Support variable-dimension input feature for 2D convolution operation.

paddle/py_paddle/dataprovider_converter.py

@@ -17,6 +17,7 @@
 import swig_paddle
 import numpy
 import itertools
+from functools import reduce
 
 __all__ = ['DataProviderConverter']
 
@@ -65,6 +66,8 @@ def finish_pre_scan(self, argument):
 
         :param argument: Output arguments object.
         :type argument: swig_paddle.Arguments
+        :param dat: Output arguments object.
+        :type dat: The Python object, numpy.array or List.
         :return:
         """
         pass
@@ -95,17 +98,35 @@ class DenseScanner(IScanner):
     def __init__(self, input_type, pos):
         IScanner.__init__(self, input_type, pos)
         self.__mat__ = None
+        self.__shape__ = None
         self.__height__ = 0
+        self.__dim__ = 0
 
     def pre_scan(self, dat):
         self.__height__ += 1
+        if self.__shape__ is None:
+            self.__shape__ = numpy.array(dat).shape
+            if len(self.__shape__) > 3:
+                raise ValueError(
+                    "The dimension of input cannot be greater than 3.")
+            self.__dim__ = reduce(lambda x, y: x * y, self.__shape__)
+            if len(self.__shape__) == 1 and self.__dim__ != self.input_type.dim:
+                raise ValueError(
+                    "The data size must be equal to it in data layer.")
+        else:
+            if self.__shape__ != numpy.array(dat).shape:
+                raise ValueError(
+                    "The data shape must be same in one mini-batch.")
 
     def finish_pre_scan(self, argument):
         self.__mat__ = numpy.ndarray(
-            shape=(self.__height__, self.input_type.dim), dtype=numpy.float32)
+            shape=(self.__height__, self.__dim__), dtype=numpy.float32)
         self.__height__ = 0
 
     def scan(self, dat):
+        # It's better to use NumPy array for speed.
+        dat = numpy.array(dat)
+        dat = dat.flatten()
         self.__mat__[self.__height__] = dat
         self.__height__ += 1
 
@@ -116,6 +137,14 @@ def finish_scan(self, argument):
         m = swig_paddle.Matrix.createDenseFromNumpy(self.__mat__, True,
                                                     self.data_in_gpu)
         argument.setSlotValue(self.pos, m)
+        if len(self.__shape__) > 1:
+            # The last-two dimenstions are the frame height and width.
+            # For example, the layout is CHW for 3-D feature of image.
+            # The H and W are the fram height and width.
+            h, w = self.__shape__[-2:]
+            argument.setSlotFrameHeight(self.pos, h)
+            argument.setSlotFrameWidth(self.pos, w)
+        self.__shape__ = None
 
 
 class SparseBinaryScanner(IScanner):

python/paddle/trainer/PyDataProvider2.py

@@ -72,9 +72,16 @@ def __init__(self, dim, seq_type, tp):
 
 def dense_slot(dim, seq_type=SequenceType.NO_SEQUENCE):
     """
-    Dense Vector. It means the input feature is dense float vector. For example,
-    if the input is an image with 28*28 pixels, the input of Paddle neural
-    network should be a dense vector with dimension 784.
+    Dense Array. It means the input feature is dense array with float type.
+    For example, if the input is an image with 28*28 pixels, the input of
+    Paddle neural network could be a dense vector with dimension 784 or a
+    numpy array with shape (28, 28).
+
+    For the 2-D convolution operation, each sample in one mini-batch must have
+    the similarly size in PaddlePaddle now. But, it supports variable-dimension
+    feature across mini-batch. For the variable-dimension, the param dim is not
+    used. While the data reader must yield numpy array and the data feeder will
+    set the data shape correctly.
 
     :param dim: dimension of this vector.
     :type dim: int
@@ -135,6 +142,10 @@ def index_slot(value_range, seq_type=SequenceType.NO_SEQUENCE):
 sparse_vector = sparse_value_slot
 integer_value = index_slot
 
+# dense_array can be used for variable-length input feature.
+# Each feature is not a vector, but a multi-dimensional array.
+dense_array = dense_slot
+
 
 def dense_vector_sequence(dim):
     """

python/paddle/v2/data_type.py

@@ -16,7 +16,8 @@
 
 import_list = [
     nm for nm in dir(pydp2)
-    if '_' in nm and nm[0] != '_' and ('value' in nm or 'vector' in nm)
+    if '_' in nm and nm[0] != '_' and ('value' in nm or 'vector' in nm or
+                                       'array' in nm)
 ]
 import_list.extend(['InputType'])
 

python/paddle/v2/tests/test_data_feeder.py

@@ -233,6 +233,30 @@ def test_multiple_features_tuple(self):
             self.assertEqual(out_sparse.getSparseRowCols(i), data[i][1])
             self.assertEqual(out_index[i], data[i][0])
 
+    def test_dense_set_shape(self):
+        # test 2-D data
+        def gen_data(batch_size, shape):
+            data = []
+            for i in xrange(batch_size):
+                each_sample = []
+                each_sample.append(np.random.random(shape))
+                data.append(each_sample)
+            return data
+
+        feeder = DataFeeder([('image', data_type.dense_array(2352))],
+                            {'image': 0})
+        arg = feeder(gen_data(32, (3, 28, 28)))
+        h = arg.getSlotFrameHeight(0)
+        w = arg.getSlotFrameWidth(0)
+        self.assertEqual(h, 28)
+        self.assertEqual(w, 28)
+
+        arg = feeder(gen_data(32, (3, 30, 32)))
+        h = arg.getSlotFrameHeight(0)
+        w = arg.getSlotFrameWidth(0)
+        self.assertEqual(h, 30)
+        self.assertEqual(w, 32)
+
 
 if __name__ == '__main__':
     api.initPaddle("--use_gpu=0")