Add Esm

keras_hub/api/models/__init__.py

@@ -189,6 +189,22 @@
 from keras_hub.src.models.electra.electra_tokenizer import (
     ElectraTokenizer as ElectraTokenizer,
 )
+from keras_hub.src.models.esm.esm_backbone import ESMBackbone as ESM2Backbone
+from keras_hub.src.models.esm.esm_backbone import ESMBackbone as ESMBackbone
+from keras_hub.src.models.esm.esm_classifier import (
+    ESMProteinClassifier as ESMProteinClassifier,
+)
+from keras_hub.src.models.esm.esm_classifier_preprocessor import (
+    ESMProteinClassifierPreprocessor as ESMProteinClassifierPreprocessor,
+)
+from keras_hub.src.models.esm.esm_masked_plm import (
+    ESMMaskedPLM as ESM2MaskedPLM,
+)
+from keras_hub.src.models.esm.esm_masked_plm import ESMMaskedPLM as ESMMaskedPLM
+from keras_hub.src.models.esm.esm_masked_plm_preprocessor import (
+    ESMMaskedPLMPreprocessor as ESMMaskedPLMPreprocessor,
+)
+from keras_hub.src.models.esm.esm_tokenizer import ESMTokenizer as ESMTokenizer
 from keras_hub.src.models.f_net.f_net_backbone import (
     FNetBackbone as FNetBackbone,
 )

keras_hub/api/tokenizers/__init__.py

@@ -28,6 +28,7 @@
 from keras_hub.src.models.electra.electra_tokenizer import (
     ElectraTokenizer as ElectraTokenizer,
 )
+from keras_hub.src.models.esm.esm_tokenizer import ESMTokenizer as ESMTokenizer
 from keras_hub.src.models.f_net.f_net_tokenizer import (
     FNetTokenizer as FNetTokenizer,
 )

keras_hub/src/models/esm/esm_attention.py

@@ -0,0 +1,94 @@
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
+from keras_hub.src.models.roformer_v2.roformer_v2_attention import (
+    RoformerAttention,
+)
+
+
+class ESMRotaryEmbedding(RotaryEmbedding):
+    def _compute_cos_sin_embedding(self, x, position=1):
+        dim = x.shape[-1]
+        inv_freq = self.scaling_factor / (
+            self.max_wavelength ** (ops.arange(0, dim, 2, dtype=x.dtype) / dim)
+        )
+        t = ops.arange(x.shape[position], dtype=x.dtype)
+        freqs = ops.outer(t, inv_freq)
+        emb = ops.concatenate((freqs, freqs), axis=-1)
+
+        cos_emb = ops.cos(emb)[None, :, None, :]
+        sin_emb = ops.sin(emb)[None, :, None, :]
+        return cos_emb, sin_emb
+
+    def call(self, q, k, position=1):
+        cos_emb, sin_emb = self._compute_cos_sin_embedding(q, position)
+
+        return (
+            self.apply_rotary_pos_emb(q, cos_emb, sin_emb),
+            self.apply_rotary_pos_emb(k, cos_emb, sin_emb),
+        )
+
+    def rotate_half(self, x):
+        x1, x2 = ops.split(x, 2, -1)
+        return ops.concatenate((-x2, x1), axis=-1)
+
+    def apply_rotary_pos_emb(self, x, cos, sin):
+        cos = cos[:, : x.shape[1], :, :]
+        sin = sin[:, : x.shape[1], :, :]
+
+        return (x * cos) + (self.rotate_half(x) * sin)
+
+
+class EsmSelfAttention(RoformerAttention):
+    """MultiHeadAttention by ESM2
+
+    Referred to the implementation of HuggingFace.
+    In fact, this part of the calculation is exactly the same as RoFormer.
+    Only the calculation of the rotary part is different.
+    """
+
+    def __init__(self, use_rotary=True, **kwargs):
+        super().__init__(**kwargs)
+        self.use_rotary = use_rotary
+
+    def build(self, input_shape):
+        super().build(input_shape)
+        if self.use_rotary:
+            self.rotary_embedding_layer = ESMRotaryEmbedding(
+                max_wavelength=self.max_wavelength, dtype=self.dtype_policy
+            )
+            self.rotary_embedding_layer.build([])
+
+    def call(self, x, attention_mask=None):
+        qw = self.q_dense(x)
+        kw = self.k_dense(x)
+        vw = self.v_dense(x)
+
+        b, s = ops.shape(qw)[:2]
+        qw = ops.reshape(qw, (b, s, self.heads, self.head_size))
+        kw = ops.reshape(kw, (b, s, self.heads, self.head_size))
+        vw = ops.reshape(vw, (b, s, self.heads, self.head_size))
+
+        if self.use_rotary:
+            qw, kw = self.rotary_embedding_layer(qw, kw)
+        if keras.__version__ < "3.6":
+            raise ("Please make sure your Keras version is >=3.6.")
+        flash_attention = keras.config.is_flash_attention_enabled()
+        attention_mask = ops.reshape(attention_mask, [b, 1, s, 1])
+        if keras.config.backend() == "torch":
+            attention_mask = ops.repeat(attention_mask, s, -1)
+            attention_mask = ops.transpose(attention_mask, [0, 1, 3, 2])
+        o = ops.dot_product_attention(
+            qw, kw, vw, mask=attention_mask, flash_attention=flash_attention
+        )
+        return self.o_dense(ops.reshape(o, [b, s, -1]))
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "use_rotary": self.use_rotary,
+            }
+        )
+        return config

keras_hub/src/models/esm/esm_backbone.py

@@ -0,0 +1,221 @@
+import keras
+from keras import activations
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.esm.esm_encoder import ESMEncoder
+
+
+def esm2_kernel_initializer(stddev=0.02):
+    return keras.initializers.TruncatedNormal(stddev=stddev)
+
+
+@keras_hub_export(
+    ["keras_hub.models.ESM2Backbone", "keras_hub.models.ESMBackbone"]
+)
+class ESMBackbone(Backbone):
+    """A ESM2 and ESM encoder network.
+
+    This class implements a bi-directional Transformer-based encoder as
+    described in ["Roformer"](https://github.com/facebookresearch/esm).
+
+    The default constructor gives a fully customizable, randomly initialized
+    ESM2 encoder with any number of layers, heads, and embed dim.To
+    load preset architectures and weights, use the `from_preset()` constructor.
+
+
+    Args:
+        vocabulary_size: int. The size of the token vocabulary.
+        num_layers: int. The number of transformer layers.
+        num_heads: int. The number of attention heads for each transformer.
+            The hidden size must be divisible by the number of attention heads.
+        hidden_dim: int. The size of the transformer encoding and pooler layers.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            a two-layer feedforward network for each transformer.
+        dropout: float. Dropout probability for the Transformer encoder.
+                    Defaults to 0.1
+        layer_norm_eps:bool.If true, then layer norm will be used before
+                        entering the transformer block.
+                        Since it's pre-norm, the default is false.
+        max_sequence_length: int. The maximum sequence length that this encoder
+            can consume. If None, `max_sequence_length` uses the value from
+            sequence length. This determines the variable shape for positional
+            embeddings.
+        position_embedding_type:esm1 use abs position embeding,esm2 use rope.
+            so this parameter is only except for absolute and rotary.
+        dtype: None or str or .keras.mixed_precision.DTypePolicy. The dtype to
+            use for model computations and weights. Note that some computations,
+            such as softmax and layer normalization, will always be done at
+            float32 precision regardless of dtype.
+
+    Examples:
+    ```python
+    input_data = {
+        "token_ids": np.ones(shape=(1, 12), dtype="int32"),
+    }
+
+    # Pretrained ESM2 encoder.
+    model = keras_hub.models.ESM2Backbone.from_preset('hf://facebook/esm2_t6_8M_UR50D')
+    model(input_data)
+
+    # Randomly initialized ESM2 encoder with a custom config.
+    model = keras_hub.models.ESM2Backbone(
+        vocabulary_size=30552,
+        num_layers=4,
+        num_heads=4,
+        hidden_dim=256,
+        intermediate_dim=512,
+        head_size = 64,
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        intermediate_dim,
+        use_bias=True,
+        activation="gelu",
+        dropout=0.1,
+        dtype=None,
+        max_sequence_length=1024,
+        max_wavelength=10000,
+        layer_norm_eps=1e-12,
+        emb_layer_norm_before=False,
+        position_embedding_type="rotary",
+        pad_token_id=0,
+        **kwargs,
+    ):
+        if position_embedding_type not in (
+            "rotary",
+            "absolute",
+        ):
+            raise ValueError(
+                '`position_embedding_type` must be either `"rotary"`, or '
+                '`"absolute"`. Received '
+                "position_embedding_type={position_embedding_type}."
+            )
+        head_size = hidden_dim // num_heads
+        # === Layers ===
+        self.token_embedding = keras.layers.Embedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            embeddings_initializer=esm2_kernel_initializer(),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        if position_embedding_type == "absolute":
+            self.position_embedding = PositionEmbedding(
+                initializer=esm2_kernel_initializer(),
+                sequence_length=max_sequence_length,
+                dtype=dtype,
+                name="position_embedding",
+            )
+            self.embeddings_add = keras.layers.Add(
+                dtype=dtype,
+                name="embeddings_add",
+            )
+
+        self.output_layer_norm = keras.layers.LayerNormalization(
+            epsilon=layer_norm_eps,
+            dtype=dtype,
+            name="output_layer_norm",
+        )
+        if emb_layer_norm_before:
+            self.emb_layer_norm = keras.layers.LayerNormalization(
+                epsilon=layer_norm_eps,
+                dtype=dtype,
+                name="emb_layer_norm",
+            )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = ESMEncoder(
+                heads=num_heads,
+                head_size=head_size,
+                intermediate_size=intermediate_dim,
+                use_bias=use_bias,
+                max_wavelength=max_wavelength,
+                dropout=dropout,
+                activation=activation,
+                kernel_initializer=esm2_kernel_initializer(),
+                layer_norm_eps=layer_norm_eps,
+                dtype=dtype,
+                use_rotary=position_embedding_type == "rotary",
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+
+        # === Functional Model ===
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+
+        attention_mask = keras.ops.not_equal(token_id_input, pad_token_id)
+
+        token_vector = self.token_embedding(token_id_input)
+        if position_embedding_type == "absolute":
+            position_vector = self.position_embedding(
+                token_vector, start_index=pad_token_id
+            )
+            x = self.embeddings_add([token_vector, position_vector])
+        else:
+            x = token_vector
+        if emb_layer_norm_before:
+            x = self.emb_layer_norm(x)
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(x, attention_mask=attention_mask)
+        output = self.output_layer_norm(x)
+        super().__init__(
+            inputs={
+                "token_ids": token_id_input,
+            },
+            outputs=output,
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.dropout = dropout
+        self.max_wavelength = max_wavelength
+        self.head_size = head_size
+        self.dropout = dropout
+        self.activation = activations.get(activation)
+        self.use_bias = use_bias
+        self.start_token_index = 0
+        self.layer_norm_eps = layer_norm_eps
+        self.max_sequence_length = max_sequence_length
+        self.emb_layer_norm_before = emb_layer_norm_before
+        self.position_embedding_type = position_embedding_type
+        self.pad_token_id = pad_token_id
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "num_layers": self.num_layers,
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout": self.dropout,
+                "max_wavelength": self.max_wavelength,
+                "use_bias": self.use_bias,
+                "activation": activations.serialize(self.activation),
+                "layer_norm_eps": self.layer_norm_eps,
+                "emb_layer_norm_before": self.emb_layer_norm_before,
+                "position_embedding_type": self.position_embedding_type,
+                "max_sequence_length": self.max_sequence_length,
+                "pad_token_id": self.pad_token_id,
+            }
+        )
+        return config

keras_hub/src/models/esm/esm_backbone_test.py

@@ -0,0 +1,30 @@
+import keras
+from keras import ops
+
+from keras_hub.src.models.esm.esm_backbone import ESMBackbone
+from keras_hub.src.tests.test_case import TestCase
+
+
+class ESMBackboneTest(TestCase):
+    def setUp(self):
+        self.init_kwargs = {
+            "vocabulary_size": 10,
+            "num_layers": 2,
+            "num_heads": 1,
+            "hidden_dim": 2,
+            "intermediate_dim": 4,
+        }
+        self.input_data = {
+            "token_ids": ops.ones((2, 5), dtype="int32"),
+            "segment_ids": ops.zeros((2, 5), dtype="int32"),
+        }
+
+    def test_backbone_basics(self):
+        if keras.__version__ < "3.6":
+            self.skipTest("Failing on keras lower version")
+        self.run_backbone_test(
+            cls=ESMBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.input_data,
+            expected_output_shape=(2, 5, 2),
+        )