np.broadcast_to extension

Author: haojin2

src/operator/numpy/np_broadcast_reduce_op_value.cc

@@ -467,17 +467,21 @@ bool NumpyBroadcastToShape(const nnvm::NodeAttrs& attrs,
   mxnet::TShape& ishape = (*in_attrs)[0];
   if (!mxnet::shape_is_known(ishape)) return false;
   const BroadcastToParam& param = nnvm::get<BroadcastToParam>(attrs.parsed);
-  CHECK(mxnet::shape_is_known(param.shape))
-      << "the objective shape for broadcasting array must be known";
   CHECK_LE(ishape.ndim(), param.shape.ndim())
       << "shape " << ishape << " is not broadcastable to " << param.shape;
+  TShape pshape = param.shape;
   for (int i = param.shape.ndim() - 1; i >= 0; --i) {
     int j = i - param.shape.ndim() + ishape.ndim();
     if (j < 0) break;
-    CHECK(ishape[j] == param.shape[i] || ishape[j] == 1)
-        << "shape " << ishape << " is not broadcastable to " << param.shape;
+    if (pshape[i] == -2) {
+      pshape[i] = ishape[j];
+    }
+    CHECK(ishape[j] == pshape[i] || ishape[j] == 1)
+        << "shape " << ishape << " is not broadcastable to " << pshape;
   }
-  SHAPE_ASSIGN_CHECK(*out_attrs, 0, param.shape);
+  CHECK(mxnet::shape_is_known(pshape))
+      << "the objective shape for broadcasting array must be known";
+  SHAPE_ASSIGN_CHECK(*out_attrs, 0, pshape);
   return true;
 }
 

tests/python/unittest/test_numpy_op.py

@@ -1550,6 +1550,7 @@ def hybrid_forward(self, F, x):
         ((4, 1), (1, 2, 3, 4, 5)),
         ((4, 1), (1, 0, 3, 4, 5))
     ]
+
     for src_shape, dst_shape in shapes:
         for hybridize in [True, False]:
             test_broadcast_to = TestBroadcastTo(dst_shape)
@@ -1578,6 +1579,32 @@ def hybrid_forward(self, F, x):
                 ret = test_scalar_broadcast_to(np.empty(()))
             assert_almost_equal(ret.asnumpy(), expected_ret, rtol=1e-5, atol=1e-6, use_broadcast=False)
 
+    # Test npx functionality
+    shapes = [
+        ((5,), (3, 4, -2), (3, 4, 5)),
+        ((5,), (0, -2), (0, 5)),
+        ((1, 0), (2, -2, -2), (2, 1, 0)),
+        ((3, 4), (1, 2, 3, -2), (1, 2, 3, 4)),
+        ((3, 4), (1, 0, -2, 4), (1, 0, 3, 4))
+    ]
+
+    for src_shape, npx_dst_shape, np_dst_shape in shapes:
+        for hybridize in [True, False]:
+            test_broadcast_to = TestBroadcastTo(npx_dst_shape)
+            if hybridize:
+                test_broadcast_to.hybridize()
+
+            a = _np.random.uniform(size=src_shape).astype(np.float32)
+            expected_ret = _np.broadcast_to(a, np_dst_shape)
+            a_mx = np.array(a, dtype=a.dtype)
+            a_mx.attach_grad()
+            with mx.autograd.record():
+                ret = test_broadcast_to(a_mx)
+            assert_almost_equal(ret.asnumpy(), expected_ret, rtol=1e-5, atol=1e-6, use_broadcast=False)
+            ret.backward()
+            expected_grad = collapse_sum_like(_np.ones_like(expected_ret), src_shape)
+            assert_almost_equal(a_mx.grad.asnumpy(), expected_grad, rtol=1e-5, atol=1e-6, use_broadcast=False)
+
 
 @with_seed()
 @use_np