liuliu

60 ms·

901.
by liuliu 12y ago
Or just use std::atomic and other std::mutex in Objective-C++. Under Objective-C world, no memory semantics are well-defined, and all these are hacks on pile of other hacks.
902.
by liuliu 12y ago
That's when learning rate changed to a smaller number. The graph mainly shows that with different initialization scheme, the network starts descending initially faster.
903.
by liuliu 12y ago
Great! Feel free to reach me at i@liuliu.me!
904.
by liuliu 12y ago
It includes a training program (./bin/image-net.c), however, it requires a bit of work to get it started (you need to initialize a new convnet with the pre-trained parameters, and then fine-tuning the whole network). It is on my t
905.
Ccv 0.7, with a sub-10% image classifier and a decent face detector(libccv.org)
53 points by liuliu 12y ago|11 comments
906.
by liuliu 12y ago
The computing model, no, not really anything fundamentally different. It comes to tooling and profiling under Linux. Also, NVidia has slightly beefer cores and fewer ones, where as AMD has more cores (as I heard). Thus, for me, CUDA is a mo
907.
by liuliu 12y ago
Hmm. I am thinking about to get additional two GPUs for my configuration. 980 indeed sounds interesting, but the memory is too small. Any rumors about when NVIDIA plan to release a 6G+ card at similar price point as Titan?
908.
by liuliu 12y ago
This year's result: http://image-net.org/challenges/LSVRC/2014/results
909.
by liuliu 12y ago
CNN layer's parameters are not interesting for compression, these are very small comparing to the massive full connect layers. In retrospective, the full connect layer's regularity probably just reflects that some neurons are more
910.
by liuliu 12y ago
Yes, it is pretty noisy. My assumption is that PNG is pretty good at encoding 2D patterns too, and there are some patterns: https://github.com/liuliu/klaus/blob/master/ccv/ccvResources... This is a
911.
How to run a state-of-the-art image classifier on an iPhone(libccv.org)
65 points by liuliu 12y ago|18 comments
912.
by liuliu 12y ago
Actually, I don't quite understand their method. At the end of the day, it is not a gigantic shared memory, as in Hogwild! paper conveniently describes. You need to synchronize slave's model with master model, in model parallelism
913.
by liuliu 12y ago
John Carmack talked about latency issues in modern graphics pipeline in length: http://www.altdev.co/2013/02/22/latency-mitigation-strategie...
914.
by liuliu 13y ago
One bits interests me is the contribution of clang-omp: http://clang-omp.github.io/ looks promising, and cannot wait to try this in trunk.
915.
by liuliu 13y ago
Not really. JPEG encodes by 8x8 blocks, thus, for the DCT transform at the beginning, the first component after the transformation is always the mean value (of the 64 elements). Therefore, if your resizing target is more than 8x smaller, yo
916.
by liuliu 13y ago
Hey, Donahue, I actually referenced Caffe extensively in detailed documentation: http://libccv.org/doc/doc-convnet/ This is a preliminary implementation, but is a complete one includes both training and testing co
917.
by liuliu 13y ago
You need 1.2TB SSD for sure to train the complete ImageNet dataset. The data is loaded into GPU memory only one batch (256 images) at a time. But the loading part will be the bottleneck if you use a rotational disk.
918.
by liuliu 13y ago
Not really the whole dataset. For ImageNet, I have mini_batch size of 256, and I need to allocate the whole network on GPU for this mini_batch (which is 256 * the neurons in the network), and plus parameters (parameters is about 200MiB * 3
919.
by liuliu 13y ago
From my understanding, you cannot deploy OverFeat in production because its license is only for research and evaluation purposes. You still need TITAN for any reasonable custom datasets (with non-trivial data). The current CPU implementatio
920.
Ccv open sources near state-of-the-art image classifier under Creative Commons(libccv.org)
199 points by liuliu 13y ago|30 comments
921.
by liuliu 13y ago
Exactly, H264 supposes to be a patent pool that as long as you pay the royalty and get a license, you are free of other patent trolling. Mess these two up can significantly hinder the positive side of patent pooling. Would you use a impleme
922.
by liuliu 13y ago
last time I checked, Microsoft has a weird restriction on their c++ compiler (they don't have a c compiler) that only allows 4k length const chars, you have to concatenate consts to get arbitrary length.
923.
by liuliu 13y ago
You mean like: https://github.com/liuliu/mopack ? But unable to inspect the embedded resource at programming time is a big show-stopper for me.
924.
by liuliu 13y ago
have you setup Exchange protocol support (for pushmail on iOS devices)? Gmail disabled Exchange protocol support for newer devices, that would be my main motivation to move away from Gmail.
925.
by liuliu 13y ago
Exactly. I would think that a layered cache scheme probably could have the best of the two worlds (a good size of cached content on disk, and fast uncompressed image hit from memory). But using mmap'ed file instead of doing own layered
926.
by liuliu 13y ago
It does seem weird that they store the uncompressed image on the disk. I think that a benchmark on this comparing with compressed / decompress on CPU would be interesting because one is bounded on I/O, another is bounded by CPU th
927.
by liuliu 13y ago
Because Accelerate framework is in obj-c? (It is in C). But in all seriousness, most cases, you want to implement your own neural network anyway because the specific choice of network structure, activation functions etc. Beyond that, it jus
928.
by liuliu 13y ago
The Titan card seems requires 600W PSU: http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-tit... . Don't know how it works with 450W PSU.
929.
An Amazon review from 2001(amazon.com)
1 points by liuliu 13y ago|0 comments
930.
Deep Learning Tutorial(deeplearning.net)
2 points by liuliu 13y ago|0 comments

More ›