Port bart transformer checkpoint

keras_nlp/src/utils/transformers/convert.py

@@ -18,6 +18,8 @@
 from keras_nlp.src.utils.transformers.convert_albert import (
     load_albert_tokenizer,
 )
+from keras_nlp.src.utils.transformers.convert_bart import load_bart_backbone
+from keras_nlp.src.utils.transformers.convert_bart import load_bart_tokenizer
 from keras_nlp.src.utils.transformers.convert_bert import load_bert_backbone
 from keras_nlp.src.utils.transformers.convert_bert import load_bert_tokenizer
 from keras_nlp.src.utils.transformers.convert_distilbert import (
@@ -70,6 +72,8 @@ def load_transformers_backbone(cls, preset, load_weights):
         return load_distilbert_backbone(cls, preset, load_weights)
     if cls.__name__ == "AlbertBackbone":
         return load_albert_backbone(cls, preset, load_weights)
+    if cls.__name__ == "BartBackbone":
+        return load_bart_backbone(cls, preset, load_weights)
     raise ValueError(
         f"{cls} has not been ported from the Hugging Face format yet. "
         "Please check Hugging Face Hub for the Keras model. "
@@ -103,6 +107,8 @@ def load_transformers_tokenizer(cls, preset):
         return load_distilbert_tokenizer(cls, preset)
     if cls.__name__ == "AlbertTokenizer":
         return load_albert_tokenizer(cls, preset)
+    if cls.__name__ == "BartTokenizer":
+        return load_bart_tokenizer(cls, preset)
     raise ValueError(
         f"{cls} has not been ported from the Hugging Face format yet. "
         "Please check Hugging Face Hub for the Keras model. "

keras_nlp/src/utils/transformers/convert_bart.py

@@ -0,0 +1,372 @@
+# Copyright 2024 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.
+import numpy as np
+
+from keras_nlp.src.utils.preset_utils import HF_CONFIG_FILE
+from keras_nlp.src.utils.preset_utils import get_file
+from keras_nlp.src.utils.preset_utils import jax_memory_cleanup
+from keras_nlp.src.utils.preset_utils import load_config
+from keras_nlp.src.utils.transformers.safetensor_utils import SafetensorLoader
+
+
+def convert_backbone_config(transformers_config):
+    return {
+        "vocabulary_size": transformers_config["vocab_size"],
+        "num_layers": transformers_config["num_hidden_layers"],
+        "num_heads": transformers_config["encoder_attention_heads"],
+        "hidden_dim": transformers_config["d_model"],
+        "intermediate_dim": transformers_config["encoder_ffn_dim"],
+        "dropout": transformers_config["dropout"],
+        "max_sequence_length": transformers_config["max_position_embeddings"],
+    }
+
+
+def convert_weights(backbone, loader):
+    # Embeddings
+    loader.port_weight(
+        keras_variable=backbone.token_embedding.embeddings,
+        hf_weight_key="shared.weight",
+    )
+
+    # Encoder blocks
+    for index in range(backbone.num_layers):
+        encoder_layer = backbone.encoder_transformer_layers[index]
+        encoder_self_attention = encoder_layer._self_attention_layer
+        hf_encoder_prefix = f"encoder.layers.{index}"
+
+        # Norm layers
+        loader.port_weight(
+            keras_variable=encoder_layer._self_attention_layer_norm.gamma,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn_layer_norm.weight",
+        )
+        loader.port_weight(
+            keras_variable=encoder_layer._self_attention_layer_norm.beta,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn_layer_norm.bias",
+        )
+        loader.port_weight(
+            keras_variable=encoder_layer._feedforward_layer_norm.gamma,
+            hf_weight_key=f"{hf_encoder_prefix}.final_layer_norm.weight",
+        )
+        loader.port_weight(
+            keras_variable=encoder_layer._feedforward_layer_norm.beta,
+            hf_weight_key=f"{hf_encoder_prefix}.final_layer_norm.bias",
+        )
+
+        # Self Attention layers
+        # Query
+        loader.port_weight(
+            keras_variable=encoder_self_attention.query_dense.kernel,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.q_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=encoder_self_attention.query_dense.bias,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.q_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Key
+        loader.port_weight(
+            keras_variable=encoder_self_attention.key_dense.kernel,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.k_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=encoder_self_attention.key_dense.bias,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.k_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Value
+        loader.port_weight(
+            keras_variable=encoder_self_attention.value_dense.kernel,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.v_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=encoder_self_attention.value_dense.bias,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.v_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Output
+        loader.port_weight(
+            keras_variable=encoder_self_attention.output_dense.kernel,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.out_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=encoder_self_attention.output_dense.bias,
+            hf_weight_key=f"{hf_encoder_prefix}.self_attn.out_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # MLP layers
+        loader.port_weight(
+            keras_variable=encoder_layer._feedforward_intermediate_dense.kernel,
+            hf_weight_key=f"{hf_encoder_prefix}.fc1.weight",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+        loader.port_weight(
+            keras_variable=encoder_layer._feedforward_intermediate_dense.bias,
+            hf_weight_key=f"{hf_encoder_prefix}.fc1.bias",
+        )
+        loader.port_weight(
+            keras_variable=encoder_layer._feedforward_output_dense.kernel,
+            hf_weight_key=f"{hf_encoder_prefix}.fc2.weight",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+        loader.port_weight(
+            keras_variable=encoder_layer._feedforward_output_dense.bias,
+            hf_weight_key=f"{hf_encoder_prefix}.fc2.bias",
+        )
+
+    # Decoder blocks
+    for index in range(backbone.num_layers):
+        decoder_layer = backbone.decoder_transformer_layers[index]
+        decoder_self_attention = decoder_layer._self_attention_layer
+        decoder_cross_attention = decoder_layer._cross_attention_layer
+        hf_decoder_prefix = f"decoder.layers.{index}"
+
+        # Norm layers
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer_norm.gamma,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn_layer_norm.weight",
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._self_attention_layer_norm.beta,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn_layer_norm.bias",
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_layer_norm.gamma,
+            hf_weight_key=f"{hf_decoder_prefix}.final_layer_norm.weight",
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_layer_norm.beta,
+            hf_weight_key=f"{hf_decoder_prefix}.final_layer_norm.bias",
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._cross_attention_layer_norm.gamma,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn_layer_norm.weight",
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._cross_attention_layer_norm.beta,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn_layer_norm.bias",
+        )
+
+        # Self Attention layers
+        # Query
+        loader.port_weight(
+            keras_variable=decoder_self_attention.query_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.q_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_self_attention.query_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.q_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Key
+        loader.port_weight(
+            keras_variable=decoder_self_attention.key_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.k_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_self_attention.key_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.k_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Value
+        loader.port_weight(
+            keras_variable=decoder_self_attention.value_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.v_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_self_attention.value_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.v_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Output
+        loader.port_weight(
+            keras_variable=decoder_self_attention.output_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.out_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_self_attention.output_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.self_attn.out_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # MLP layers
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_intermediate_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.fc1.weight",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_intermediate_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.fc1.bias",
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_output_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.fc2.weight",
+            hook_fn=lambda hf_tensor, _: np.transpose(hf_tensor, axes=(1, 0)),
+        )
+        loader.port_weight(
+            keras_variable=decoder_layer._feedforward_output_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.fc2.bias",
+        )
+
+        # Cross Attention Layers
+        # Query
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.query_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.q_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.query_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.q_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Key
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.key_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.k_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.key_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.k_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Value
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.value_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.v_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.value_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.v_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+        # Output
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.output_dense.kernel,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.out_proj.weight",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+        loader.port_weight(
+            keras_variable=decoder_cross_attention.output_dense.bias,
+            hf_weight_key=f"{hf_decoder_prefix}.encoder_attn.out_proj.bias",
+            hook_fn=lambda hf_tensor, keras_shape: np.reshape(
+                np.transpose(hf_tensor), keras_shape
+            ),
+        )
+
+    # Normalization
+    loader.port_weight(
+        keras_variable=backbone.encoder_embeddings_layer_norm.gamma,
+        hf_weight_key="encoder.layernorm_embedding.weight",
+    )
+    loader.port_weight(
+        keras_variable=backbone.encoder_embeddings_layer_norm.beta,
+        hf_weight_key="encoder.layernorm_embedding.bias",
+    )
+    loader.port_weight(
+        keras_variable=backbone.decoder_embeddings_layer_norm.gamma,
+        hf_weight_key="decoder.layernorm_embedding.weight",
+    )
+    loader.port_weight(
+        keras_variable=backbone.decoder_embeddings_layer_norm.beta,
+        hf_weight_key="decoder.layernorm_embedding.bias",
+    )
+
+    return backbone
+
+
+def load_bart_backbone(cls, preset, load_weights):
+    transformers_config = load_config(preset, HF_CONFIG_FILE)
+    keras_config = convert_backbone_config(transformers_config)
+    backbone = cls(**keras_config)
+    if load_weights:
+        jax_memory_cleanup(backbone)
+        with SafetensorLoader(preset) as loader:
+            convert_weights(backbone, loader)
+    return backbone
+
+
+def load_bart_tokenizer(cls, preset):
+    vocab_file = get_file(preset, "vocab.json")
+    merges_file = get_file(preset, "merges.txt")
+    return cls(
+        vocabulary=vocab_file,
+        merges=merges_file,
+    )