Aapo Kyrola 1e5140aa76 option to recompute blobs backward pass with massive memory savings ... Summary: This diff adds an option to recurrent_net to define some cell blobs to be recomputed on backward step, and thus they don't need to be stored in the step workspace. This is done by modifying the backward step to automatically include all operators that are needed to produce the output that is to be recomputed, and by storing those blobs in a shared workspace. To enable the shared workspace, i had to modify the stepworkspaces blob to also store a forward shared workspace. Making it a class field won't work since the lifecycle of the blob does not match the lifecycle of the operator. For basic LSTM, the performance hit is quite modest (about 15% with one setting, but your mileage might vary. For Attention models, I am sure this is beneficial as computing the attention blobs is not expensive. For basic LSTM, the memory saving is wonderful: each forward workspace only has 4 bytes (for timestep). I also modified the neural_mt LSTM Cells, but there is no test available, so I am not 100% sure I did it correctly. Please have a look. Added options to LSTM, MILSTM and LSTMAttention to enable memory mode. Reviewed By: urikz Differential Revision: D4853890 fbshipit-source-id: d8d0e0e75a5330d174fbfa39b96d8e4e8c446baa		2017-04-11 13:03:48 -07:00
..
docs	doxygen python block added	2017-03-29 06:46:16 -07:00
examples	Configurable CuDNN workspace limit in resnet50_trainer	2017-04-05 10:50:00 -07:00
helpers	create helpers package and add dropout	2017-04-07 17:33:49 -07:00
layers	create bucket-based calibration - layer	2017-04-11 12:30:26 -07:00
mint	doxygen python block added	2017-03-29 06:46:16 -07:00
models	Downloader fix	2017-04-07 10:16:58 -07:00
operator_test	option to recompute blobs backward pass with massive memory savings	2017-04-11 13:03:48 -07:00
predictor	Predictor exporter open-sourcing	2017-04-06 10:01:42 -07:00
_import_c_extension.py	doxygen python block added	2017-03-29 06:46:16 -07:00
attention.py	option to recompute blobs backward pass with massive memory savings	2017-04-11 13:03:48 -07:00
caffe_translator_test.py	Allow test discovery in caffe2/python/	2017-03-14 18:16:41 -07:00
caffe_translator.py	Add Reduction layer in caffe_translator	2017-04-07 16:17:07 -07:00
checkpoint_test.py	Skips the initialization phase of the individual checkpoint objects.	2017-03-31 10:10:56 -07:00
checkpoint.py	Skips the initialization phase of the individual checkpoint objects.	2017-03-31 10:10:56 -07:00
CMakeLists.txt	CMake completions work	2017-01-11 16:59:22 -08:00
cnn.py	create helpers package and add dropout	2017-04-07 17:33:49 -07:00
context_test.py	Make ContextManager thread-safe	2017-02-13 19:45:35 -08:00
context.py	doxygen python block added	2017-03-29 06:46:16 -07:00
control_test.py	fbsync. TODO: check if build files need update.	2016-11-15 00:00:46 -08:00
control.py	doxygen python block added	2017-03-29 06:46:16 -07:00
convnet_benchmarks_test.py	chunky sync - build scripts to be written	2016-07-21 10:16:42 -07:00
convnet_benchmarks.py	doxygen python block added	2017-03-29 06:46:16 -07:00
core_gradients_test.py	add inference for gradient ops + a couple of missing shape inference functions + fix to scalars	2017-02-28 23:33:32 -08:00
core_test.py	NextScopedBlob with well-defined behavior and respect namescope	2017-02-16 17:16:36 -08:00
core.py	Caffe2/Recurrent] recurrent.py API to cuDNN LSTM	2017-04-05 14:20:23 -07:00
crf.py	cuDNN version of TransposeOp	2017-04-03 13:33:10 -07:00
cudnn_recurrent_test.py	Caffe2/Recurrent] recurrent.py API to cuDNN LSTM	2017-04-05 14:20:23 -07:00
data_parallel_model_test.py	make memonger work with RecurrentNetwork(Gradient)	2017-04-05 09:48:25 -07:00
data_parallel_model.py	doxygen python block added	2017-03-29 06:46:16 -07:00
data_workers_test.py	close blobs queues when stopping + test	2017-02-27 10:07:57 -08:00
data_workers.py	doxygen python block added	2017-03-29 06:46:16 -07:00
dataio_test.py	Stop multi_reader if we run out of data before max_examples	2017-03-10 18:03:57 -08:00
dataio.py	doxygen python block added	2017-03-29 06:46:16 -07:00
dataset.py	doxygen python block added	2017-03-29 06:46:16 -07:00
db_test.py	Fix db_test under tsan	2016-11-29 15:18:37 -08:00
device_checker.py	doxygen python block added	2017-03-29 06:46:16 -07:00
dyndep.py	doxygen python block added	2017-03-29 06:46:16 -07:00
experiment_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
extension_loader.py	Make extension loader properly handle visibility.	2017-03-30 14:38:38 -07:00
gradient_check_test.py	gradient checker for nets	2017-03-28 13:03:14 -07:00
gradient_checker.py	doxygen python block added	2017-03-29 06:46:16 -07:00
hsm_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
hypothesis_test_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
hypothesis_test.py	feature processing ops	2017-04-11 07:07:51 -07:00
layer_model_helper.py	Remove unused optimizers	2017-04-05 21:18:29 -07:00
layer_model_instantiator.py	doxygen python block added	2017-03-29 06:46:16 -07:00
layer_test_util.py	uniform_sampling layer	2017-03-29 14:36:12 -07:00
layers_test.py	Add option to subtract log odd from sampled trained prediction.	2017-04-03 17:50:58 -07:00
load_save_test.py	Improve error message from LogFileDB on missing file	2017-03-10 23:31:28 -08:00
lstm_benchmark.py	option to recompute blobs backward pass with massive memory savings	2017-04-11 13:03:48 -07:00
memonger_test.py	Added a DP + recursion algorithm for finding optimal blob assignments based on blob sizes.	2017-04-07 02:18:08 -07:00
memonger.py	Added a DP + recursion algorithm for finding optimal blob assignments based on blob sizes.	2017-04-07 02:18:08 -07:00
mkl_test_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
model_device_test.py	Comment out NHWC Alexnet test for now	2017-01-23 13:59:29 -08:00
model_helper.py	ability to disable inputs for extract predictor net	2017-04-06 17:05:32 -07:00
model_helpers_test.py	create helpers package and add dropout	2017-04-07 17:33:49 -07:00
model_helpers.py	create helpers package and add dropout	2017-04-07 17:33:49 -07:00
mpi_python.cc	Move mpi_python.cc to the python folder to be more consistent about source file locations.	2017-01-09 10:59:39 -08:00
muji_test.py	chunky sync - build scripts to be written	2016-07-21 10:16:42 -07:00
muji.py	doxygen python block added	2017-03-29 06:46:16 -07:00
net_builder_test.py	Allow test discovery in caffe2/python/	2017-03-14 18:16:41 -07:00
net_builder.py	doxygen python block added	2017-03-29 06:46:16 -07:00
net_drawer.py	doxygen python block added	2017-03-29 06:46:16 -07:00
net_printer_test.py	Debug/Analysis tools for Jobs/ExecutionSteps	2017-02-06 17:31:20 -08:00
net_printer.py	doxygen python block added	2017-03-29 06:46:16 -07:00
optimizer_test_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
optimizer_test.py	Allow test discovery in caffe2/python/	2017-03-14 18:16:41 -07:00
optimizer.py	doxygen python block added	2017-03-29 06:46:16 -07:00
pipeline.py	doxygen python block added	2017-03-29 06:46:16 -07:00
predictor_constants.py	Constant string is generated from Protobuf instead of Thrift	2017-04-04 15:03:39 -07:00
pybind_state_gpu.cc	Cudnn v6	2017-02-28 17:46:33 -08:00
pybind_state_mkl.cc	Expose MKLMemory to the Python Feed and Fetch interface, and misc changes	2016-11-29 15:18:36 -08:00
pybind_state.cc	Add CAFFE_ENFORCE to protobuf parsing	2017-04-06 14:34:30 -07:00
pybind_state.h	bugfix for Windows, esp. VS 2017	2017-03-21 05:17:59 -07:00
python_op_test.py	Allow PythonOp to access the workspace	2016-12-05 11:53:26 -08:00
queue_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
record_queue.py	doxygen python block added	2017-03-29 06:46:16 -07:00
recurrent.py	option to recompute blobs backward pass with massive memory savings	2017-04-11 13:03:48 -07:00
schema_test.py	Struct nested field name lookup supports List	2017-03-24 18:17:19 -07:00
schema.py	distributed training for dper2	2017-03-30 19:04:50 -07:00
scope_test.py	fbsync. TODO: check if build files need update.	2016-11-15 00:00:46 -08:00
scope.py	doxygen python block added	2017-03-29 06:46:16 -07:00
session_test.py	NextScopedBlob with well-defined behavior and respect namescope	2017-02-16 17:16:36 -08:00
session.py	doxygen python block added	2017-03-29 06:46:16 -07:00
sparse_to_dense_mask_test.py	Fix few more operators to handle empty batches correctly.	2016-11-29 15:18:37 -08:00
task.py	doxygen python block added	2017-03-29 06:46:16 -07:00
test_util.py	doxygen python block added	2017-03-29 06:46:16 -07:00
text_file_reader.py	doxygen python block added	2017-03-29 06:46:16 -07:00
timeout_guard.py	doxygen python block added	2017-03-29 06:46:16 -07:00
toy_regression_test.py	sync	2016-08-10 11:02:15 -07:00
tt_core_test.py	sync	2016-08-10 11:02:15 -07:00
tt_core.py	doxygen python block added	2017-03-29 06:46:16 -07:00
utils.py	doxygen python block added	2017-03-29 06:46:16 -07:00
visualize.py	doxygen python block added	2017-03-29 06:46:16 -07:00
workspace_test.py	Added predictor bindings to python interface	2017-03-15 11:17:54 -07:00
workspace.py	doxygen python block added	2017-03-29 06:46:16 -07:00