fixing breaking change introduced in #17123 when batch_axis=0

python/mxnet/gluon/utils.py

@@ -77,10 +77,15 @@ def split_data(data, num_slice, batch_axis=0, even_split=True):
         slices = _mx_np.split(data, indices_or_sections=list(div_points[1: -1]), axis=batch_axis)
     else:
         slices = []
-        for i in range(num_slice):
-            st = div_points[i]
-            end = div_points[i + 1]
-            slices.append(ndarray.slice_axis(data, axis=batch_axis, begin=st, end=end))
+        if batch_axis is not 0:
+            for i in range(num_slice):
+                st = div_points[i]
+                end = div_points[i + 1]
+                slices.append(ndarray.slice_axis(data, axis=batch_axis, begin=st, end=end))
+        else:
+            # Fixes issue: https://github.com/apache/incubator-mxnet/issues/19268
+            slices = [data[div_points[i]:div_points[i + 1]] if i < num_slice - 1 else data[div_points[i]:size]
+                      for i in range(num_slice)]
     return slices
 
 

tests/python/unittest/test_gluon.py

@@ -29,6 +29,8 @@
 import mxnet.numpy as _mx_np
 from common import (setup_module, with_seed, assertRaises, teardown,
                     assert_raises_cudnn_not_satisfied, environment)
+import mxnet.ndarray.sparse as mxsps
+import scipy.sparse as sps
 import numpy as np
 from numpy.testing import assert_array_equal
 from nose.tools import raises, assert_raises
@@ -3229,6 +3231,16 @@ def hybrid_forward(self, F, x):
 
     mx.test_utils.assert_almost_equal(grad1, grad2)
 
+
+@with_seed
+def test_split_and_load():
+    ctx_list = (mx.gpu(0), mx.gpu(0))
+    x = mxsps.csr_matrix(sps.coo_matrix(([2.0], ([99], [999]))).tocsr(), ctx=mx.gpu(0))
+    y = nn.utils.split_and_load(x, cts_list)
+    assert hasattr(y[0], 'indices')
+    assert isintanceof(y[0], ndarray.CSRNDArray)
+
+
 if __name__ == '__main__':
     import nose
     nose.runmodule()