nnForge v1.1.5

I ported performance improvemetns for forward propagation of convolutional layers in GPU backend to hessian calculators and weight updaters. Get latest release of nnForge v1.1.5:

• Performance of weight updaters for convolutional layers is improved in CUDA backend. Mostly for Kepler architecture. The increase is much smaller than for forward prop, I got >800 GFLOPs on training single Galaxy Zoo network on GeForce Titan
• Convolutional 1D, 2D, 3D, and 4D layers are fully supported by CUDA backend (it was 2D and 3D only before)
• Fixed training multiple networks with CPU backend
• Fixed supervised_data_mem_reader for float input data

nnForge v1.1.4

Here is new v1.1.4 release of nnForge. It contains mostly performance improvements but they are rather significant: the performance of convolutional layer in GPU (CUDA) backend improved from 1.15 TFLOPs to 2 TFLOPs for Galaxy Zoo example when running on NVIDIA GeForce GTX Titan, and the whole network performance improved from 1 TFLOPs to 1.55 TFLOPs. See the NSight VSE profiling screenshot:

2 TFLOPs on Titan running at 784 MHz... how efficient is it? Let's see: 2 TFLOPs / (784 Mhz * 14 SM * 192 fma/SM * 2 op/fma) = 47% of theoretical peak, which I consider a pretty good number. And there is certaily room for improvement here. Training could also benefit from these improvements; I plan to port these changes to hessian calculators and updaters soon.

Here are all the changes in this release:

• C++11 limited support added: you can build everything except for CUDA backend - this is due to NVCC not yet supporting C++11
• Improved testing and validating (feed forward) performance of convolutional layers in CUDA backend for Kepler at the same time greatly simplifying the code
• Improved performance of max subsampling 2d tester for CUDA backend. The implementation is far from optimal yet

Galaxy Zoo

I took the second place in Galazy Zoo competition. Organizers requested the report from all prize winners, here is mine. Sander Dieleman won the challenge with a large margin. He used convolutional neural networks too, although his approach was more sophisticated. Team 6789, which took the thrird place, used convnets too!

nnForge v1.1.3

I labelled the latest changes in nnForge with v1.1.3:

• Snapshot functionality is redesigned fully - it is now doing backpropagation, the feature is still in beta
• Ability to define custom error functions is added
• Cross-entropy error function is added, use with care - not tested yet
• Galaxy Zoo example added - see Galaxy Zoo challenge
• cuda_max_global_memory_usage_ratio is set to 0.8 by default - This should help those running code on a primary videocard
• per_layer_mu mode is added - More robust training in some cases
• Fixes:
• Fixed crash when using output transformer
• Fixed backprop for local_contrast_subtractive_2d_updater in CUDA backend
• Fixed build with Boost 1.55

nnForge v1.1.2

I brushed up parameters for nnForge toolset. I also changed default values for some of them; if you run GTSRB you will probably need to update config file. Here is the full change list:

• Deterministic transformator added for testing and validating
• snapshots are made on ANNs from batch directory
• Toolset parameters changed:
• learning_rate_decay_rate is exposed as a command line parameter
• training_speed parameter renamed to learning_rate, training_speed_degradation is dropped
• training_iteration_count renamed to training_epoch_count
• train command does batch train, batch_train command is removed
• validate and test now work in batch mode, validate_batch and test_batch removed
• mu_increase_factor is set to 1.0 by default
• max_mu set to 1.0 by default
• Bug-fixes

nnForge v1.1.1

I've just published new nnForge release v1.1.1:

• Using space-filling curve for all the convolutional updaters, testers and hessians in CUDA backend, training large networks performance improved 
• Improved concurrent training and loading/processing input data for all the stages by loading data in a separate host thread, CUDA backend only
• In-memory supervised data reader added
• Added NVTX profiling for reading input data, CUDA backend only
• Fixed:
• Binding texture to too large linear buffer
• Average subsampling backprop in CUDA backend is wrong for non-even configs
• Fixed performance in Windws with WDDM driver

Moved to blogger

Moved the blog from Zoho to Blogger platform for a number of reasons including better uptime, design, ability to edit posts in place. All posts are copied to the new platform.