note bug in cifar10 full with CPU computation

diff --git a/examples/cifar10/cifar10_full_solver.prototxt b/examples/cifar10/cifar10_full_solver.prototxt
index b985b65..cd62790 100644 (file)
--- a/examples/cifar10/cifar10_full_solver.prototxt
+++ b/examples/cifar10/cifar10_full_solver.prototxt
@@ -25,4 +25,9 @@ max_iter: 60000
 snapshot: 10000
 snapshot_prefix: "cifar10_full"
 # solver mode: 0 for CPU and 1 for GPU
+# Note: there seems to be a bug with CPU computation in the pooling layers,
+# and changing to solver_mode: 0 (for CPU computation) may result in NaNs
+# on this example.  If you want to train a variant of this architecture on the
+# CPU, try changing the pooling regions from WITHIN_CHANNEL to ACROSS_CHANNELS
+# in both cifar_full_train.prototxt and cifar_full_test.prototxt.
 solver_mode: 1