Fix weight tying in TF-ESM (#22839)

Fix weight tying in ESM

src/transformers/models/esm/modeling_tf_esm.py

@@ -14,6 +14,7 @@
 # limitations under the License.
 """ PyTorch ESM model."""
 
+import os
 from typing import Optional, Tuple, Union
 
 import numpy as np
@@ -1102,6 +1103,11 @@ class TFEsmForMaskedLM(TFEsmPreTrainedModel, TFMaskedLanguageModelingLoss):
 
         self.esm = TFEsmMainLayer(config, add_pooling_layer=False, name="esm")
         self.lm_head = TFEsmLMHead(config, name="lm_head")
+        if config.tie_word_embeddings:
+            # Ensure word embeddings are built so that we actually have something to tie
+            with tf.name_scope(os.path.join(self._name_scope(), "esm", "embeddings", "word_embeddings")):
+                self.esm.embeddings.word_embeddings.build((None, None))
+            self.lm_head.decoder = self.esm.embeddings.word_embeddings.weights[0]
 
     def get_output_embeddings(self):
         return self.lm_head.decoder
@@ -1211,18 +1217,22 @@ class TFEsmLMHead(Layer):
 
         self.layer_norm = LayerNormalization(epsilon=config.layer_norm_eps, name="layer_norm")
 
-        self.decoder = Dense(
+        self.decoder = None
-            config.vocab_size,
-            use_bias=False,
-            kernel_initializer=get_initializer(config.initializer_range),
-            name="decoder",
-        )
         self.config = config
 
     def build(self, input_shape):
         super().build(input_shape)
         # Separate bias to match the PT model and allow weight cross-loading to work
         # Put it in the build so it gets the right name when adding it as a weight
+        if not self.config.tie_word_embeddings:
+            if self.decoder is not None:
+                raise ValueError("Expected decoder not to be initialized before build when not tying weights!")
+            self.decoder = self.add_weight(
+                "decoder.weight",
+                shape=(self.config.hidden_size, self.config.vocab_size),
+                initializer=get_initializer(self.config.initializer_range),
+                trainable=True,
+            )
         self.bias = self.add_weight("bias", shape=(self.config.vocab_size,), initializer="zeros", trainable=True)
 
     def get_bias(self):
@@ -1234,8 +1244,7 @@ class TFEsmLMHead(Layer):
         x = self.layer_norm(x)
 
         # project back to size of vocabulary with bias
-        x = self.decoder(x)
+        x = tf.matmul(x, self.decoder, transpose_b=True) + self.bias
-        x = x + self.bias
         return x
 
 

tests/models/esm/test_modeling_tf_esm.py

@@ -262,6 +262,24 @@ class TFEsmModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase)
     def test_save_load_after_resize_token_embeddings(self):
         pass
 
+    def test_model_common_attributes(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
+            if model_class is TFEsmForMaskedLM:
+                # Output embedding test differs from the main test because they're a matrix, not a layer
+                name = model.get_bias()
+                assert isinstance(name, dict)
+                for k, v in name.items():
+                    assert isinstance(v, tf.Variable)
+            else:
+                x = model.get_output_embeddings()
+                assert x is None
+                name = model.get_bias()
+                assert name is None
+
 
 @require_tf
 class TFEsmModelIntegrationTest(unittest.TestCase):