=x_enc,
+                decoder_padding_mask=decoder_padding_mask,
+                encoder_padding_mask=encoder_padding_mask,
+            )
+
+        super().__init__(
+            inputs={
+                "encoder_token_ids": encoder_token_ids,
+                "decoder_token_ids": decoder_token_ids,
+                "encoder_padding_mask": encoder_padding_mask,
+                "decoder_padding_mask": decoder_padding_mask,
+            },
+            outputs={
+                "encoder_sequence_output": x_enc,
+                "decoder_sequence_output": x_dec,
+            },
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocab_size = vocab_size
+        self.num_layers_enc = num_layers_enc
+        self.num_layers_dec = num_layers_dec
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.epsilon = epsilon
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocab_size": self.vocab_size,
+                "num_layers_enc": self.num_layers_enc,
+                "num_layers_dec": self.num_layers_dec,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "num_heads": self.num_heads,
+                "dropout": self.dropout,
+                "epsilon": self.epsilon,
+            }
+        )
+        return config
+
+    @property
+    def token_embedding(self):
+        return self.embedding

keras_hub_nightly-0.26.0.dev202601170438/keras_hub/src/models/edrec/edrec_layers.py

@@ -0,0 +1,434 @@
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.cached_multi_head_attention import (
+    CachedMultiHeadAttention,
+)
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    compute_causal_mask,
+)
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    merge_padding_and_attention_mask,
+)
+
+
+class EdRecRMSNormalization(keras.layers.Layer):
+    """RMSNorm layer that matches JAX EdRec implementation.
+
+    Attributes:
+        epsilon: float, epsilon value for numerical stability.
+    """
+
+    def __init__(self, epsilon=1e-6, **kwargs):
+        super().__init__(**kwargs)
+        self.epsilon = epsilon
+
+    def build(self, input_shape):
+        self.scale = self.add_weight(
+            name="scale",
+            shape=(input_shape[-1],),
+            initializer="ones",
+            trainable=True,
+        )
+        super().build(input_shape)
+
+    def call(self, x):
+        # JAX: rms = jnp.sqrt(jnp.mean(jnp.square(x), axis=-1, keepdims=True)
+        #                     + self.eps)
+        # JAX: normed = x / rms
+        # JAX: normed = normed * (1 + scale)
+
+        # Standard RMSNorm is x * scale / rms.
+        # EdRec RMSNorm is x * (1 + scale) / rms.
+        # Note: If scale is initialized to ones, (1+scale) starts at 2.
+
+        mean_square = ops.mean(ops.square(x), axis=-1, keepdims=True)
+        rms = ops.sqrt(mean_square + self.epsilon)
+        normed = x / rms
+        return normed * ops.cast(1.0 + self.scale, x.dtype)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update({"epsilon": self.epsilon})
+        return config
+
+
+class EdRecGatedFeedForward(keras.layers.Layer):
+    """Gated FeedForward (GLU-style) layer.
+
+    y = GELU(up_proj(x)) * gate_proj(x)
+    y = down_proj(y)
+    """
+
+    def __init__(
+        self,
+        intermediate_dim,
+        hidden_dim,
+        dropout_rate=0.0,
+        activation="gelu",
+        kernel_initializer="glorot_uniform",
+        bias_initializer="zeros",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.intermediate_dim = intermediate_dim
+        self.hidden_dim = hidden_dim  # The output dimension (d_model)
+        self.dropout_rate = dropout_rate
+        self.activation = activation
+        self.kernel_initializer = kernel_initializer
+        self.bias_initializer = bias_initializer
+
+    def build(self, input_shape):
+        self.up_proj = keras.layers.Dense(
+            self.intermediate_dim,
+            use_bias=False,
+            kernel_initializer=self.kernel_initializer,
+            dtype=self.dtype_policy,
+            name="up_proj",
+        )
+        self.gate_proj = keras.layers.Dense(
+            self.intermediate_dim,
+            use_bias=False,
+            kernel_initializer=self.kernel_initializer,
+            dtype=self.dtype_policy,
+            name="gate_proj",
+        )
+        self.down_proj = keras.layers.Dense(
+            self.hidden_dim,
+            use_bias=False,
+            kernel_initializer=self.kernel_initializer,
+            dtype=self.dtype_policy,
+            name="down_proj",
+        )
+        self.dropout = keras.layers.Dropout(
+            self.dropout_rate, dtype=self.dtype_policy, name="dropout"
+        )
+
+    def call(self, x, training=False):
+        # Up projection + activation (GELU)
+        h = self.up_proj(x)
+        if self.activation == "gelu":
+            h = keras.activations.gelu(h, approximate=True)
+        else:
+            h = keras.activations.get(self.activation)(h)
+
+        # Gate projection
+        g = self.gate_proj(x)
+
+        # Elementwise gating
+        y = h * g
+
+        # Down projection
+        y = self.down_proj(y)
+
+        # Dropout
+        if self.dropout_rate > 0.0:
+            y = self.dropout(y, training=training)
+
+        return y
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "intermediate_dim": self.intermediate_dim,
+                "hidden_dim": self.hidden_dim,
+                "dropout_rate": self.dropout_rate,
+                "activation": self.activation,
+                "kernel_initializer": self.kernel_initializer,
+                "bias_initializer": self.bias_initializer,
+            }
+        )
+        return config
+
+
+class EdRecEncoderBlock(keras.layers.Layer):
+    """EdRec Encoder Block.
+
+    Pre-norm: x = x + Dropout(Attention(RMSNorm(x))) x = x +
+    GatedFeedForward(RMSNorm(x))
+    """
+
+    def __init__(
+        self,
+        hidden_dim,
+        num_heads,
+        intermediate_dim,
+        dropout_rate=0.0,
+        epsilon=1e-6,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.hidden_dim = hidden_dim
+        self.num_heads = num_heads
+        self.intermediate_dim = intermediate_dim
+        self.dropout_rate = dropout_rate
+        self.epsilon = epsilon
+        self.head_dim = hidden_dim // num_heads
+
+    def build(self, input_shape):
+        self.pre_attention_norm = EdRecRMSNormalization(
+            epsilon=self.epsilon,
+            dtype=self.dtype_policy,
+            name="pre_attention_norm",
+        )
+        self.attention = keras.layers.MultiHeadAttention(
+            num_heads=self.num_heads,
+            key_dim=self.head_dim,
+            use_bias=False,
+            output_shape=self.hidden_dim,
+            dtype=self.dtype_policy,
+            name="attention",
+        )
+        self.dropout1 = keras.layers.Dropout(
+            self.dropout_rate, dtype=self.dtype_policy, name="dropout1"
+        )
+
+        self.pre_ffw_norm = EdRecRMSNormalization(
+            epsilon=self.epsilon, dtype=self.dtype_policy, name="pre_ffw_norm"
+        )
+        self.mlp = EdRecGatedFeedForward(
+            intermediate_dim=self.intermediate_dim,
+            hidden_dim=self.hidden_dim,
+            dropout_rate=self.dropout_rate,
+            dtype=self.dtype_policy,
+            name="mlp",
+        )
+
+    def call(self, x, padding_mask=None, training=False):
+        # Self Attention
+        residual = x
+        x_norm = self.pre_attention_norm(x)
+
+        # padding_mask is [B, L]
+        # We need to expand it to [B, 1, 1, L] for broadcasting against
+        # [B, H, L, L]
+        if padding_mask is not None:
+            padding_mask = merge_padding_and_attention_mask(
+                x, padding_mask, None
+            )
+
+        attn_out = self.attention(
+            query=x_norm,
+            value=x_norm,
+            attention_mask=padding_mask,
+            training=training,
+        )
+        attn_out = self.dropout1(attn_out, training=training)
+        x = residual + attn_out
+
+        # Feed Forward
+        residual = x
+        ff_norm = self.pre_ffw_norm(x)
+        ff_out = self.mlp(ff_norm, training=training)
+        x = residual + ff_out
+
+        return x
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "hidden_dim": self.hidden_dim,
+                "num_heads": self.num_heads,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout_rate": self.dropout_rate,
+                "epsilon": self.epsilon,
+            }
+        )
+        return config
+
+
+class EdRecDecoderBlock(keras.layers.Layer):
+    """EdRec Decoder Block.
+
+    x = x + Dropout(SelfAttention(RMSNorm(x)))
+    x = x + Dropout(CrossAttention(RMSNorm(x), encoder_outputs))
+    x = x + GatedFeedForward(RMSNorm(x))
+    """
+
+    def __init__(
+        self,
+        hidden_dim,
+        num_heads,
+        intermediate_dim,
+        dropout_rate=0.0,
+        epsilon=1e-6,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.hidden_dim = hidden_dim
+        self.num_heads = num_heads
+        self.intermediate_dim = intermediate_dim
+        self.dropout_rate = dropout_rate
+        self.epsilon = epsilon
+        self.head_dim = hidden_dim // num_heads
+
+    def build(self, input_shape):
+        self.pre_self_attn_norm = EdRecRMSNormalization(
+            epsilon=self.epsilon,
+            dtype=self.dtype_policy,
+            name="pre_self_attn_norm",
+        )
+        self.self_attention = CachedMultiHeadAttention(
+            num_heads=self.num_heads,
+            key_dim=self.head_dim,
+            use_bias=False,
+            output_shape=self.hidden_dim,
+            dtype=self.dtype_policy,
+            name="self_attention",
+        )
+        self.dropout1 = keras.layers.Dropout(
+            self.dropout_rate, dtype=self.dtype_policy, name="dropout1"
+        )
+
+        self.pre_cross_attn_norm = EdRecRMSNormalization(
+            epsilon=self.epsilon,
+            dtype=self.dtype_policy,
+            name="pre_cross_attn_norm",
+        )
+        self.cross_attention = CachedMultiHeadAttention(
+            num_heads=self.num_heads,
+            key_dim=self.head_dim,
+            use_bias=False,
+            output_shape=self.hidden_dim,
+            dtype=self.dtype_policy,
+            name="cross_attention",
+        )
+        self.dropout2 = keras.layers.Dropout(
+            self.dropout_rate, dtype=self.dtype_policy, name="dropout2"
+        )
+
+        self.pre_ffw_norm = EdRecRMSNormalization(
+            epsilon=self.epsilon, dtype=self.dtype_policy, name="pre_ffw_norm"
+        )
+        self.mlp = EdRecGatedFeedForward(
+            intermediate_dim=self.intermediate_dim,
+            hidden_dim=self.hidden_dim,
+            dropout_rate=self.dropout_rate,
+            dtype=self.dtype_policy,
+            name="mlp",
+        )
+
+    def call(
+        self,
+        x,
+        encoder_outputs,
+        decoder_padding_mask=None,
+        encoder_padding_mask=None,
+        self_attention_cache=None,
+        self_attention_cache_update_index=None,
+        cross_attention_cache=None,
+        cross_attention_cache_update_index=None,
+        use_causal_mask=True,
+        training=False,
+    ):
+        # Self Attention
+        residual = x
+        x_norm = self.pre_self_attn_norm(x)
+
+        batch_size = ops.shape(x)[0]
+        input_length = ops.shape(x)[1]
+
+        total_length = input_length
+        if self_attention_cache is not None:
+            total_length = ops.shape(self_attention_cache)[2]
+
+        # Compute causal mask
+        causal_mask = None
+        if use_causal_mask:
+            causal_mask = compute_causal_mask(
+                batch_size,
+                total_length,
+                input_length,
+                0
+                if self_attention_cache_update_index is None
+                else self_attention_cache_update_index,
+            )
+
+        # Merge with padding mask
+        self_attn_mask = causal_mask
+        if decoder_padding_mask is not None:
+            # decoder_padding_mask is [B, L_dec]
+            # merge_padding_and_attention_mask gives [B, 1, L, L]
+            padding_mask_merged = merge_padding_and_attention_mask(
+                x, decoder_padding_mask, None
+            )
+
+            if causal_mask is not None:
+                self_attn_mask = ops.minimum(padding_mask_merged, causal_mask)
+            else:
+                self_attn_mask = padding_mask_merged
+
+        self_attn_out = self.self_attention(
+            query=x_norm,
+            value=x_norm,
+            attention_mask=self_attn_mask,
+            cache=self_attention_cache,
+            cache_update_index=self_attention_cache_update_index,
+            training=training,
+        )
+
+        if self_attention_cache is not None:
+            self_attn_out, self_attention_cache = self_attn_out
+
+        self_attn_out = self.dropout1(self_attn_out, training=training)
+        x = residual + self_attn_out
+
+        # Cross Attention
+        residual = x
+        x_norm = self.pre_cross_attn_norm(x)
+
+        cross_mask = None
+        if encoder_padding_mask is not None:
+            cross_mask = merge_padding_and_attention_mask(
+                encoder_outputs, encoder_padding_mask, None
+            )
+
+        cross_attn_out = self.cross_attention(
+            query=x_norm,
+            value=encoder_outputs,
+            attention_mask=cross_mask,
+            cache=cross_attention_cache,
+            cache_update_index=cross_attention_cache_update_index,
+            training=training,
+        )
+
+        if cross_attention_cache is not None:
+            cross_attn_out, cross_attention_cache = cross_attn_out
+
+        cross_attn_out = self.dropout2(cross_attn_out, training=training)
+        x = residual + cross_attn_out
+
+        # Feed Forward
+        residual = x
+        ff_norm = self.pre_ffw_norm(x)
+        ff_out = self.mlp(ff_norm, training=training)
+        x = residual + ff_out
+
+        if self_attention_cache is not None:
+            if cross_attention_cache is not None:
+                return x, self_attention_cache, cross_attention_cache
+            return (
+                x,
+                self_attention_cache,
+                ops.zeros((0,), dtype=self.compute_dtype),
+            )
+        return (
+            x,
+            ops.zeros((0,), dtype=self.compute_dtype),
+            ops.zeros((0,), dtype=self.compute_dtype),
+        )
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "hidden_dim": self.hidden_dim,
+                "num_heads": self.num_heads,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout_rate": self.dropout_rate,
+                "epsilon": self.epsilon,
+            }
+        )
+        return config