MyCaffe 0.10.1.221-beta1

CUDA 10.1.243, cuDNN 7.6.3, nvapi 430, Native Caffe up to 10/24/2018, Windows 10-1903, Driver 430.86

MyCaffe[1] (a complete C# re-write of CAFFE[2]) now supports Deep Q-Learning[3][4] with a NoisyNet[5] and Prioritized Replay Buffer[6], all supported by the new DQN trainer and do so with the newly released CUDA 10.1.243, CuDNN 7.6.3 and the dual RNN/RL training that allows multi-pass training where the first pass involves RNN training and the second pass involves RL training that uses the already trained RNN side of the model.

IMPORTANT NOTE: When using TCC mode, we recommend that ALL headless GPU’s are placed in TCC mode for we have experienced stability issues when using a mix of TCC and WDM modes with headless GPU’s.

REQUIRED SOFTWARE: 1.) Install NVIDIA CUDA 10.1.243 which you can download from https://developer.nvidia.com/cuda-downloads 2.) Install NVIDIA cuDNN 7.6.3 which you can download from https://developer.nvidia.com/cudnn 3.) Download and install Microsoft SQL Express 2016 (or later).

This release of the MyCaffe AI Platform and Test Applications has the following new additions: • CUDA 10.1.243/cuDNN 7.6.3 supported (with driver 430.86 or above). • Windows 1903, OS Build 18362.356 now supported. • Added CudaDnn.matrix_mean support. • Added CudaDnn.matrix_stdev support. • Added CudaDnn.matrix_correlation support. • Added CudaDnn.mulbsx matrix vector multiplication support. • Added CudaDnn.divbsx matrix vector division support. • Added new CudaDnn.permute support. • Added new Normalization Layer for SSD. • Added new PriorBox Layer for SSD. • Added new MultiBoxLoss Layer for SSD. • Added legacy compute 3.5 support to cuDNN 10.1 version of CudaDnn DLL. • Added compute checks to the MyCaffe Test Application. • Added data load control to Image Database allowing optional loading of Image Criteria and Debug data. • Added new LoadLite method to MyCaffeControl to load model and solver without MyCaffeImageDatabase.

The following bug fixes are in this release: • Fixed bugs related to allocating a large number of items.

Easily run Neural Style, train Deep Q-Learning[3][4] or Policy Gradient[1] models to beat Pong or Cart-Pole, or create the CIFAR-10 and MNIST datasets using the MyCaffe Test Application which you can download from the MyCaffe GitHub site.

Create and train the Deep Q-Learning[3][4], Policy Gradient[1], Neural Style Transfer, Recurrent Learning, Policy Gradient Reinforcement Learning, Auto-Encoder, DANN and ResNet models by following step-by-step instructions in the SignalPop Tutorials. And, to see other cool examples that show what MyCaffe can do, see the SignalPop Examples.

If you would like to visually design, develop, test and debug your models, see the SignalPop AI Designer specifically designed to enhance your MyCaffe deep learning.

Also, check out the SignalPop Universal Miner that not only keeps your GPU's cool as you train, but also gives you detailed information on each of your GPU's (such as temperature, fan speed, overclock, and usage), and allows you to easily mine Ethereum. When not training AI, put those GPU's to use making some Ether - never let a good GPU go to waste!

Happy ‘deep’ learning!

[1] MyCaffe: A Complete C# Re-Write of Caffe with Reinforcement Learning by D. Brown, 2018.

[2] Caffe: Convolutional Architecture for Fast Feature Embedding by Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, Trevor Darrell, 2014, arXiv:1408.5093

[3] GitHub: Google/dopamine licensed under the Apache 2.0 License;

[4] Dopamine: A Research Framework for Deep Reinforcement Learning by Pablo Samuel Castro, Subhodeep Moitra, Carles Gelada, Saurabh Kumar, Marc G. Bellemare, 2018, arXiv:1812.06110

[5] Noisy Networks for Exploration by Meire Fortunato, Mohammad Gheshlaghi Azar, Bilal Piot, Jacob Menick, Ian Osband, Alex Graves, Vlad Mnih, Remi Munos, Demis Hassabis, Olivier Pietquin, Charles Blundell, Shane Legg, 2018, arXiv:1706.10295

[6] Prioritized Experience Replay by Tom Schaul, John Quan, Ioannis Antonoglou, David Silver, 2016, arXiv:1511.05952