make bench_gen.py work for 3d conv (pytorch#12433)

Summary:
Pull Request resolved: pytorch#12433

To test 3d conv, we need to pass lists in spec argument. We also don't want to set use_cudnn=True which is the default in brew.

Reviewed By: llyfacebook, csummersea

Differential Revision: D10234315

fbshipit-source-id: 96a39992a97e020d6e9dac103e6d64df0cc1020b

binaries/bench_gen/bench_gen.py

@@ -6,6 +6,7 @@
 from __future__ import unicode_literals
 
 import argparse
+import ast
 
 from caffe2.python.model_helper import ModelHelper
 from caffe2.python.predictor import mobile_exporter
@@ -15,18 +16,15 @@
 def parse_kwarg(kwarg_str):
     key, value = kwarg_str.split('=')
     try:
-        value = int(value)
+        value = ast.literal_eval(value)
     except ValueError:
-        try:
-            value = float(value)
-        except ValueError:
-            pass
+        pass
     return key, value
 
 
 def main(args):
     # User defined keyword arguments
-    kwargs = {"order": "NCHW"}
+    kwargs = {"order": "NCHW", "use_cudnn": False}
     kwargs.update(dict(args.kwargs))
 
     model = ModelHelper(name=args.benchmark_name)