Update Transformers to 4.54

For higher Transformers versions than 4.36, models need a build()
function that ensures correctly loading of models. This commit adds
those functions.

Author: ZJaume

CHANGELOG.md

@@ -4,6 +4,11 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## Bicleaner AI 3.3.0:
+### Changed:
+- Update Transformers to 4.52
+- Add build methods to BCAI XLMR custom layers to load the models correctly.
+
 ## Bicleaner AI 3.2.0:
 ### Added:
 - Option to require GPU.

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "bicleaner-ai"
-version = "3.2.1"
+version = "3.3.0"
 license = {file = "LICENSE"}
 authors = [
     { "name" = "Prompsit Language Engineering", "email" = "[email protected]" }
@@ -19,15 +19,15 @@ dependencies = [
     "toolwrapper",
     "joblib",
     "sacremoses",
-    "bicleaner-hardrules==2.10.4",
+    "bicleaner-hardrules==2.10.6",
     "sentencepiece",
     "protobuf==3.20.3",
     "tensorflow>=2.6.5,<2.16",
     "bicleaner-ai-glove==0.2.1",
     "fuzzywuzzy",
     "python-Levenshtein",
-    "transformers==4.36.1",
-    "huggingface-hub>=0.22,<0.23",
+    "transformers==4.52.4",
+    "huggingface-hub>=0.30,<0.31",
     "zstandard",
     "psutil",
     "regex",

src/bicleaner_ai/layers.py

@@ -88,6 +88,7 @@ def __init__(self, config, **kwargs):
             kernel_initializer=get_initializer(config.initializer_range),
             name="out_proj"
         )
+        self.config = config
 
     def call(self, features, training=False):
         x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
@@ -96,3 +97,20 @@ def call(self, features, training=False):
         x = self.dropout(x, training=training)
         x = self.out_proj(x)
         return x
+
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "dense", None) is not None:
+            with tf.name_scope(self.dense.name):
+                # This was failing because our dense layer last input dimension is
+                # actually config.hidden_size and not config.head_hidden_size
+                # self.hidden_size is the original parameter and is 768, just like the last XLMR state
+                # we use a different hidden size with config.head_hidden_size (typically 2048)
+                # so even if we use config.head_hidden_size as the number of neurones in self.dense
+                # the shape of dense is (self.head_hidden_size,self.hidden_size) (2048,768)
+                self.dense.build([None, None, self.config.hidden_size])
+        if getattr(self, "out_proj", None) is not None:
+            with tf.name_scope(self.out_proj.name):
+                self.out_proj.build([None, None, self.config.head_hidden_size])

src/bicleaner_ai/models.py

@@ -646,3 +646,13 @@ def __init__(self, config, *inputs, **kwargs):
         self.classifier = BicleanerAIClassificationHead(config,
                             name=name)
 
+    def build(self, input_shape=None):
+        if self.built:
+            return
+        self.built = True
+        if getattr(self, "roberta", None) is not None:
+            with tf.name_scope(self.roberta.name):
+                self.roberta.build(None)
+        if getattr(self, "classifier", None) is not None:
+            with tf.name_scope(self.classifier.name):
+                self.classifier.build(None)