ADD: Optional strategy

ai/src/itwinai/backend/tensorflow/trainer.py

@@ -14,7 +14,7 @@ def __init__(
             optimizer,
             model_func,
             metrics_func,
-            strategy=tf.distribute.MirroredStrategy()
+            strategy
     ):
         self.strategy = strategy
         self.epochs = epochs
@@ -24,7 +24,11 @@ def __init__(
         self.optimizer = optimizer
 
         # Create distributed TF vars
-        with self.strategy.scope():
+        if self.strategy:
+            with self.strategy.scope():
+                self.model = model_func()
+                self.model.compile(loss=self.loss, optimizer=self.optimizer, metrics=metrics_func())
+        else:
             self.model = model_func()
             self.model.compile(loss=self.loss, optimizer=self.optimizer, metrics=metrics_func())
         print(f"Strategy is working with: {strategy.num_replicas_in_sync} devices")
@@ -41,8 +45,9 @@ def train(self, data):
         n_test = test.cardinality().numpy()
 
         # Distribute dataset
-        train = self.strategy.experimental_distribute_dataset(train)
-        test = self.strategy.experimental_distribute_dataset(test)
+        if self.strategy:
+            train = self.strategy.experimental_distribute_dataset(train)
+            test = self.strategy.experimental_distribute_dataset(test)
 
         # compute the steps per epoch for train and valid
         train_steps = n_train // self.batch_size

use-cases/mnist/tensorflow/env-files/tensorflow-env.yml

@@ -3,7 +3,7 @@ channels:
   - tensorflow
   - conda-forge
 dependencies:
-  - tensorflow
+  - tensorflow-gpu
   - python=3.9.12
   - pip:
       - tfx

use-cases/mnist/tensorflow/trainer.py

@@ -27,7 +27,7 @@ def __init__(
             optimizer=keras.optimizers.get(optimizer),
             model_func=lambda: model,
             metrics_func=lambda: [],
-            strategy=tf.distribute.MirroredStrategy()
+            strategy=None
         )
 
     def train(self, data):