Do not merge: demonstrate mshadow size cast build.

.gitignore

@@ -66,6 +66,7 @@ __pycache__
 build
 cmake-build*
 data
+model
 recommonmark
 deps
 

cpp-package/include/mxnet-cpp/monitor.h

@@ -70,8 +70,9 @@ class Monitor {
   /*!
   * \brief install callback to executor. Supports installing to multiple executors.
   * \param exe The executor to install to.
+  * \param monitor_all If true, monitor both input and output, otherwise monitor output only.
   */
-  void install(Executor *exe);
+  void install(Executor *exe, bool monitor_all = false);
 
   /*!
   * \brief Start collecting stats for current batch. Call before calling forward.

cpp-package/include/mxnet-cpp/monitor.hpp

@@ -43,10 +43,10 @@ inline Monitor::Monitor(int interval, std::regex pattern, StatFunc stat_func)
   : interval(interval), pattern(pattern), stat_func(stat_func), step(0) {
 }
 
-inline void Monitor::install(Executor *exe) {
+inline void Monitor::install(Executor *exe, bool monitor_all) {
   MXExecutorSetMonitorCallback(exe->handle_,
-      static_cast<ExecutorMonitorCallback>(&Monitor::executor_callback),
-      this);
+                               static_cast<ExecutorMonitorCallback>(&Monitor::executor_callback),
+                               this, monitor_all);
   exes.push_back(exe);
 }
 

example/quantization/imagenet_gen_qsym_mkldnn.py

@@ -55,24 +55,24 @@ def convert_from_gluon(model_name, image_shape, classes=1000, logger=None):
     symnet = mx.symbol.load_json(y.tojson())
     params = net.collect_params()
     args = {}
-    auxs = {}    
+    auxs = {}
     for param in params.values():
         v = param._reduce()
         k = param.name
         if 'running' in k:
             auxs[k] = v
         else:
-            args[k] = v            
+            args[k] = v
     mod = mx.mod.Module(symbol=symnet, context=mx.cpu(),
                         label_names = ['softmax_label'])
-    mod.bind(for_training=False, 
-             data_shapes=[('data', (1,) + 
+    mod.bind(for_training=False,
+             data_shapes=[('data', (1,) +
                           tuple([int(i) for i in image_shape.split(',')]))])
     mod.set_params(arg_params=args, aux_params=auxs)
     dst_dir = os.path.join(dir_path, 'model')
     prefix = os.path.join(dir_path, 'model', model_name)
     if not os.path.isdir(dst_dir):
-        os.mkdir(dst_dir)       
+        os.mkdir(dst_dir)
     mod.save_checkpoint(prefix, 0)
     return prefix
 
@@ -104,7 +104,7 @@ def save_params(fname, arg_params, aux_params, logger=None):
                              'you can set to custom to load your pre-trained model.')
     parser.add_argument('--use-gluon-model', type=bool, default=False,
                         help='If enabled, will download pretrained model from Gluon-CV '
-                             'and convert to symbolic model ')    
+                             'and convert to symbolic model ')
     parser.add_argument('--batch-size', type=int, default=32)
     parser.add_argument('--label-name', type=str, default='softmax_label')
     parser.add_argument('--calib-dataset', type=str, default='data/val_256_q90.rec',
@@ -114,7 +114,7 @@ def save_params(fname, arg_params, aux_params, logger=None):
                         help='number of threads for data decoding')
     parser.add_argument('--num-calib-batches', type=int, default=10,
                         help='number of batches for calibration')
-    parser.add_argument('--exclude-first-conv', action='store_true', default=True,
+    parser.add_argument('--exclude-first-conv', action='store_true', default=False,
                         help='excluding quantizing the first conv layer since the'
                              ' input data may have negative value which doesn\'t support at moment' )
     parser.add_argument('--shuffle-dataset', action='store_true', default=True,
@@ -140,8 +140,8 @@ def save_params(fname, arg_params, aux_params, logger=None):
                              ' thresholds. This mode is expected to produce the best inference accuracy of all three'
                              ' kinds of quantized models if the calibration dataset is representative enough of the'
                              ' inference dataset.')
-    parser.add_argument('--quantized-dtype', type=str, default='uint8',
-                        choices=['int8', 'uint8'],
+    parser.add_argument('--quantized-dtype', type=str, default='auto',
+                        choices=['auto', 'int8', 'uint8'],
                         help='quantization destination data type for input data')
     parser.add_argument('--enable-calib-quantize', type=bool, default=True,
                         help='If enabled, the quantize op will '
@@ -203,35 +203,30 @@ def save_params(fname, arg_params, aux_params, logger=None):
     if args.model == 'imagenet1k-resnet-152':
         rgb_mean = '0,0,0'
         rgb_std = '1,1,1'
-        calib_layer = lambda name: name.endswith('_output')
-        excluded_sym_names += ['flatten0', 'fc1', 'pooling0']
+        excluded_sym_names += ['flatten0', 'fc1']
         if exclude_first_conv:
             excluded_sym_names += ['conv0']
     elif args.model == 'imagenet1k-inception-bn':
         rgb_mean = '123.68,116.779,103.939'
         rgb_std = '1,1,1'
-        calib_layer = lambda name: name.endswith('_output')
         excluded_sym_names += ['flatten', 'fc1']
         if exclude_first_conv:
             excluded_sym_names += ['conv_1']
     elif args.model in ['resnet50_v1', 'resnet101_v1']:
         rgb_mean = '123.68,116.779,103.939'
         rgb_std = '58.393, 57.12, 57.375'
-        calib_layer = lambda name: name.endswith('_output')
-        excluded_sym_names += ['resnetv10_dense0_fwd', 'resnetv10_pool0_fwd']
+        excluded_sym_names += ['resnetv10_dense0_fwd']
         if exclude_first_conv:
             excluded_sym_names += ['resnetv10_conv0_fwd']
     elif args.model == 'squeezenet1.0':
         rgb_mean = '123.68,116.779,103.939'
         rgb_std = '58.393, 57.12, 57.375'
-        calib_layer = lambda name: name.endswith('_output')
         excluded_sym_names += ['squeezenet0_flatten0_flatten0']
         if exclude_first_conv:
             excluded_sym_names += ['squeezenet0_conv0_fwd']
     elif args.model == 'mobilenet1.0':
         rgb_mean = '123.68,116.779,103.939'
         rgb_std = '58.393, 57.12, 57.375'
-        calib_layer = lambda name: name.endswith('_output')
         excluded_sym_names += ['mobilenet0_flatten0_flatten0',
                                'mobilenet0_dense0_fwd',
                                'mobilenet0_pool0_fwd']
@@ -240,16 +235,13 @@ def save_params(fname, arg_params, aux_params, logger=None):
     elif args.model == 'inceptionv3':
         rgb_mean = '123.68,116.779,103.939'
         rgb_std = '58.393, 57.12, 57.375'
-        calib_layer = lambda name: name.endswith('_output')
-        excluded_sym_names += ['inception30_dense0_fwd',
-                               'inception30_pool0_fwd']
+        excluded_sym_names += ['inception30_dense0_fwd']
         if exclude_first_conv:
             excluded_sym_names += ['inception30_conv0_fwd']
     elif args.model == 'custom':
         # add rgb mean/std of your model.
         rgb_mean = '0,0,0'
         rgb_std = '0,0,0'
-        calib_layer = lambda name: name.endswith('_output')
         # add layer names you donnot want to quantize.
         # add conv/pool layer names that has negative inputs
         # since Intel MKL-DNN only support uint8 quantization temporary.
@@ -302,9 +294,8 @@ def save_params(fname, arg_params, aux_params, logger=None):
                                                         ctx=ctx, excluded_sym_names=excluded_sym_names,
                                                         calib_mode=calib_mode, calib_data=data,
                                                         num_calib_examples=num_calib_batches * batch_size,
-                                                        calib_layer=calib_layer, quantized_dtype=args.quantized_dtype,
-                                                        label_names=(label_name,), calib_quantize_op = True,
-                                                        logger=logger)
+                                                        calib_layer=None, quantized_dtype=args.quantized_dtype,
+                                                        label_names=(label_name,), logger=logger)
         if calib_mode == 'entropy':
             suffix = '-quantized-%dbatches-entropy' % num_calib_batches
         elif calib_mode == 'naive':

example/ssd/quantization.py

@@ -51,7 +51,7 @@ def save_params(fname, arg_params, aux_params, logger=None):
     parser.add_argument('--batch-size', type=int, default=32)
     parser.add_argument('--num-calib-batches', type=int, default=5,
                         help='number of batches for calibration')
-    parser.add_argument('--exclude-first-conv', action='store_true', default=True,
+    parser.add_argument('--exclude-first-conv', action='store_true', default=False,
                         help='excluding quantizing the first conv layer since the'
                              ' number of channels is usually not a multiple of 4 in that layer'
                              ' which does not satisfy the requirement of cuDNN')
@@ -78,8 +78,8 @@ def save_params(fname, arg_params, aux_params, logger=None):
                              ' thresholds. This mode is expected to produce the best inference accuracy of all three'
                              ' kinds of quantized models if the calibration dataset is representative enough of the'
                              ' inference dataset.')
-    parser.add_argument('--quantized-dtype', type=str, default='uint8',
-                        choices=['int8', 'uint8'],
+    parser.add_argument('--quantized-dtype', type=str, default='auto',
+                        choices=['auto', 'int8', 'uint8'],
                         help='quantization destination data type for input data')
 
     args = parser.parse_args()
@@ -119,12 +119,9 @@ def save_params(fname, arg_params, aux_params, logger=None):
     exclude_first_conv = args.exclude_first_conv
     excluded_sym_names = []
     rgb_mean = '123,117,104'
-    calib_layer = lambda name: name.endswith('_output')
     for i in range(1,19):
         excluded_sym_names += ['flatten'+str(i)]
-    excluded_sym_names += ['relu4_3_cls_pred_conv',
-                            'relu7_cls_pred_conv',
-                            'relu4_3_loc_pred_conv']
+
     if exclude_first_conv:
         excluded_sym_names += ['conv1_1']
 
@@ -156,10 +153,8 @@ def save_params(fname, arg_params, aux_params, logger=None):
                                                         ctx=ctx, excluded_sym_names=excluded_sym_names,
                                                         calib_mode=calib_mode, calib_data=eval_iter,
                                                         num_calib_examples=num_calib_batches * batch_size,
-                                                        calib_layer=calib_layer, quantized_dtype=args.quantized_dtype,
-                                                        label_names=(label_name,),
-                                                        calib_quantize_op = True,
-                                                        logger=logger)
+                                                        calib_layer=None, quantized_dtype=args.quantized_dtype,
+                                                        label_names=(label_name,), logger=logger)
         sym_name = '%s-symbol.json' % ('./model/cqssd_vgg16_reduced_300')
         param_name = '%s-%04d.params' % ('./model/cqssd_vgg16_reduced_300', epoch)
     qsym = qsym.get_backend_symbol('MKLDNN_POST_QUANTIZE')

include/mxnet/c_api.h

@@ -1556,13 +1556,12 @@ MXNET_DLL int MXSymbolInferType(SymbolHandle sym,
  * \param num_offline number of parameters that are quantized offline
  * \param offline_params array of c strings representing the names of params quantized offline
  * \param quantized_dtype the quantized destination type for input data.
- * \param calib_quantize whether calibrate quantize op with offline calibration data.
  */
 MXNET_DLL int MXQuantizeSymbol(SymbolHandle sym_handle, SymbolHandle *ret_sym_handle,
                                const mx_uint num_excluded_symbols,
                                const char **excluded_symbols,
                                const mx_uint num_offline, const char **offline_params,
-                               const char *quantized_dtype, const bool calib_quantize);
+                               const char *quantized_dtype);
 
 /*!
  * \brief Set calibration table to node attributes in the sym
@@ -1833,10 +1832,12 @@ MXNET_DLL int MXExecutorGetOptimizedSymbol(ExecutorHandle handle,
 
 /*!
  * \brief set a call back to notify the completion of operation
+ * \param monitor_all If true, monitor both input and output, otherwise monitor output only.
  */
 MXNET_DLL int MXExecutorSetMonitorCallback(ExecutorHandle handle,
                                            ExecutorMonitorCallback callback,
-                                           void* callback_handle);
+                                           void* callback_handle,
+                                           bool monitor_all);
 //--------------------------------------------
 // Part 5: IO Interface
 //--------------------------------------------

include/mxnet/executor.h

@@ -174,7 +174,7 @@ class Executor {
   /*!
    * \brief Install a callback to notify the completion of operation.
    */
-  virtual void SetMonitorCallback(const MonitorCallback& callback) {}
+  virtual void SetMonitorCallback(const MonitorCallback& callback, bool monitor_all) {}
 };  // class executor
 }  // namespace mxnet
 #endif  // MXNET_EXECUTOR_H_

include/mxnet/ndarray.h

@@ -694,9 +694,13 @@ class NDArray {
   /*
    * Create NDArray from mkldnn memory.
    * mkldnn_mem The mkldnn memory to be managed.
-   * static_data If true, mkldnn memory won't be freed on destruction.
    */
-  explicit NDArray(const mkldnn::memory *mkldnn_mem, bool static_data = true);
+  explicit NDArray(const std::shared_ptr<mkldnn::memory> &mkldnn_mem);
+  /*
+   * Create NDArray from mkldnn memory descriptor.
+   * mem_pd The mkldnn memory descriptor to be created.
+   */
+  explicit NDArray(mkldnn::memory::primitive_desc mem_pd);
   /*
    * Test if the data is stored in one of special MKLDNN format.
    */
@@ -776,7 +780,7 @@ class NDArray {
    /*!
    * \ Fix mkldnn memory descriptor mismatch from NDArray.
    */
-  void UpdateMKLDNNMemDesc();
+  void UpdateMKLDNNMemDesc(mkldnn::memory::format format);
 #endif
 
   /*!

include/mxnet/tensor_blob.h

@@ -287,7 +287,7 @@ class TBlob {
     CHECK(Device::kDevMask == this->dev_mask())
       << "TBlob.get: device type do not match specified type";
     CHECK_EQ(this->CheckContiguous(), true) << "TBlob.get_reshape: must be contiguous";
-    CHECK_EQ(this->shape_.Size(), shape.Size())
+    CHECK_EQ(this->shape_.Size(), static_cast<size_t>(shape.Size()))
       << "TBlob.get_with_shape: new and old shape do not match total elements";
     return mshadow::Tensor<Device, dim, DType>(dptr<DType>(), shape,
                                                shape[dim - 1], stream);

perl-package/AI-MXNetCAPI/mxnet.i

@@ -1614,10 +1614,12 @@ int MXExecutorReshape(int partial_shaping,
 
 /*!
  * \brief set a call back to notify the completion of operation
+ * \param monitor_all If true, monitor both input and output, otherwise monitor output only.
  */
 int MXExecutorSetMonitorCallback(ExecutorHandle handle,
                                            ExecutorMonitorCallback callback,
-                                           void* callback_handle);
+                                           void* callback_handle,
+                                           bool monitor_all);
 //--------------------------------------------
 // Part 5: IO Interface
 //--------------------------------------------
@@ -2167,4 +2169,3 @@ int MXRtcCudaKernelCall(CudaKernelHandle handle, int dev_id, void** cuda_kernel_
                                   mx_uint grid_dim_z, mx_uint block_dim_x,
                                   mx_uint block_dim_y, mx_uint block_dim_z,
                                   mx_uint shared_mem);
-