Add GPT-2 Model and its Variants (#354)

abheesht17 · web-flow · commit 3df60d65be4f · 2022-09-16T12:17:54.000-07:00
* Add GPT-2

* Small edit

* Fix doc-string

* Fix some typos

* Rename GPTXLarge to GPTExtraLarge

* Small changes

* Small fix

* Format

* Address comments - II

* Reverse other file changes

* Small edit

* Update doc-string

* Remove param ct
diff --git a/keras_nlp/models/__init__.py b/keras_nlp/models/__init__.py
@@ -20,6 +20,11 @@
 from keras_nlp.models.bert import BertPreprocessor
 from keras_nlp.models.bert import BertSmall
 from keras_nlp.models.bert import BertTiny
+from keras_nlp.models.gpt2 import Gpt2Base
+from keras_nlp.models.gpt2 import Gpt2Custom
+from keras_nlp.models.gpt2 import Gpt2ExtraLarge
+from keras_nlp.models.gpt2 import Gpt2Large
+from keras_nlp.models.gpt2 import Gpt2Medium
 from keras_nlp.models.roberta import RobertaBase
 from keras_nlp.models.roberta import RobertaClassifier
 from keras_nlp.models.roberta import RobertaCustom
diff --git a/keras_nlp/models/gpt2.py b/keras_nlp/models/gpt2.py
@@ -0,0 +1,286 @@
+# Copyright 2022 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""GPT-2 model configurable class, preconfigured versions, and task heads."""
+
+import tensorflow as tf
+from tensorflow import keras
+
+from keras_nlp.layers import PositionEmbedding
+from keras_nlp.layers import TransformerDecoder
+
+
+def _gpt_2_kernel_initializer(stddev=0.02):
+    return keras.initializers.RandomNormal(stddev=stddev)
+
+
+class Gpt2Custom(keras.Model):
+    """GPT-2 core network with customizable hyperparameters.
+
+    This network implements a Transformer-based decoder network,
+    Generative Pretrained Transformer-2 (GPT-2), as described in
+    ["Language Models are Unsupervised Multitask Learners"](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf).
+    It includes the embedding lookups and transformer layers.
+
+    This class gives a fully customizable GPT-2 model with any number of layers,
+    heads, and embedding dimensions. For specific GPT-2 architectures
+    defined in the paper, see, for example, `keras_nlp.models.Gpt2Base`.
+
+    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.
+        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.
+        name: string, optional. Name of the model.
+        trainable: boolean, optional. If the model's variables should be
+            trainable.
+
+    Example usage:
+    ```python
+    # Randomly initialized GPT-2 decoder
+    model = keras_nlp.models.Gpt2Custom(
+        vocabulary_size=50257,
+        num_layers=12,
+        num_heads=12,
+        hidden_dim=768,
+        intermediate_dim=3072,
+        max_sequence_length=1024,
+        name="encoder",
+    )
+
+    # Call encoder on the inputs
+    input_data = {
+        "token_ids": tf.random.uniform(
+            shape=(1, 12), dtype=tf.int64, maxval=model.vocabulary_size
+        ),
+        "padding_mask": tf.constant(
+            [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0], shape=(1, 12)
+        ),
+    }
+    output = model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        intermediate_dim,
+        dropout=0.1,
+        max_sequence_length=1024,
+        name=None,
+        trainable=True,
+    ):
+
+        # Inputs
+        token_ids = keras.Input(shape=(None,), dtype="int32", name="token_ids")
+        padding_mask = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+
+        # Embed tokens, positions.
+        token_embedding = keras.layers.Embedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            embeddings_initializer=_gpt_2_kernel_initializer(stddev=0.01),
+            name="token_embedding",
+        )(token_ids)
+
+        # Can't use `TokenAndPositionEmbedding` layer here because of different
+        # initializers.
+        position_embedding = PositionEmbedding(
+            initializer=_gpt_2_kernel_initializer(stddev=0.02),
+            sequence_length=max_sequence_length,
+            name="position_embedding",
+        )(token_embedding)
+
+        # Sum and apply dropout to embeddings.
+        x = keras.layers.Add()((token_embedding, position_embedding))
+        x = keras.layers.Dropout(
+            dropout,
+            name="embeddings_dropout",
+        )(x)
+
+        # Apply successive transformer decoder blocks.
+        for i in range(num_layers):
+            x = TransformerDecoder(
+                intermediate_dim=intermediate_dim,
+                num_heads=num_heads,
+                dropout=dropout,
+                activation=lambda x: keras.activations.gelu(
+                    x, approximate=True
+                ),
+                layer_norm_epsilon=1e-05,
+                kernel_initializer=_gpt_2_kernel_initializer(stddev=0.02),
+                normalize_first=True,
+                name=f"transformer_layer_{i}",
+            )(x, decoder_padding_mask=padding_mask)
+
+        sequence_output = keras.layers.LayerNormalization(
+            name="layer_norm",
+            axis=-1,
+            epsilon=1e-05,
+            dtype=tf.float32,
+        )(x)
+
+        # Instantiate using Functional API Model constructor
+        super().__init__(
+            inputs={
+                "token_ids": token_ids,
+                "padding_mask": padding_mask,
+            },
+            outputs=sequence_output,
+            name=name,
+            trainable=trainable,
+        )
+        # All references to `self` below this line
+        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_sequence_length = max_sequence_length
+
+    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_sequence_length": self.max_sequence_length,
+            }
+        )
+        return config
+
+
+MODEL_DOCSTRING = """GPT-2 "{type}" architecture.
+
+    This network implements a Transformer-based decoder as
+    described in
+    ["Language Models are Unsupervised Multitask Learners"](https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf).
+    It includes the embedding lookups and transformer layers.
+
+    Args:
+        vocabulary_size: int, optional. The size of the token vocabulary.
+        name: String, optional. Name of the model.
+        trainable: boolean, optional. If the model's variables should be
+            trainable.
+
+    Example usage:
+    ```python
+    # Randomly initialized Gpt2{type} encoder
+    model = keras_nlp.models.Gpt2{type}(vocabulary_size=10000)
+
+    # Call encoder on the inputs.
+    input_data = {{
+        "token_ids": tf.random.uniform(
+            shape=(1, 1024), dtype=tf.int64, maxval=model.vocabulary_size
+        ),
+        "padding_mask": tf.constant([1] * 1024, shape=(1, 1024)),
+    }}
+    output = model(input_data)
+"""
+
+
+def Gpt2Base(vocabulary_size, name=None, trainable=True):
+    return Gpt2Custom(
+        vocabulary_size=vocabulary_size,
+        num_layers=12,
+        num_heads=12,
+        hidden_dim=768,
+        intermediate_dim=3072,
+        dropout=0.1,
+        max_sequence_length=1024,
+        name=name,
+        trainable=trainable,
+    )
+
+
+def Gpt2Medium(vocabulary_size, name=None, trainable=True):
+    return Gpt2Custom(
+        vocabulary_size=vocabulary_size,
+        num_layers=24,
+        num_heads=16,
+        hidden_dim=1024,
+        intermediate_dim=4096,
+        dropout=0.1,
+        max_sequence_length=1024,
+        name=name,
+        trainable=trainable,
+    )
+
+
+def Gpt2Large(vocabulary_size, name=None, trainable=True):
+    return Gpt2Custom(
+        vocabulary_size=vocabulary_size,
+        num_layers=36,
+        num_heads=20,
+        hidden_dim=1280,
+        intermediate_dim=5120,
+        dropout=0.1,
+        max_sequence_length=1024,
+        name=name,
+        trainable=trainable,
+    )
+
+
+def Gpt2ExtraLarge(vocabulary_size, name=None, trainable=True):
+    return Gpt2Custom(
+        vocabulary_size=vocabulary_size,
+        num_layers=48,
+        num_heads=25,
+        hidden_dim=1600,
+        intermediate_dim=6400,
+        dropout=0.1,
+        max_sequence_length=1024,
+        name=name,
+        trainable=trainable,
+    )
+
+
+setattr(
+    Gpt2Base,
+    "__doc__",
+    MODEL_DOCSTRING.format(type="Base"),
+)
+setattr(
+    Gpt2Medium,
+    "__doc__",
+    MODEL_DOCSTRING.format(type="Medium"),
+)
+setattr(
+    Gpt2Large,
+    "__doc__",
+    MODEL_DOCSTRING.format(type="Large"),
+)
+setattr(
+    Gpt2ExtraLarge,
+    "__doc__",
+    MODEL_DOCSTRING.format(type="ExtraLarge"),
+)
diff --git a/keras_nlp/models/gpt2_test.py b/keras_nlp/models/gpt2_test.py
@@ -0,0 +1,92 @@
+# Copyright 2022 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for GPT-2 model."""
+
+import os
+
+import tensorflow as tf
+from absl.testing import parameterized
+from tensorflow import keras
+
+from keras_nlp.models import gpt2
+
+
+class Gpt2Test(tf.test.TestCase, parameterized.TestCase):
+    def setUp(self):
+        self.model = gpt2.Gpt2Custom(
+            vocabulary_size=1000,
+            num_layers=2,
+            num_heads=2,
+            hidden_dim=64,
+            intermediate_dim=128,
+            max_sequence_length=128,
+            name="gpt2_test",
+        )
+        self.batch_size = 8
+        self.input_batch = {
+            "token_ids": tf.ones(
+                (self.batch_size, self.model.max_sequence_length), dtype="int32"
+            ),
+            "padding_mask": tf.ones(
+                (self.batch_size, self.model.max_sequence_length), dtype="int32"
+            ),
+        }
+
+        self.input_dataset = tf.data.Dataset.from_tensor_slices(
+            self.input_batch
+        ).batch(2)
+
+    def test_valid_call_gpt2(self):
+        self.model(self.input_batch)
+
+    def test_variable_sequence_length_call_gpt2(self):
+        for seq_length in (25, 50, 75):
+            input_data = {
+                "token_ids": tf.ones(
+                    (self.batch_size, seq_length), dtype="int32"
+                ),
+                "padding_mask": tf.ones(
+                    (self.batch_size, seq_length), dtype="int32"
+                ),
+            }
+            self.model(input_data)
+
+    def test_valid_call_gpt2_base(self):
+        model = gpt2.Gpt2Base(vocabulary_size=1000, name="gpt2_base_test")
+        model(self.input_batch)
+
+    @parameterized.named_parameters(
+        ("jit_compile_false", False), ("jit_compile_true", True)
+    )
+    def test_gpt2_base_compile(self, jit_compile):
+        model = gpt2.Gpt2Base(vocabulary_size=1000, name="gpt2_base_test")
+        model.compile(jit_compile=jit_compile)
+        model.predict(self.input_batch)
+
+    @parameterized.named_parameters(
+        ("jit_compile_false", False), ("jit_compile_true", True)
+    )
+    def test_gpt2_base_compile_batched_ds(self, jit_compile):
+        model = gpt2.Gpt2Base(vocabulary_size=1000, name="gpt2_base_test")
+        model.compile(jit_compile=jit_compile)
+        model.predict(self.input_dataset)
+
+    def test_saving_model(self):
+        model_output = self.model(self.input_batch)
+        save_path = os.path.join(self.get_temp_dir(), "model")
+        self.model.save(save_path)
+        restored_model = keras.models.load_model(save_path)
+
+        restored_output = restored_model(self.input_batch)
+        self.assertAllClose(model_output, restored_output)
