Gradient accumulation for TFTrainer (#9585)

* gradient accumulation for tftrainer

* label naming
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* label naming
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/trainer_tf.py

@@ -638,7 +638,9 @@ class TFTrainer:
                 reduced_features = {
                     k: ft[: self.args.train_batch_size // self.args.n_replicas] for k, ft in features.items()
                 }
-                reduced_labels = labels[: self.args.train_batch_size // self.args.n_replicas]
+                reduced_labels = {
+                    k: lbl[: self.args.train_batch_size // self.args.n_replicas] for k, lbl in labels.items()
+                }
 
                 self.training_step(reduced_features, reduced_labels, nb_instances_in_global_batch)
 
@@ -650,9 +652,13 @@ class TFTrainer:
                     for k, ft in features.items()
                 }
 
-                labels = tf.concat(
-                    [labels[self.args.train_batch_size // self.args.n_replicas :], reduced_labels], axis=0
+                labels = {
+                    k: tf.concat(
+                        [lbl[self.args.train_batch_size // self.args.n_replicas :], reduced_labels[k]],
+                        axis=0,
                     )
+                    for k, lbl in labels.items()
+                }
 
             gradients = self.gradient_accumulator.gradients
             gradients = [