Cntk Speed Tests
On the following images, we have plotted the time it took in our experiments to train/evaluate a 3-layer network with different minibatch sizes. In all the cases, we used 10^6 tuples. Clearly, batching tuples results in much faster training/evaluation:
- Training
- Evaluation
- DQN update rule
