Merge pull request apache#1070 from mli/master

fix apache#1060

example/image-classification/README.md

@@ -180,8 +180,9 @@ recommend to use CUDNN.
 
   | name | hardware | software |
   | --- | --- | --- |
-  | GTX980 | dual Xeon E5-2680 v2, dual GTX 980, 1G Ethernet | GCC 4.8, CUDA 7.5, CUDNN v3 |
-  | EC2-g2.8x | Xeon E5-2670, dual GRID K520, 10G Ethernet | GCC 4.8, CUDA 7.5, CUDNN v3 |
+  | GTX980 | Xeon E5-1650 v3, 4 x GTX 980 | GCC 4.8, CUDA 7.5, CUDNN 3 |
+  | TitanX | dual Xeon E5-2630 v3, 4 x GTX Titan X | GCC 4.8, CUDA 7.5, CUDNN 3 |
+  | EC2-g2.8x | Xeon E5-2670, 2 x GRID K520, 10G Ethernet | GCC 4.8, CUDA 7.5, CUDNN 3 |
 
 - Datasets
 
@@ -210,24 +211,48 @@ python train_cifar10.py --batch-size 128 --lr 0.1 --lr-factor .94 --num-epoch 50
 
 ### ILSVRC 12
 
-#### `train_imagenet.py` with `--network alexnet`
+<!-- #### Alexnet -->
 
-- time for one epoch:
+<!-- `train_imagenet.py` with `--network alexnet` -->
 
-  | 1 x GTX 980 | 2 x GTX 980  | 4 x GTX 980  |
-  | ----------- | ------------ | ------------ |
-  | 2,413 sec | 1,244 sec | 906 sec |
+<!-- - time for one epoch: -->
 
-#### `train_imagenet.py` with `--network inception-bn`
+<!--   | 1 x GTX 980 | 2 x GTX 980  | 4 x GTX 980  | -->
+<!--   | ----------- | ------------ | ------------ | -->
+<!--   | 2,413 sec | 1,244 sec | 906 sec | -->
+
+#### VGG
+
+`train_imagenet.py` with `--network vgg`
+
+- Performance
+
+  | Cluster | # machines | # GPUs | batch size | kvstore | epoch time |
+  | --- | --- | --- | --- | --- | ---: |
+  | TitanX | 1 | 1 | 96 | `none` | 14,545 |
+  | - | - | 2 | - | `local` | 19,692 |
+  | - | - | 4 | - | - | 20,014 |
+  | - | - | 2 | - | `local_allreduce_device` | 9,142 |
+  | - | - | 4 | - | - | 8,533 |
+  | - | - | - | 384 | - | 5,161 |
+
+#### Inception with Batch Normalization
+
+`train_imagenet.py` with `--network inception-bn`
 
 - Performance
 
   | Cluster | # machines | # GPUs | batch size | kvstore | epoch time |
   | --- | --- | --- | --- | --- | ---: |
   | GTX980 | 1 | 1 |  32 | `local` | 13,210 |
-  | - | 1 | 2 |  64 | `local` | 7,198 |
-  | - | 1 | 3 |  128 | `local` | 4,952 |
-  | - | 1 | 4 |  128 | `local` | 3,589 |
+  | - | - | 2 |  64 | - | 7,198 |
+  | - | - | 3 |  128 | - | 4,952 |
+  | - | - | 4 |  - | - | 3,589 |
+  | TitanX | 1 | 1 | 128 | `none` | 10,666 |
+  | - | - | 2 | - | `local` | 5,161 |
+  | - | - | 3 | - | - | 3,460 |
+  | - | - | 4 | - | - | 2,844 |
+  | - | - | - | 512 | - | 2,495 |
   | EC2-g2.8x | 1 | 4 | 144 |  `local` | 14,203 |
   | - | 10 | 40 | 144 |  `dist_sync` | 1,422 |
 
@@ -236,8 +261,8 @@ python train_cifar10.py --batch-size 128 --lr 0.1 --lr-factor .94 --num-epoch 50
   - `single machine` :
 
   ```bash
-  python train_imagenet.py --network inception-bn \
-      --batch-size 128 --lr 0.05 --num-epoch 60 --lr-factor .94 \
+  python train_imagenet.py --batch-size 144 --lr 0.05 --lr-factor .94 \
+      --gpus 0,1,2,3 --num-epoch 60 --network inception-bn \
       --data-dir ilsvrc12/ --model-prefix model/ilsvrc12
   ```
 
@@ -251,7 +276,8 @@ python train_cifar10.py --batch-size 128 --lr 0.1 --lr-factor .94 --num-epoch 50
         --data-dir s3://dmlc/ilsvrc12/  --model-prefix s3://dmlc/model/ilsvrc12
   ```
 
-  *Note: S3 is unstable sometimes, before fixing this problem, we recommend to download data to `/mnt` first*
+  *Note: S3 is unstable sometimes, if your training hangs or getting error
+   freqently, you cant download data to `/mnt` first*
 
   Accuracy vs epoch ([the interactive figure](https://docs.google.com/spreadsheets/d/1AEesHjWUZOzCN0Gp_PYI1Cw4U1kZMKot360p9Fowmjw/pubchart?oid=1740787404&format=interactive)):
 

include/mxnet/c_predict_api.h

@@ -37,6 +37,7 @@ typedef void *NDListHandle;
  * \return The last error happened at the predictor.
  */
 MXNET_DLL const char* MXGetLastError();
+
 /*!
  * \brief create a predictor
  * \param symbol_json_str The JSON string of the symbol.
@@ -65,6 +66,41 @@ MXNET_DLL int MXPredCreate(const char* symbol_json_str,
                            const mx_uint* input_shape_indptr,
                            const mx_uint* input_shape_data,
                            PredictorHandle* out);
+
+/*!
+ * \brief create a predictor wich customized outputs
+ * \param symbol_json_str The JSON string of the symbol.
+ * \param param_bytes The in-memory raw bytes of parameter ndarray file.
+ * \param param_size The size of parameter ndarray file.
+ * \param dev_type The device type, 1: cpu, 2:gpu
+ * \param dev_id The device id of the predictor.
+ * \param num_input_nodes Number of input nodes to the net,
+ *    For feedforward net, this is 1.
+ * \param input_keys The name of input argument.
+ *    For feedforward net, this is {"data"}
+ * \param input_shape_indptr Index pointer of shapes of each input node.
+ *    The length of this array = num_input_nodes + 1.
+ *    For feedforward net that takes 4 dimensional input, this is {0, 4}.
+ * \param input_shape_data A flatted data of shapes of each input node.
+ *    For feedforward net that takes 4 dimensional input, this is the shape data.
+ * \param num_output_nodes Number of output nodes to the net,
+ * \param output_keys The name of output argument.
+ *    For example {"global_pool"}
+ * \param out The created predictor handle.
+ * \return 0 when success, -1 when failure.
+ */
+
+MXNET_DLL int MXPredCreatePartialOut(const char* symbol_json_str,
+                                     const void* param_bytes,
+                                     int param_size,
+                                     int dev_type, int dev_id,
+                                     mx_uint num_input_nodes,
+                                     const char** input_keys,
+                                     const mx_uint* input_shape_indptr,
+                                     const mx_uint* input_shape_data,
+                                     mx_uint num_output_nodes,
+                                     const char** output_keys,
+                                     PredictorHandle* out);
 /*!
  * \brief Get the shape of output node.
  *  The returned shape_data and shape_ndim is only valid before next call to MXPred function.

matlab/+mxnet/model.m

@@ -15,8 +15,12 @@
   predictor
 % the previous input size
   prev_input_size
-% the previous device id, -1 means cpu
+% the previous device id
   prev_dev_id
+% the previous device type (cpu or gpu)
+  prev_dev_type
+% the previous output layers
+  prev_out_layers
 end
 
 methods
@@ -26,6 +30,7 @@
   obj.prev_input_size = zeros(1,4);
   obj.verbose = 1;
   obj.prev_dev_id = -1;
+  obj.prev_dev_type = -1;
   end
 
   function delete(obj)
@@ -64,21 +69,11 @@ function load(obj, model_prefix, num_epoch)
   end
 
 
-  function outputs = forward(obj, imgs, varargin)
+  function outputs = forward(obj, input, varargin)
   %FORWARD perform forward
   %
-  % OUT = MODEL.FORWARD(imgs) returns the forward (prediction) outputs of a list
-  % of images, where imgs can be either a single image with the format
-  %\
-  %   width x height x channel
-  %
-  % which is return format of `imread` or a list of images with format
-  %
-  %   width x height x channel x num_images
-  %
-  % MODEL.FORWARD(imgs, 'gpu', [0, 1]) uses GPU 0 and 1 for prediction
-  %
-  % MODEL.FORWARD(imgs, {'conv4', 'conv5'}) extract outputs for two internal layers
+  % OUT = MODEL.FORWARD(input) returns the forward (prediction) outputs of a list
+  % of input examples
   %
   % Examples
   %
@@ -96,79 +91,135 @@ function load(obj, model_prefix, num_epoch)
   %   imgs(:,:,:,2) = img2
   %   out = model.forward(imgs, 'gpu', [0,1])
 
-  % check arguments
-  assert(length(varargin) == 0, 'sorry, not implemented yet..');
+  % parse arguments
+  dev_type = 1; % cpu in default
+  dev_id = 0;
+  out_layers = {};
+  while length(varargin) > 0
+    if ischar(varargin{1}) && strcmp(varargin{1}, 'gpu')
+      assert(length(varargin) > 1, 'arg error: no gpu id')
+      assert(isnumeric(varargin{2}))
+      dev_type = 2;
+      dev_id = varargin{2};
+      varargin = varargin(3:end);
+      continue
+    end
 
-  % convert from matlab order (col-major) into c order (row major):
-  siz = size(imgs);
-  if length(siz) == 2
-    imgs = permute(imgs, [2, 1]);
-    siz = [siz, 1, 1];
-  elseif length(siz) == 3
-    imgs = permute(imgs, [2, 1, 3]);
-    siz = [siz, 1];
-  elseif length(siz) == 4
-    imgs = permute(imgs, [2, 1, 3, 4]);
-  else
-    error('imgs shape error')
+    if ischar(varargin{1})
+      out_layers{end+1} = varargin{1};
+      varargin = varargin(2:end);
+      continue
+    end
+
+    if iscell(varargin{1})
+      out_layers = varargin{1};
+      varargin = varargin(2:end);
+      continue
+    end
   end
 
-  if any(siz ~= obj.prev_input_size)
+  siz = size(input);
+  assert(length(siz) >= 2);
+
+  % convert from matlab order (col-major) into c order (row major):
+  input = obj.convert_ndarray(input);
+
+  if obj.changed(siz, dev_type, dev_id, out_layers)
     obj.free_predictor()
   end
-  obj.prev_input_size = siz;
 
-  dev_type = 1;
   if obj.predictor.Value == 0
-    if obj.verbose
-      fprintf('create predictor with input size ');
-      fprintf('%d ', siz);
-      fprintf('\n');
-    end
-    callmxnet('MXPredCreate', obj.symbol, ...
+    fprintf('create predictor with input size ');
+    fprintf('%d ', siz);
+    fprintf('\n');
+    csize = [ones(1, 4-length(siz)), siz(end:-1:1)];
+    callmxnet('MXPredCreatePartialOut', obj.symbol, ...
               libpointer('voidPtr', obj.params), ...
               length(obj.params), ...
-              dev_type, 0, ...
+              int32(dev_type), int32(dev_id), ...
               1, {'data'}, ...
               uint32([0, 4]), ...
-              uint32(siz(end:-1:1)), ...
+              uint32(csize), ...
+              uint32(length(out_layers)), out_layers, ...
               obj.predictor);
   end
 
   % feed input
-  callmxnet('MXPredSetInput', obj.predictor, 'data', single(imgs(:)), uint32(numel(imgs)));
+  callmxnet('MXPredSetInput', obj.predictor, 'data', single(input(:)), uint32(numel(input)));
   % forward
   callmxnet('MXPredForward', obj.predictor);
 
-  % get output size
-  out_dim = libpointer('uint32Ptr', 0);
-  out_shape = libpointer('uint32PtrPtr', ones(4,1));
-  callmxnet('MXPredGetOutputShape', obj.predictor, 0, out_shape, out_dim);
-  assert(out_dim.Value <= 4);
-  out_siz = out_shape.Value(1:out_dim.Value);
-  out_siz = double(out_siz(:)');
-
   % get output
-  out = libpointer('singlePtr', single(ones(out_siz)));
+  num_out = 1;
+  if ~isempty(out_layers), num_out = length(out_layers); end
 
-  callmxnet('MXPredGetOutput', obj.predictor, 0, ...
-            out, uint32(prod(out_siz)));
+  if num_out == 1
+    outputs = obj.get_output(0);
+  else
+    outputs = cell(num_out,1);
+    for i = 1 : num_out
+      outputs{i} = obj.get_output(i-1);
+    end
+  end
 
-  % TODO convert from c order to matlab order...
-  outputs = out.Value;
   end
 end
 
 methods (Access = private)
   function free_predictor(obj)
   % free the predictor
   if obj.predictor.Value ~= 0
-    if obj.verbose
-      fprintf('destroy predictor\n')
-    end
     callmxnet('MXPredFree', obj.predictor);
+    obj.predictor = libpointer('voidPtr', 0);
+  end
+  end
+
+  function Y = convert_ndarray(obj, X)
+  % convert between matlab's col major and c's row major
+  siz = size(X);
+  Y = permute(X, [2 1 3:length(siz)]);
+  end
+
+  function ret = changed(obj, input_size, dev_type, dev_id, out_layers)
+  % check if arguments changed since last call
+  ret = 0;
+  if length(input_size) ~= length(obj.prev_input_size) || ...
+        any(input_size ~= obj.prev_input_size) || ...
+        dev_type ~= obj.prev_dev_type || ...
+        length(dev_id) ~= length(obj.prev_dev_id) || ...
+        any(dev_id ~= obj.prev_dev_id) || ...
+        length(out_layers) ~= length(obj.prev_out_layers) || ...
+        ~all(cellfun(@strcmp, out_layers, obj.prev_out_layers))
+    ret = 1;
+  end
+  obj.prev_input_size = input_size;
+  obj.prev_dev_type = dev_type;
+  obj.prev_dev_id = dev_id;
+  obj.prev_out_layers = out_layers;
+  end
+
+  function out = get_output(obj, index)
+  % get the i-th output
+  out_dim = libpointer('uint32Ptr', 0);
+  out_shape = libpointer('uint32PtrPtr', ones(4,1));
+  callmxnet('MXPredGetOutputShape', obj.predictor, index, out_shape, out_dim);
+  assert(out_dim.Value <= 4);
+  out_siz = out_shape.Value(1:out_dim.Value);
+  out_siz = double(out_siz(end:-1:1))';
+
+  % get output
+  out = libpointer('singlePtr', single(zeros(out_siz)));
+
+  callmxnet('MXPredGetOutput', obj.predictor, index, ...
+            out, uint32(prod(out_siz)));
+
+  % TODO convert from c order to matlab order...
+  out = reshape(out.Value, out_siz);
+  if length(out_siz) > 2
+    out = obj.convert_ndarray(out);
   end
   end
+
 end
 
 end