Demonstration of "static subgraph optimizations" feature

The subfolders in this directory contain versions of the following Chainer examples that were modified to support this feature. Since the static graph optimizations feature is experimental, these examples might be unstable.

• MNIST: This example includes two models: A version in which @static_graph was added to the model chain, and a version that demonstrates the use of "side effect" code.

• CIFAR: This is mostly unchanged from the existing CIFAR example except that @static_graph was added to the model chain.

• ptb: This example was modified so that the RNN is explicitly unrolled inside of a static chain. This was necessary because the LSTM link is only partially compatible with this feature.

You may find it interesting to run and time the examples both with and without static subgraph optimizations (i.e., by commenting out the decorator) to see the effect on runtime performance and memory usage.

Depending on the model and other hyperparameters such as the batch size, enabling graph optimizations might not always result in a speedup. An example of this is the CIFAR example, which does not run any faster with this feature enabled. This is because GPU kernels such as the convolution and multiplication operations already account for most of the execution time. As a general rule, if the GPU utilization is already close to 100%, the model is unlikely to benefit from this feature.