I want to vent a bit about the current landscape of Deep Learning frameworks / libraries / middlewares. Why do all of them have different file formats? Its like having web browsers that expect different types of responses from web servers. Hua? Neural Networks are pretty simple. Its a graph with weights. Basically an ordered set of linear algerbra operations. Now, with dropout and resampling we see amazing non-linear results but the core of what the network is / does isn’t that complicated.
What is a NVIDIA Jetson TX1 Development Kit? In short is a quad-core ARM and a 256-core Maxwell GPU packed into a single chip with pretty much all the accessories you need in a breakout board. Here is an unboxing video. I have had the NVIDIA Jetson TK1 Development Kit for a while now and the TX1 it is quite a bit more powerful, though 3x the cost. Getting Torch Setup First step is to get the TX1 flashed with the latest Linux 4 Tegra (L4T) along with the full CUDA toolchain.
OSX Pickup dependencies: brew tap caskroom/cask brew cask install anaconda brew install zeromq openssl luarocks install luacrypto OPENSSL_DIR=/usr/local/opt/openssl/ Add to your shell if you use zsh: echo "export PATH=\"$HOME/anaconda3/bin:$PATH\"" >> ~/.zshrc source ~/.zshrc Or if you use bash: echo "export PATH=\"$HOME/anaconda3/bin:$PATH\"" >> ~/.bashrc source ~/.bashrc Now we can install Jupyter and iTorch conda install jupyter git clone https://github.com/facebook/iTorch.git cd iTorch && luarocks make && popd Hello World Lets grab this great example:
Disable Integrity Protection OSX El Capitan needs the disable integrity protection to run the modified Nvidia Drivers Shutdown your Mac. Press the power button and immediately hold down Command + R until you see the apple logo Now your Mac will be in Recovery Mode. Open up Utilities -> Terminal Run the following command: csrutil disable You should see text confirming that integrity protection is now disabled. Now restart your Mac.
Equipment MacBookPro or MacBookAir with Thunderbolt 2 port(s) Akitio Thunder2 PCIe Box GeForce GPU capable of CUDA, Maxwell generation or newer http://www.geforce.com/hardware/technology/cuda/supported-gpus Bplus SWEX : ATX Power Supply Power ON switch ATX PC Power Supply (>400W) StarTech BEZELWRKIT ATX Case Front Bezel Wire Kit 4 Pcs Green Black Latching On/Off SPST Round Rocker Switch The Box Basically throw away the Akitio chassis but keep the power cord from the transformer to the chassis (the DC power part).
Starting with John Kaufhold‘s DeepLearning presentation (pdf) to the DataCommunity DC back in 2014. Marvin Galvin did a blog post covering the highlights. Here is the video: With that history lesson, I want to zoom down a level of detail and understand all these papers. Here are the papers in a handy script to download them. View the code on Gist.