update example

example/cifar10/README.md

@@ -12,6 +12,7 @@ run `cifar.py`:
 | --- | --- | --- | --- |
 | cxxnet | 362 img/sec | 675 img/sec | 1282 img/sec |
 | mxnet | 420 img/sec | 804 img/sec | 1436 img/sec |
+| mxnet + cudnn v3 | 835 img/sec | 1525 img/sec | 2270 img/sec |
 
 sample output
 

example/mnist/README.md

@@ -17,9 +17,10 @@ Cortes, and Christopher J.C. Burges.
 
 Using 100 minibatch size and 20 data passes (not fine tuned.)
 
-Machine: Dual Xeon E5-2680 2.8GHz, Dual GTX 980, Intel MKL, and CUDA 7.0
+Machine: Dual Xeon E5-2680 2.8GHz, Dual GTX 980, Ubuntu 14.0, GCC 4.8. Intel MKL, and CUDA 7.0
 
 | | val accuracy | 2 x E5-2680 | 1 x GTX 980 | 2 x GTX 980 |
 | --- | ---: | ---: | ---: | ---: |
-| `mlp.py` | 97.8% | 40K img/sec | 103K img/sec | 60K img/sec |
-| `lenet.py` | 99% | 368 img/sec | 22.5K img/sec  | 33K img/sec |
+| `mlp.py` MKL + CUDA 7 | 97.8% | 40K img/sec | 103K img/sec | 60K img/sec |
+| `lenet.py` MKL + CUDA 7 | 99% | 368 img/sec | 22.5K img/sec  | 33K img/sec |
+| `lenet.py` MKL + CUDA 7 + CUDNN v3 | - | - | 19K img/sec | 29 K img/sec |

example/mnist/lenet.py

@@ -41,4 +41,5 @@
     ctx = dev, symbol = lenet, num_round = 20,
     learning_rate = 0.01, momentum = 0.9, wd = 0.00001)
 
-model.fit(X=train, eval_data=val)
+model.fit(X=train, eval_data=val,
+          epoch_end_callback=mx.callback.Speedometer(100))

make/config.mk

@@ -43,7 +43,8 @@ USE_CUDNN = 0
 
 # add the path to CUDNN libary to link and compile flag
 # if you do not need that, or do not have that, leave it as NONE
-USE_CUDNN_PATH = NONE
+# (NOTE: not enable at this moment)
+# USE_CUDNN_PATH = NONE
 
 # whether use opencv during compilation
 # you can disable it, however, you will not able to use