Fix the elementwise_div bug in the broadcast process, test=develop (#24987)

Fix the bug for elementwise_div op, when the first var is scalar; Use the shape 1 replace the -1 in shape

python/paddle/fluid/layers/math_op_patch.py

@@ -72,17 +72,23 @@ def monkey_patch_variable():
         block = current_block(ref_var)
         var = create_new_tmp_var(block, dtype)
         batch_dim = -1
+        out_shape = []
         for i, d in enumerate(ref_var.shape):
             if d < 0:
+                if batch_dim < 0:
                     batch_dim = i
-                break
+                    out_shape.append(d)
+                else:
+                    out_shape.append(1)
+            else:
+                out_shape.append(d)
         assert batch_dim != -1
         block.append_op(
             type='fill_constant_batch_size_like',
             outputs={'Out': [var]},
             inputs={'Input': [ref_var]},
             attrs={
-                'shape': ref_var.shape,
+                'shape': out_shape,
                 'value': value,
                 'input_dim_idx': batch_dim,
                 'output_dim_idx': batch_dim

python/paddle/fluid/tests/unittests/test_elementwise_div_op.py

@@ -227,5 +227,18 @@ class TestElementwiseDivOpFp16(ElementwiseDivOp):
             ['X'], 'Out', max_relative_error=1, no_grad_set=set('Y'))
 
 
+class TestElementwiseDivBroadcast(unittest.TestCase):
+    def test_shape_with_batch_sizes(self):
+        with fluid.program_guard(fluid.Program()):
+            x_var = fluid.data(
+                name='x', dtype='float32', shape=[None, 3, None, None])
+            one = 2.
+            out = one / x_var
+            exe = fluid.Executor(fluid.CPUPlace())
+            x = np.random.uniform(0.1, 0.6, (1, 3, 32, 32)).astype("float32")
+            out_result, = exe.run(feed={'x': x}, fetch_list=[out])
+            self.assertEqual((out_result == (2 / x)).all(), True)
+
+
 if __name__ == '__main__':
     unittest.main()