MultiGPU version of "mnist-keras.py"

examples/keras/README.md

@@ -15,9 +15,14 @@ are the only two tools I know that can scale the training of a large Keras model
 
 ### Simple Examples:
 
-[mnist-keras.py](mnist-keras.py): a simple MNIST model written mostly in tensorpack style, but use Keras model as symbolic functions.
+There are two flavors where you can use a Keras model inside tensorpack:
 
-[mnist-keras-v2.py](mnist-keras-v2.py): the same MNIST model written in Keras style.
+1. Write the tower function similar to a standard tensorpack program, but use some Keras layers in
+	 between. See [mnist-keras.py](mnist-keras.py) on how to do this.
+	 It does not support all tensorpack trainers.
+
+2. The entire model to train is a Keras model (and there will be no `ModelDesc`, etc).
+	 See [mnist-keras-v2.py](mnist-keras-v2.py).
 
 ### ImageNet Example:
 

examples/keras/mnist-keras.py

@@ -10,6 +10,7 @@ from tensorpack import *
 from tensorpack.contrib.keras import KerasPhaseCallback
 from tensorpack.dataflow import dataset
 from tensorpack.utils.argtools import memoized
+from tensorpack.utils.gpu import get_num_gpu
 
 KL = keras.layers
 
@@ -23,8 +24,9 @@ Note: this example does not work for replicated-style data-parallel trainers.
 IMAGE_SIZE = 28
 
 
-@memoized        # this is necessary for sonnet/Keras to work under tensorpack
+@memoized  # this is necessary for sonnet/keras to work under tensorpack
 def get_keras_model():
+    with tf.name_scope('/'):
         M = keras.models.Sequential()
         M.add(KL.Conv2D(32, 3, activation='relu', input_shape=[IMAGE_SIZE, IMAGE_SIZE, 1], padding='same'))
         M.add(KL.MaxPooling2D())
@@ -96,4 +98,11 @@ if __name__ == '__main__':
         max_epoch=100,
     )
 
+    if get_num_gpu() <= 1:
+        # single GPU:
         launch_train_with_config(cfg, QueueInputTrainer())
+    else:
+        # multi GPU:
+        launch_train_with_config(cfg, SyncMultiGPUTrainerParameterServer(2))
+        # "Replicated" multi-gpu trainer is not supported for Keras model
+        # since Keras does not respect variable scopes.

tensorpack/contrib/keras.py

@@ -40,7 +40,7 @@ class KerasModelCaller(object):
         self.get_model = get_model
         self.cached_model = None
 
-    def __call__(self, input_tensors):
+    def __call__(self, *input_tensors):
         """
         Args:
             input_tensors ([tf.Tensor])
@@ -106,6 +106,8 @@ class KerasModelCaller(object):
 
         post_process_model(model)
 
+        if isinstance(outputs, list) and len(outputs) == 1:
+            return outputs[0]
         return outputs
 
 
@@ -167,7 +169,7 @@ def setup_keras_trainer(
         target_tensors = list(inputs[nr_inputs:])
         # TODO mapping between target tensors & output tensors
 
-        outputs = model_caller(input_tensors)
+        outputs = model_caller(*input_tensors)
 
         if isinstance(outputs, tf.Tensor):
             outputs = [outputs]