PNG summary: default to non-parallel graphs

Data parallel models may execute faster on multiple GPUs, but rendering
them creates visually complex and illegible graphs.
Therefore, when creating models for a PNG summary, we opt to use
non-parallel models.

examples/classifier_compression/compress_classifier.py

@@ -196,7 +196,8 @@ def main():
     args.dataset = 'cifar10' if 'cifar' in args.arch else 'imagenet'
 
     # Create the model
-    model = create_model(args.pretrained, args.dataset, args.arch, device_ids=args.gpus)
+    is_parallel = args.summary != 'png' # For PNG summary, parallel graphs are illegible
+    model = create_model(args.pretrained, args.dataset, args.arch, parallel=is_parallel, device_ids=args.gpus)
 
     compression_scheduler = None
     # Create a couple of logging backends.  TensorBoardLogger writes log files in a format