Merge remote-tracking branch 'upstream/master'

.travis.yml

@@ -27,8 +27,6 @@ matrix:
   exclude:
     - os: osx
       env: TASK=lint
-    - os: osx
-      env: TASK=doc
     - os: linux
       env: TASK=r_test
     - os: osx
@@ -73,7 +71,7 @@ cache:
 
 before_cache:
   - dmlc-core/scripts/travis/travis_before_cache.sh
-  
+
 after_failure:
   - tests/travis/travis_after_failure.sh
 

LICENSE

@@ -1,4 +1,4 @@
-Copyright (c) 2015 by Contributors 
+Copyright (c) 2015-2016 by Contributors 
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.

Makefile

@@ -94,10 +94,11 @@ endif
 all: lib/libmxnet.a lib/libmxnet.so $(BIN)
 
 SRC = $(wildcard src/*.cc src/*/*.cc)
-OBJ = $(patsubst src/%.cc, build/%.o, $(SRC))
+OBJ = $(patsubst %.cc, build/%.o, $(SRC))
 CUSRC = $(wildcard src/*/*.cu)
-CUOBJ = $(patsubst src/%.cu, build/%_gpu.o, $(CUSRC))
+CUOBJ = $(patsubst %.cu, build/%_gpu.o, $(CUSRC))
 
+# extra operators
 ifneq ($(EXTRA_OPERATORS),)
 	EXTRA_SRC = $(wildcard $(EXTRA_OPERATORS)/*.cc $(EXTRA_OPERATORS)/*/*.cc)
 	EXTRA_OBJ = $(patsubst $(EXTRA_OPERATORS)/%.cc, $(EXTRA_OPERATORS)/build/%.o, $(EXTRA_SRC))
@@ -110,10 +111,16 @@ else
 	EXTRA_CUOBJ =
 endif
 
+# plugin
+PLUGIN_OBJ =
+PLUGIN_CUOBJ =
+include $(MXNET_PLUGINS)
+
+# all dep
 LIB_DEP += $(DMLC_CORE)/libdmlc.a
-ALL_DEP = $(OBJ) $(EXTRA_OBJ) $(LIB_DEP)
+ALL_DEP = $(OBJ) $(EXTRA_OBJ) $(PLUGIN_OBJ) $(LIB_DEP)
 ifeq ($(USE_CUDA), 1)
-	ALL_DEP += $(CUOBJ) $(EXTRA_CUOBJ)
+	ALL_DEP += $(CUOBJ) $(EXTRA_CUOBJ) $(PLUGIN_CUOBJ)
 	LDFLAGS += -lcuda
 endif
 
@@ -125,14 +132,26 @@ else
 endif
 
 
-build/%.o: src/%.cc
+build/src/%.o: src/%.cc
+	@mkdir -p $(@D)
+	$(CXX) -std=c++0x $(CFLAGS) -MM -MT build/src/$*.o $< >build/src/$*.d
+	$(CXX) -std=c++0x -c $(CFLAGS) -c $< -o $@
+
+build/src/%_gpu.o: src/%.cu
+	@mkdir -p $(@D)
+	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" -M -MT build/src/$*_gpu.o $< >build/src/$*_gpu.d
+	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $<
+
+build/plugin/%.o: plugin/%.cc
 	@mkdir -p $(@D)
-	$(CXX) -std=c++0x $(CFLAGS) -MM -MT build/$*.o $< >build/$*.d
+	$(CXX) -std=c++0x $(CFLAGS) -MM -MT build/plugin/$*.o $< >build/plugin/$*.d
 	$(CXX) -std=c++0x -c $(CFLAGS) -c $< -o $@
 
-build/%_gpu.o: src/%.cu
+# A nvcc bug cause it to generate "generic/xxx.h" dependencies from torch headers.
+# Use CXX to generate dependency instead.
+build/plugin/%_gpu.o: plugin/%.cu
 	@mkdir -p $(@D)
-	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" -M -MT build/$*_gpu.o $< >build/$*_gpu.d
+	$(CXX) -std=c++0x $(CFLAGS) -MM -MT build/plugin/$*_gpu.o $< >build/plugin/$*_gpu.d
 	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $<
 
 $(EXTRA_OPERATORS)/build/%.o: $(EXTRA_OPERATORS)/%.cc
@@ -173,7 +192,7 @@ include tests/cpp/unittest.mk
 test: $(TEST)
 
 lint: rcpplint
-	python2 dmlc-core/scripts/lint.py mxnet ${LINT_LANG} include src scripts python predict/python
+	python2 dmlc-core/scripts/lint.py mxnet ${LINT_LANG} include src plugin scripts python predict/python
 
 doc: doxygen
 
@@ -216,6 +235,7 @@ clean_all: clean
 
 -include build/*.d
 -include build/*/*.d
+-include build/*/*/*.d
 ifneq ($(EXTRA_OPERATORS),)
 	-include $(EXTRA_OPERATORS)/build/*.d
 endif

R-package/DESCRIPTION

@@ -24,7 +24,8 @@ Suggests:
     mlbench,
     knitr,
     rmarkdown,
-    imager
+    imager,
+    roxygen2
 LinkingTo: Rcpp
 RoxygenNote: 5.0.1
 VignetteBuilder: knitr

R-package/R/executor.R

@@ -2,7 +2,7 @@
 #' with information from input shapes.
 #'
 #' @export
-mx.simple.bind <- function(symbol, ctx, grad.req=FALSE, ...) {
+mx.simple.bind <- function(symbol, ctx, grad.req = "null", ...) {
   if (!is.MXSymbol(symbol)) stop("symbol need to be MXSymbol")
   slist <- symbol$infer.shape(list(...))
 
@@ -16,9 +16,11 @@ mx.simple.bind <- function(symbol, ctx, grad.req=FALSE, ...) {
     mx.nd.zeros(shape, ctx)
   }, simplify = FALSE, USE.NAMES = TRUE)
   grad.reqs <- lapply(names(slist$arg.shapes), function(nm) {
-    grad.req &&
-    !mx.util.str.endswith(nm, "label") &&
-    !mx.util.str.endswith(nm, "data")
+    if (!mx.util.str.endswith(nm, "label") && !mx.util.str.endswith(nm, "data")) {
+      grad.req
+    } else {
+      "null"
+    }
   })
   mx.symbol.bind(symbol, ctx,
                  arg.arrays=arg.arrays,

R-package/R/model.R

@@ -104,7 +104,7 @@ mx.model.train <- function(symbol, ctx, input.shape,
   # create the executors
   sliceinfo <- mx.model.slice.shape(input.shape, ndevice)
   train.execs <- lapply(1:ndevice, function(i) {
-    mx.simple.bind(symbol, ctx=ctx[[i]], data=sliceinfo[[i]]$shape, grad.req=TRUE)
+    mx.simple.bind(symbol, ctx=ctx[[i]], data=sliceinfo[[i]]$shape, grad.req="write")
   })
   # set the parameters into executors
   for (texec in train.execs) {
@@ -469,7 +469,7 @@ predict.MXFeedForwardModel <- function(model, X, ctx=NULL, array.batch.size=128,
   X$reset()
   if (!X$iter.next()) stop("Cannot predict on empty iterator")
   dlist = X$value()
-  pexec <- mx.simple.bind(model$symbol, ctx=ctx, data=dim(dlist$data), grad.req=FALSE)
+  pexec <- mx.simple.bind(model$symbol, ctx=ctx, data=dim(dlist$data), grad.req="null")
   mx.exec.update.arg.arrays(pexec, model$arg.params, match.name=TRUE)
   mx.exec.update.aux.arrays(pexec, model$aux.params, match.name=TRUE)
   packer <- mx.nd.arraypacker()

R-package/R/mxnet_generated.R

@@ -413,6 +413,8 @@ mx.symbol.Cast <- function(...) {
 
 #' Perform an feature concat on channel dim (dim 1) over all the inputs.
 #' 
+#' @param data  Symbol[]
+#'     List of tensors to concatenate
 #' @param num.args  int, required
 #'     Number of inputs to be concated.
 #' @param dim  int, optional, default='1'
@@ -792,6 +794,8 @@ mx.symbol.SoftmaxActivation <- function(...) {
 #' 
 #' @param data  Symbol
 #'     Input data to softmax.
+#' @param label  Symbol
+#'     Label data.
 #' @param grad.scale  float, optional, default=1
 #'     Scale the gradient by a float factor
 #' @param ignore.label  float, optional, default=-1
@@ -828,6 +832,8 @@ mx.symbol.SwapAxis <- function(...) {
 
 #' Perform nearest neighboor/bilinear up sampling to inputs
 #' 
+#' @param data  Symbol[]
+#'     Array of tensors to upsample
 #' @param scale  int (non-negative), required
 #'     Up sampling scale
 #' @param num.filter  int (non-negative), optional, default=0

R-package/demo/basic_executor.R

@@ -19,7 +19,7 @@ exec = mxnet:::mx.symbol.bind(
   ctx=mx.cpu(),
   arg.arrays = list(A=a, B=b),
   aux.arrays = list(),
-  grad.reqs = list(FALSE, FALSE))
+  grad.reqs = list("null", "null"))
 
 # calculate outputs
 mx.exec.forward(exec)

R-package/src/executor.cc

@@ -141,14 +141,22 @@ inline Rcpp::List* CreateGradList(const Rcpp::List& source_array,
     ret->names() = names;
     handles->resize(grad_reqs.size(), nullptr);
     grad_req_type->resize(grad_reqs.size(), 0);
+    std::map<std::string, int> req_map;
+    req_map["null"] = 0;
+    req_map["write"] = 1;
+    req_map["add"] = 3;
 
     for (size_t i = 0; i < grad_reqs.size(); ++i) {
-      RCHECK(Rcpp::is<bool>(grad_reqs[i]))
-          << "Expect input grad_reqs to be list of booleans";
-      if (Rcpp::as<bool>(grad_reqs[i])) {
+      if (Rcpp::as<std::string>(grad_reqs[i]) != "null"
+          && Rcpp::as<std::string>(grad_reqs[i]) != "write"
+          && Rcpp::as<std::string>(grad_reqs[i]) != "add") {
+        RLOG_FATAL << "grad_req must be one of 'null', 'write' or 'add'";
+      }
+
+      if (Rcpp::as<std::string>(grad_reqs[i]) != "null") {
         ret->at(i) = NDArray::Empty(NDArray::FromRObject(source_array[i]).dim(), ctx);
         handles->at(i) = NDArray::FromRObject(ret->at(i))->handle;
-        grad_req_type->at(i) = 1;
+        grad_req_type->at(i) = req_map[Rcpp::as<std::string>(grad_reqs[i])];
       }
     }
   } catch(const Rcpp::exception& ex) {

README.md

@@ -20,6 +20,7 @@ deep learning system, and interesting insights of DL systems for hackers.
 
 What's New
 ----------
+* [Embedding Torch layers and functions in MXNet](https://mxnet.readthedocs.org/en/latest/tutorial/torch_howto.html)
 * [MXNet.js: Javascript Package for Deep Learning in Browser (without server)
 ](https://github.com/dmlc/mxnet.js/)
 * [Design Note: Design Efficient Deep Learning Data Loading Module](http://mxnet.readthedocs.org/en/latest/developer-guide/note_data_loading.html)
@@ -34,6 +35,7 @@ Contents
 * [Documentation and Tutorials](http://mxnet.readthedocs.org/en/latest/)
 * [Open Source Design Notes](http://mxnet.readthedocs.org/en/latest/#open-source-design-notes)
 * [Code Examples](example)
+* [Installation](http://mxnet.readthedocs.org/en/latest/build.html)
 * [Pretrained Models](https://github.com/dmlc/mxnet-model-gallery)
 * [Contribute to MXNet](http://mxnet.readthedocs.org/en/latest/contribute.html)
 * [Frequent Asked Questions](http://mxnet.readthedocs.org/en/latest/faq.html)

doc/build.md

@@ -24,7 +24,7 @@ even better to send pull request if you can fix the problem.
 - [Python Package Installation](#python-package-installation)
 - [R Package Installation](#r-package-installation)
 - [Docker Images](#docker-images)
-- [Frequently asked questions](#frequently-asked-questions)
+- [Trouble Shooting](#trouble-shooting)
 
 ## Build the Shared Library
 
@@ -240,10 +240,10 @@ R CMD INSTALL mxnet_0.5.tar.gz
 ```
 
 
-To install the package using GPU on Windows without building the package from scratch. Note that you need a couple of programs installed already:  
+To install the package using GPU on Windows without building the package from scratch. Note that you need a couple of programs installed already:
 - You'll need the [CUDA Toolkit](https://developer.nvidia.com/cuda-toolkit). This depends on Visual Studio, and a free compatible version would be [Visual Studio Community 2013](https://www.visualstudio.com/en-us/news/vs2013-community-vs.aspx). For instructions and compatibility checks, read http://docs.nvidia.com/cuda/cuda-getting-started-guide-for-microsoft-windows/ .
 
-- You will also need to register as a developer at nvidia and download CUDNN V3, https://developer.nvidia.com/cudnn . 
+- You will also need to register as a developer at nvidia and download CUDNN V3, https://developer.nvidia.com/cudnn .
 
 
 1. Download the mxnet package as a ZIP from the Github repository https://github.com/dmlc/mxnet and unpack it. You will be editing the `/mxnet/R-package` folder.
@@ -256,7 +256,7 @@ To install the package using GPU on Windows without building the package from sc
 
 4. Create the folder `R-package/inst/libs/x64`. We only support 64-bit operating system now, so you need the x64 folder;
 
-5. Put dll files in `R-package/inst/libs/x64`. 
+5. Put dll files in `R-package/inst/libs/x64`.
 
 The first dll file you need is `nocudnn/lib/libmxnet.dll`. The other dll files you need are the ones in all 4 subfolders of `nocudnn/3rdparty/`, for the `cudnn` and `openblas` you'll need to look in the `/bin` folders. There should be 11 dll files now in `R-package/inst/libs/x64`.
 
@@ -284,17 +284,17 @@ sudo docker run -it --device /dev/nvidiactl --device /dev/nvidia-uvm --device /d
 For a guide to Docker, see the [official docs](https://docs.docker.com/userguide/). For more details on how to use the
 MXNet Docker images, including requirements for CUDA support, consult the [source project](https://github.com/Kaixhin/dockerfiles).
 
-## Frequently Asked Questions
+## Trouble Shooting
 
-1. **Compile failed after `git pull`**
+### Compile failed after git pull
 
    Please first update the submodules, clean all and recompile:
 
    ```bash
    git submodule update && make clean_all && make -j4
    ```
 
-2. **Compile failed after `config.mk` is modified**
+### Compile failed after config.mk is modified
 
    This often happens if `USE_CUDA` or `USE_DIST_KVSTORE` has been changed. You
    need to clean all first:
@@ -303,7 +303,9 @@ MXNet Docker images, including requirements for CUDA support, consult the [sourc
     make clean_all && make -j4
     ```
 
-3. **Still get the error message e.g. `compile with USE_DIST_KVSTORE=1 to use
+### Still get the error message after re-installation
+
+   e.g. `compile with USE_DIST_KVSTORE=1 to use
    dist` after recomplied with `USE_DIST_KVSTORE=1`**
 
    It is often because mxnet is failed to load the new built library. If you