New TF loading weights (#8490)

* New TF loading weights

* apply style

* Better naming

* Largely comment the loading method

* Apply style

* Address Patrick's comments

* Remove useless line of code

* Update Docstring

* Address Sylvain's and Lysandre's comments

* Simplify the names computation

* Typos

src/transformers/modeling_tf_utils.py

@@ -236,9 +236,9 @@ class TFNextSentencePredictionLoss:
         return loss_fn(next_sentence_label, next_sentence_reduced_logits)
 
 
-def detect_tf_missing_unexpected_layers(model, resolved_archive_file):
+def load_tf_weights(model, resolved_archive_file):
     """
-    Detect missing and unexpected layers.
+    Detect missing and unexpected layers and load the TF weights accordingly to their names and shapes.
 
     Args:
         model (:obj:`tf.keras.models.Model`):
@@ -252,62 +252,60 @@ def detect_tf_missing_unexpected_layers(model, resolved_archive_file):
     missing_layers = []
     unexpected_layers = []
 
+    # Read the H5 file
     with h5py.File(resolved_archive_file, "r") as f:
-        saved_layer_names = set(hdf5_format.load_attributes_from_hdf5_group(f, "layer_names"))
-        model_layer_names = set(layer.name for layer in model.layers)
-        missing_layers = list(model_layer_names - saved_layer_names)
-        unexpected_layers = list(saved_layer_names - model_layer_names)
-
-        for layer in model.layers:
-            if layer.name in saved_layer_names:
-                g = f[layer.name]
-                saved_weight_names = hdf5_format.load_attributes_from_hdf5_group(g, "weight_names")
-                saved_weight_names_set = set(
-                    "/".join(weight_name.split("/")[2:]) for weight_name in saved_weight_names
-                )
-                symbolic_weights = layer.trainable_weights + layer.non_trainable_weights
-                symbolic_weights_names = set(
-                    "/".join(symbolic_weight.name.split("/")[2:]) for symbolic_weight in symbolic_weights
-                )
-                missing_layers.extend(list(symbolic_weights_names - saved_weight_names_set))
-                unexpected_layers.extend(list(saved_weight_names_set - symbolic_weights_names))
+        # Retrieve the name of each layer from the H5 file
+        saved_h5_model_layers_name = set(hdf5_format.load_attributes_from_hdf5_group(f, "layer_names"))
 
-    return missing_layers, unexpected_layers
-
-
-def load_tf_weights(model, resolved_archive_file):
-    """
-    Load the TF weights from a H5 file.
+        # Find the missing layers from the high level list of layers
+        missing_layers = list(set([layer.name for layer in model.layers]) - saved_h5_model_layers_name)
 
-    Args:
-        model (:obj:`tf.keras.models.Model`):
-            The model to load the weights into.
-        resolved_archive_file (:obj:`str`):
-            The location of the H5 file.
-    """
-    with h5py.File(resolved_archive_file, "r") as f:
-        saved_layer_names = set(hdf5_format.load_attributes_from_hdf5_group(f, "layer_names"))
+        # Find the unexpected layers from the high level list of layers
+        unexpected_layers = list(saved_h5_model_layers_name - set([layer.name for layer in model.layers]))
+        saved_weight_names_set = set()
+        symbolic_weights_names = set()
         weight_value_tuples = []
 
+        # Compute missing and unexpected sub layers
+        # Store the weights in list of tuples that looks like [(weight_object, value_of_weight),...]
         for layer in model.layers:
-            if layer.name in saved_layer_names:
-                g = f[layer.name]
-                saved_weight_names = hdf5_format.load_attributes_from_hdf5_group(g, "weight_names")
+            # if layer_name from the H5 file belongs to the layers from the instantiated model
+            if layer.name in saved_h5_model_layers_name:
+                # Get the H5 layer object from its name
+                h5_layer_object = f[layer.name]
+                # Get all the weights as a list from the layer object
                 symbolic_weights = layer.trainable_weights + layer.non_trainable_weights
-                saved_weight_names_values = {}
+                saved_weights = {}
 
-                for weight_name in saved_weight_names:
+                # Create a dict from the H5 saved model that looks like {"weight_name": weight_value}
+                # And a set with only the names
+                for weight_name in hdf5_format.load_attributes_from_hdf5_group(h5_layer_object, "weight_names"):
+                    # TF names always start with the model name so we ignore it
                     name = "/".join(weight_name.split("/")[1:])
-                    saved_weight_names_values[name] = np.asarray(g[weight_name])
+                    saved_weights[name] = np.asarray(h5_layer_object[weight_name])
 
+                    # Add the updated name to the final list for computing missing/unexpected values
+                    saved_weight_names_set.add(name)
+
+                # Loop over each weights from the instantiated model and compare with the weights from the H5 file
                 for symbolic_weight in symbolic_weights:
-                    splited_layers = symbolic_weight.name.split("/")[1:]
-                    symbolic_weight_name = "/".join(splited_layers)
+                    # TF names always start with the model name so we ignore it
+                    symbolic_weight_name = "/".join(symbolic_weight.name.split("/")[1:])
+
+                    # here we check if the current weight is among the weights from the H5 file
+                    # If yes, get the weight_value of the corresponding weight from the H5 file
+                    # If not, make the value to None
+                    saved_weight_value = saved_weights.get(symbolic_weight_name, None)
 
-                    if symbolic_weight_name in saved_weight_names_values:
-                        saved_weight_value = saved_weight_names_values[symbolic_weight_name]
+                    # Add the updated name to the final list for computing missing/unexpected values
+                    symbolic_weights_names.add(symbolic_weight_name)
 
+                    # If the current weight is found
+                    if saved_weight_value is not None:
+                        # Check if the shape of the current weight and the one from the H5 file are different
                         if K.int_shape(symbolic_weight) != saved_weight_value.shape:
+                            # If yes we reshape the weight from the H5 file accordingly to the current weight
+                            # If the two shapes are not compatible we raise an issue
                             try:
                                 array = np.reshape(saved_weight_value, K.int_shape(symbolic_weight))
                             except AssertionError as e:
@@ -316,10 +314,18 @@ def load_tf_weights(model, resolved_archive_file):
                         else:
                             array = saved_weight_value
 
+                        # We create the tuple that will be loaded and add it to the final list
                         weight_value_tuples.append((symbolic_weight, array))
 
+    # Load all the weights
     K.batch_set_value(weight_value_tuples)
 
+    # Compute the missing and unexpected layers
+    missing_layers.extend(list(symbolic_weights_names - saved_weight_names_set))
+    unexpected_layers.extend(list(saved_weight_names_set - symbolic_weights_names))
+
+    return missing_layers, unexpected_layers
+
 
 class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin):
     r"""
@@ -727,7 +733,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin):
         # 'by_name' allow us to do transfer learning by skipping/adding layers
         # see https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/tensorflow/python/keras/engine/network.py#L1339-L1357
         try:
-            load_tf_weights(model, resolved_archive_file)
+            missing_keys, unexpected_keys = load_tf_weights(model, resolved_archive_file)
         except OSError:
             raise OSError(
                 "Unable to load weights from h5 file. "
@@ -736,8 +742,6 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin):
 
         model(model.dummy_inputs, training=False)  # Make sure restore ops are run
 
-        missing_keys, unexpected_keys = detect_tf_missing_unexpected_layers(model, resolved_archive_file)
-
         if cls.authorized_missing_keys is not None:
             for pat in cls.authorized_missing_keys:
                 missing_keys = [k for k in missing_keys if re.search(pat, k) is None]
@@ -1033,18 +1037,18 @@ class TFSequenceSummary(tf.keras.layers.Layer):
         return output
 
 
-def shape_list(x: tf.Tensor) -> List[int]:
+def shape_list(tensor: tf.Tensor) -> List[int]:
     """
     Deal with dynamic shape in tensorflow cleanly.
 
     Args:
-        x (:obj:`tf.Tensor`): The tensor we want the shape of.
+        tensor (:obj:`tf.Tensor`): The tensor we want the shape of.
 
     Returns:
         :obj:`List[int]`: The shape of the tensor as a list.
     """
-    static = x.shape.as_list()
-    dynamic = tf.shape(x)
+    static = tensor.shape.as_list()
+    dynamic = tf.shape(tensor)
     return [dynamic[i] if s is None else s for i, s in enumerate(static)]