Add prefetcher

Makefile

@@ -12,12 +12,7 @@ ifndef DMLC_CORE
 	DMLC_CORE = dmlc-core
 endif
 
-
-ifneq ($(USE_OPENMP_ITER), 1)
-	export NO_OPENMP = 1
-endif
-
-ifneq ($(USE_OPENMP_ITER), 1)
+ifneq ($(USE_OPENMP), 1)
 	export NO_OPENMP = 1
 endif
 

example/cifar10/cifar10.py

@@ -201,18 +201,21 @@ def Update(grad, weight, mom):
         mean_img="data/cifar/cifar_mean.bin",
         rand_crop=True,
         rand_mirror=True,
+        shuffle=False,
         input_shape=(3,28,28),
         batch_size=batch_size,
-        nthread=1)
+        nthread=4,
+        prefetch_capacity=6)
 test_dataiter = mx.io.ImageRecordIter(
         path_imgrec="data/cifar/test.rec",
         mean_img="data/cifar/cifar_mean.bin",
         rand_crop=False,
         rand_mirror=False,
+        shuffle=False,
         input_shape=(3,28,28),
         batch_size=batch_size,
-        nthread=1)
-
+        nthread=4,
+        prefetch_capacity=6)
 
 def progress(count, total, epoch, toc):
     bar_len = 50

example/cifar10/cifar10_multi_gpus.py

@@ -148,24 +148,25 @@ def momentum_update(key, grad, weight):
 get_data.GetCifar10()
 
 train_dataiter = mx.io.ImageRecordIter(
-    path_imgrec="data/cifar/train.rec",
-    mean_img="data/cifar/cifar_mean.bin",
-    rand_crop=True,
-    rand_mirror=True,
-    shuffle=True,
-    input_shape=(3,28,28),
-    batch_size=batch_size,
-    nthread=1)
-
-val_dataiter = mx.io.ImageRecordIter(
-    path_imgrec="data/cifar/test.rec",
-    mean_img="data/cifar/cifar_mean.bin",
-    rand_crop=False,
-    rand_mirror=False,
-    input_shape=(3,28,28),
-    batch_size=batch_size,
-    nthread=1)
-
+        path_imgrec="data/cifar/train.rec",
+        mean_img="data/cifar/cifar_mean.bin",
+        rand_crop=True,
+        rand_mirror=True,
+        shuffle=False,
+        input_shape=(3,28,28),
+        batch_size=batch_size,
+        nthread=4,
+        prefetch_capacity=6)
+test_dataiter = mx.io.ImageRecordIter(
+        path_imgrec="data/cifar/test.rec",
+        mean_img="data/cifar/cifar_mean.bin",
+        rand_crop=False,
+        rand_mirror=False,
+        shuffle=False,
+        input_shape=(3,28,28),
+        batch_size=batch_size,
+        nthread=4,
+        prefetch_capacity=6)
 
 def progress(count, total, epoch, tic):
     bar_len = 50

example/mnist/mlp_gpu.py

@@ -67,14 +67,15 @@ def Update(grad, weight):
 
 #check data
 get_data.GetMNIST_ubyte()
-
 train_dataiter = mx.io.MNISTIter(
         image="data/train-images-idx3-ubyte",
         label="data/train-labels-idx1-ubyte",
+        input_shape=(784,),
         batch_size=batch_size, shuffle=True, flat=True, silent=False, seed=10)
 val_dataiter = mx.io.MNISTIter(
         image="data/t10k-images-idx3-ubyte",
         label="data/t10k-labels-idx1-ubyte",
+        input_shape=(784,),
         batch_size=batch_size, shuffle=True, flat=True, silent=False)
 
 tmp_label = mx.nd.zeros(name2shape["sm_label"])

example/mnist/mlp_multi_gpu.py

@@ -60,10 +60,12 @@ def updater(key, grad, weight):
 train_dataiter = mx.io.MNISTIter(
         image="data/train-images-idx3-ubyte",
         label="data/train-labels-idx1-ubyte",
+        input_shape=(784,),
         batch_size=batch_size, shuffle=True, flat=True, silent=False, seed=10)
 val_dataiter = mx.io.MNISTIter(
         image="data/t10k-images-idx3-ubyte",
         label="data/t10k-labels-idx1-ubyte",
+        input_shape=(784,),
         batch_size=batch_size, shuffle=True, flat=True, silent=False)
 
 def cal_acc(out, label):

include/mxnet/io.h

@@ -11,7 +11,9 @@
 #include <vector>
 #include <string>
 #include <utility>
+#include <queue>
 #include "./base.h"
+#include "./ndarray.h"
 
 namespace mxnet {
 /*!
@@ -59,27 +61,16 @@ struct DataInst {
  *      data and label, how we use them is to see the DNN implementation.
  */
 struct DataBatch {
- public:
-  /*! \brief unique id for instance, can be NULL, sometimes is useful */
-  unsigned *inst_index;
-  /*! \brief number of instance */
-  mshadow::index_t batch_size;
-  /*! \brief number of padding elements in this batch,
-       this is used to indicate the last elements in the batch are only padded up to match the batch, and should be discarded */
-  mshadow::index_t num_batch_padd;
  public:
   /*! \brief content of dense data, if this DataBatch is dense */
-  std::vector<TBlob> data;
+  std::vector<NDArray> data;
   /*! \brief extra data to be fed to the network */
   std::string extra_data;
  public:
   /*! \brief constructor */
-  DataBatch(void) {
-    inst_index = NULL;
-    batch_size = 0; num_batch_padd = 0;
-  }
-  /*! \brief giving name to the data */
-  void Naming(std::vector<std::string> names);
+  DataBatch(void) {}
+  /*! \brief destructor */
+  ~DataBatch() {}
 };  // struct DataBatch
 
 /*! \brief typedef the factory function of data iterator */
@@ -121,10 +112,29 @@ struct DataIteratorReg
  * \endcode
  */
 #define MXNET_REGISTER_IO_CHAINED_ITER(name, ChainedDataIterType, HoldingDataIterType)          \
-  static ::mxnet::IIterator<DataBatch>* __create__ ## ChainedDataIteratorType ## __() { \
+  static ::mxnet::IIterator<DataBatch>* __create__ ## ChainedDataIterType ## __() { \
     return new HoldingDataIterType(new ChainedDataIterType);                                    \
   }                                                                     \
   DMLC_REGISTRY_REGISTER(::mxnet::DataIteratorReg, DataIteratorReg, name) \
-  .set_body(__create__ ## ChainedDataIteratorType ## __)
+  .set_body(__create__ ## ChainedDataIterType ## __)
+/*!
+ * \brief Macro to register three chained Iterators
+ *
+ * \code
+ * // example of registering a imagerec iterator
+ * MXNET_REGISTER_IO_CHAINED_ITERATOR(ImageRecordIter, 
+ * ImageRecordIter, ImageRecBatchLoader, Prefetcher)
+ * .describe("batched image record data iterator");
+ *
+ * \endcode
+ */
+#define MXNET_REGISTER_IO_THREE_CHAINED_ITER(\
+        name, FirstIterType, SecondIterType, ThirdIterType)          \
+  static ::mxnet::IIterator<DataBatch>* __create__ ## ThirdIterType ## __() { \
+    return new FirstIterType(new SecondIterType(new ThirdIterType));             \
+  }                                                                     \
+  DMLC_REGISTRY_REGISTER(::mxnet::DataIteratorReg, DataIteratorReg, name) \
+  .set_body(__create__ ## ThirdIterType ## __)
+
 }  // namespace mxnet
 #endif  // MXNET_IO_H_

python/mxnet/io.py

@@ -73,15 +73,15 @@ def getdata(self):
         """
         hdl = NDArrayHandle()
         check_call(_LIB.MXDataIterGetData(self.handle, ctypes.byref(hdl)))
-        return NDArray(hdl)
+        return NDArray(hdl, False)
 
     def getlabel(self):
         """get label from batch
 
         """
         hdl = NDArrayHandle()
         check_call(_LIB.MXDataIterGetLabel(self.handle, ctypes.byref(hdl)))
-        return NDArray(hdl)
+        return NDArray(hdl, False)
 
 def _make_io_iterator(handle):
     """Create an io iterator by handle."""

python/mxnet/ndarray.py

@@ -51,7 +51,7 @@ class NDArray(object):
     NDArray is basic ndarray/Tensor like data structure in mxnet.
     """
     # pylint: disable= no-member
-    def __init__(self, handle):
+    def __init__(self, handle, writable=True):
         """initialize a new NDArray
 
         Parameters
@@ -61,6 +61,7 @@ def __init__(self, handle):
         """
         assert isinstance(handle, NDArrayHandle)
         self.handle = handle
+        self.writable = writable
 
     def __del__(self):
         check_call(_LIB.MXNDArrayFree(self.handle))
@@ -555,6 +556,8 @@ def binary_ndarray_function(lhs, rhs, out=None):
         if out:
             if isinstance(out, NDArray) == False:
                 raise TypeError('out must be NDArray')
+            if not out.writable:
+                raise TypeError('out must be writable')
         else:
             if not accept_empty_mutate:
                 raise TypeError('argument out is required to call %s' % func_name)
@@ -570,6 +573,8 @@ def unary_ndarray_function(src, out=None):
         if out:
             if isinstance(out, NDArray) == False:
                 raise TypeError('out must be NDArray')
+            if not out.writable:
+                raise TypeError('out must be writable')
         else:
             if not accept_empty_mutate:
                 raise TypeError('argument out is required to call %s' % func_name)

src/c_api.cc

@@ -872,15 +872,19 @@ int MXDataIterNext(DataIterHandle handle, int *out) {
 
 int MXDataIterGetLabel(DataIterHandle handle, NDArrayHandle *out) {
   API_BEGIN();
-  DataBatch db = static_cast<IIterator<DataBatch>* >(handle)->Value();
-  *out = new NDArray(db.data[1], 0);
+  const DataBatch& db = static_cast<IIterator<DataBatch>* >(handle)->Value();
+  NDArray* pndarray = new NDArray();
+  *pndarray = db.data[1];
+  *out = pndarray;
   API_END();
 }
 
 int MXDataIterGetData(DataIterHandle handle, NDArrayHandle *out) {
   API_BEGIN();
-  DataBatch db = static_cast<IIterator<DataBatch>* >(handle)->Value();
-  *out = new NDArray(db.data[0], 0);
+  const DataBatch& db = static_cast<IIterator<DataBatch>* >(handle)->Value();
+  NDArray* pndarray = new NDArray();
+  *pndarray = db.data[0];
+  *out = pndarray;
   API_END();
 }
 

src/io/image_augmenter.h

@@ -2,12 +2,13 @@
  *  Copyright (c) 2015 by Contributors
  * \file image_augmenter_opencv.hpp
  * \brief threaded version of page iterator
- * \author Naiyan Wang, Tianqi Chen, Tianjun Xiao
  */
 #ifndef MXNET_IO_IMAGE_AUGMENTER_H_
 #define MXNET_IO_IMAGE_AUGMENTER_H_
 
+#if MXNET_USE_OPENCV
 #include <opencv2/opencv.hpp>
+#endif
 #include <utility>
 #include <string>
 #include <algorithm>
@@ -122,6 +123,8 @@ struct ImageAugmentParam : public dmlc::Parameter<ImageAugmentParam> {
         .describe("Maximum ratio of contrast variation");
     DMLC_DECLARE_FIELD(max_random_illumination).set_default(0.0f)
         .describe("Maximum value of illumination variation");
+    DMLC_DECLARE_FIELD(silent).set_default(true)
+        .describe("Whether to print augmentor info");
   }
 };
 
@@ -130,8 +133,10 @@ class ImageAugmenter {
  public:
   // contructor
   ImageAugmenter(void)
-      : tmpres_(false),
-        rotateM_(2, 3, CV_32F) {
+      : tmpres_(false) {
+#if MXNET_USE_OPENCV
+    rotateM_ = cv::Mat(2, 3, CV_32F);
+#endif
   }
   virtual ~ImageAugmenter() {
   }
@@ -164,6 +169,7 @@ class ImageAugmenter {
       }
     }
   }
+#if MXNET_USE_OPENCV
   /*!
    * \brief augment src image, store result into dst
    *   this function is not thread safe, and will only be called by one thread
@@ -174,6 +180,7 @@ class ImageAugmenter {
    */
   virtual cv::Mat OpencvProcess(const cv::Mat &src,
                           common::RANDOM_ENGINE *prnd) {
+    if (!NeedOpencvProcess()) return src;
     // shear
     float s = NextDouble(prnd) * param_.max_shear_ratio * 2 - param_.max_shear_ratio;
     // rotate
@@ -276,8 +283,9 @@ class ImageAugmenter {
     }
     return tmpres_;
   }
-
-  void TensorProcess(mshadow::TensorContainer<cpu, 3> *p_data,
+#endif
+  void TensorProcess(mshadow::Tensor<cpu, 3> *p_data,
+                     mshadow::TensorContainer<cpu, 3> *dst_data,
                        common::RANDOM_ENGINE *prnd) {
     // Check Newly Created mean image
     if (meanfile_ready_ == false && param_.mean_img.length() != 0) {
@@ -291,7 +299,8 @@ class ImageAugmenter {
         meanfile_ready_ = true;
       }
     }
-    img_.Resize(mshadow::Shape3((*p_data).shape_[0], param_.input_shape[1], param_.input_shape[2]));
+    img_.Resize(mshadow::Shape3((*p_data).shape_[0],
+                param_.input_shape[1], param_.input_shape[2]));
     if (param_.input_shape[1] == 1) {
       img_ = (*p_data) * param_.scale;
     } else {
@@ -355,27 +364,7 @@ class ImageAugmenter {
         }
       }
     }
-    (*p_data) = img_;
-  }
-
-  virtual void Process(unsigned char *dptr, size_t sz,
-                       mshadow::TensorContainer<cpu, 3> *p_data,
-                       common::RANDOM_ENGINE *prnd) {
-    cv::Mat buf(1, sz, CV_8U, dptr);
-    cv::Mat res = cv::imdecode(buf, 1);
-    if (NeedOpencvProcess())
-        res = this->OpencvProcess(res, prnd);
-    p_data->Resize(mshadow::Shape3(3, res.rows, res.cols));
-    for (index_t i = 0; i < p_data->size(1); ++i) {
-      for (index_t j = 0; j < p_data->size(2); ++j) {
-        cv::Vec3b bgr = res.at<cv::Vec3b>(i, j);
-        (*p_data)[0][i][j] = bgr[2];
-        (*p_data)[1][i][j] = bgr[1];
-        (*p_data)[2][i][j] = bgr[0];
-      }
-    }
-    res.release();
-    this->TensorProcess(p_data, prnd);
+    (*dst_data) = img_;
   }
 
  private:
@@ -392,11 +381,13 @@ class ImageAugmenter {
   mshadow::TensorContainer<cpu, 3> meanimg_;
   /*! \brief temp space */
   mshadow::TensorContainer<cpu, 3> img_;
+#if MXNET_USE_OPENCV
   // temporal space
   cv::Mat temp_;
   // rotation param
   cv::Mat rotateM_;
   // whether the mean file is ready
+#endif
   bool meanfile_ready_;
   // parameters
   ImageAugmentParam param_;