Results for unet

lamblin · notoraptor · commit 86d6e1609389 · 2017-12-14T11:52:47.000-05:00
diff --git a/code/unet/train_unet.py b/code/unet/train_unet.py
@@ -386,7 +386,7 @@ def main():
     parser.add_argument('-batch_size',
                         type=int,
                         nargs='+',
-                        default=[5, 1, 1],
+                        default=[5, 5, 1],
                         help='Batch size [train, val, test]')
     parser.add_argument('-data_augmentation',
                         type=dict,
diff --git a/doc/unet.txt b/doc/unet.txt
@@ -156,6 +156,14 @@ And finally the output path (to obtain *number of classes* feature maps):
   :start-after: start-snippet-output
   :end-before: end-snippet-output
 
+Running ``train_unet.py`` on a Titan X lasted for around 60 minutes, ending with the following:
+
+.. code-block:: text
+
+    $ THEANO_FLAGS=device=cuda0,floatX=float32,dnn.conv.algo_fwd=time_once,dnn.conv.algo_bwd_data=time_once,dnn.conv.algo_bwd_filter=time_once,gpuarray.preallocate=1 python train_unet.py
+    [...]
+    EPOCH 364: Avg epoch training cost train 0.160667, cost val 0.265909, acc val 0.888796, jacc val class 0  0.636058, jacc val class 1 0.861970, jacc val 0.749014 took 4.379772 s
+
 
 References
 ++++++++++
