OpenCV, LMDB, LevelDB and Snappy are made optional via switches
(USE_OPENCV, USE_LMDB, USE_LEVELDB) available for Make and CMake
builds. Since Snappy is a LevelDB dependency, its use is determined by
USE_LEVELDB. HDF5 is left bundled because it is used for serializing
weights and solverstates.
Add signal handler and early exit/snapshot to Solver.
Add signal handler and early exit/snapshot to Solver.
Also check for exit and snapshot when testing.
Skip running test after early exit.
Fix more lint.
Rebase on master.
Finish rebase on master.
Fixups per review comments.
Redress review comments.
Lint.
Correct error message wording.
- Parallelize batches among GPUs and tree-reduce the gradients
- The effective batch size scales with the number of devices
- Batch size is multiplied by the number of devices
- Split batches between GPUs, and tree-reduce the gradients
- Detect machine topology (twin-GPU boards, P2P connectivity)
- Track device in syncedmem (thanks @thatguymike)
- Insert a callback in the solver for minimal code change
- Accept list for gpu flag of caffe tool, e.g. '-gpu 0,1' or '-gpu all'.
Run on default GPU if no ID given.
- Add multi-GPU solver test
- Deterministic architecture for reproducible runs
Over version introduced in https://github.com/BVLC/caffe/pull/1384
Highlights:
* Interface change: column order is now determined by using a list of `OrderedDict` objects instead of `dict` objects, which obviates the need to pass around a tuple with the column orders.
* The outputs are now named according to their names in the network protobuffer; e.g., if your top is named `loss`, then the corresponding column header will also be `loss`; we no longer rename it to, e.g., `TrainingLoss` or `TestLoss`.
* Fixed the bug/feature of the first version where the initial learning rate was always NaN.
* Add optional parameter to specify output table delimiter. It's still a comma by default.
You can use Matlab code from [this gist](https://gist.github.com/drdan14/d8b45999c4a1cbf7ad85) to verify that your results are the same before and after the changes introduced in this pull request. That code assumes that your `top` names are `accuracy` and `loss`, but you can modify the code if that's not true.
Was previously using `Train net output #M: loss = X` lines, but there may not be exactly one of those (e.g., for GoogLeNet, which has multiple loss layers); I believe that `Iteration N, loss = X` is the aggregated loss.
If there's only one loss layer, these two values will be equal and it won't matter. Otherwise, we presumably want to report the aggregated loss.
Fisher Yu
Yangqing Jia
T.E.A de Souza
Evan Shelhamer
Kai Li
Ronghang Hu
Dmytro Mishkin
Jonathan L Long
Lumin Zhou
Luke Yeager
J Yegerlehner
Cyprien Noel
Jeff Donahue
Manuel
Mohammad Norouzi
Takuma Wakamori
Daniel Golden
philkr
Anatoly Baksheev