@@ -52,6 +52,14 @@ Note that the above methods only prevent variables being updated by SGD.
Some variables may be updated by other means,
e.g., BatchNorm statistics are updated through the `UPDATE_OPS` collection and the [RunUpdateOps](../modules/callbacks.html#tensorpack.callbacks.RunUpdateOps) callback.
## The model does not run on CPUs?
Some TensorFlow ops are not implemented on CPUs.
For example, it does not support many ops in NCHW format on CPUs.
Note that if you use MKL-enabled version of TensorFlow, it supports more NCHW ops.
In general, you need to implement the model in a way your version of TensorFlow supports.
## My training seems slow. Why?
Checkout the [Performance Tuning tutorial](performance-tuning.html)