Added a MirroredStrategy to train models on multiple GPUs if they are available

deepof/data.py

@@ -907,6 +907,7 @@ class coordinates:
         entropy_knn: int = 100,
         input_type: str = False,
         run: int = 0,
+        strategy: tf.distribute.Strategy = tf.distribute.MirroredStrategy(),
     ) -> Tuple:
         """
         Annotates coordinates using an unsupervised autoencoder.
@@ -974,6 +975,7 @@ class coordinates:
             entropy_knn=entropy_knn,
             input_type=input_type,
             run=run,
+            strategy=strategy,
         )
 
         # returns a list of trained tensorflow models

deepof/train_utils.py

@@ -306,6 +306,7 @@ def autoencoder_fitting(
     entropy_knn: int,
     input_type: str,
     run: int = 0,
+    strategy: tf.distribute.Strategy = tf.distribute.MirroredStrategy(),
 ):
     """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""
 
@@ -378,7 +379,7 @@ def autoencoder_fitting(
         return_list = (encoder, decoder, ae)
 
     else:
-
+        with strategy.scope():
             (
                 encoder,
                 generator,
@@ -388,7 +389,7 @@ def autoencoder_fitting(
                 posterior,
             ) = deepof.models.SEQ_2_SEQ_GMVAE(
                 architecture_hparams=({} if hparams is None else hparams),
-            batch_size=batch_size,
+                batch_size=batch_size * strategy.num_replicas_in_sync,
                 compile_model=True,
                 encoding=encoding_size,
                 kl_annealing_mode=kl_annealing_mode,
@@ -478,16 +479,29 @@ def autoencoder_fitting(
                 ys += [y_train[-Xs.shape[0] :]]
                 yvals += [y_val[-Xvals.shape[0] :]]
 
+            # Convert data to tf.data.Dataset objects
+            options = tf.data.Options()
+            options.experimental_distribute.auto_shard_policy = (
+                tf.data.experimental.AutoShardPolicy.DATA
+            )
+
+            train_dataset = (
+                tf.data.Dataset.from_tensor_slices((Xs, *ys))
+                .with_options(options)
+                .batch(batch_size)
+            )
+            val_dataset = (
+                tf.data.Dataset.from_tensor_slices((Xvals, *yvals))
+                .with_options(options)
+                .batch(batch_size)
+            )
+
             ae.fit(
-                x=Xs,
-                y=ys,
+                x=train_dataset,
                 epochs=epochs,
-                batch_size=batch_size,
+                batch_size=batch_size * strategy.num_replicas_in_sync,
                 verbose=1,
-                validation_data=(
-                    Xvals,
-                    yvals,
-                ),
+                validation_data=val_dataset,
                 callbacks=callbacks_,
             )