Skip to content

update recurrent branch with latest BVLC/master#9

Open
dribnet wants to merge 27 commits intojeffdonahue:recurrentfrom
dribnet:recurrent-net-fix
Open

update recurrent branch with latest BVLC/master#9
dribnet wants to merge 27 commits intojeffdonahue:recurrentfrom
dribnet:recurrent-net-fix

Conversation

@dribnet
Copy link
Copy Markdown

@dribnet dribnet commented Apr 27, 2015

Merged in latest BVLC/master to keep BVLC#2033 up to date and mergeable. Resolved seemingly straightforward conflict introduced by BVLC#2370 vs a7eaaf5. Confirmed make runtest completes successfully (659 tests).

longjon and others added 27 commits March 13, 2015 13:12
(With layers whose backwards accumlate gradients), this effectively
decouples the computational batch from the SGD minibatch. Each
iteration accumulates gradients over iter_size batches, then parameters
are updated.
(double impl from NVIDIA dev docs; float impl included in CUDA as
"atomicAdd")
Removed the CPU_ONLY fix introduced in
BVLC#2370 because the surrounding
Net<Dtype>::Update() logic was previously
removed in this feature branch at a7eaaf5.

Merge remote-tracking branch 'jeff/recurrent'

* jeff/recurrent: (26 commits)
  RecurrentLayer bugfix: params still need backprop
  Prototxts + script for training LRCN COCO image captioning model
  Prototxts + script for training COCO caption language model
  Add scripts to create HDF5 datasets from COCO captions
  Add scripts for downloading COCO2014 tools & data
  Add LSTMLayer and LSTMUnitLayer, with tests
  Add RNNLayer, with tests
  Add RecurrentLayer: an abstract superclass for other recurrent layer types
  TestNet fixes for Net weight sharing modifications
  Modifications to Net to facilitate unrolled recurrent networks
  Allow ConcatLayer to take a single bottom Blob (for testing)
  Allow SliceLayer to have a single top Blob (for testing)
  EltwiseLayer with coeff blob GPU kernel
  EltwiseLayer can take a blob of per-num coefficients
  AccuracyLayer: add 'denominator' param
  FlattenLayer fix -- top should always Share* from bottom (and do everything in Reshape)
  Add (very simple version of) ReshapeLayer
  EmbedBackward with no loops -- use caffe_gpu_atomic_add instead
  Add EmbedLayer for inner products with sparse input (one-hot vectors), with unit tests
  test_gradient_check_util: check_bottom < -1 only checks params
  ...

Conflicts:
  src/caffe/net.cpp
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants