Fix broadcast add and subtract grad (#2465)

CMakeLists.txt

@@ -83,7 +83,7 @@ else(MSVC)
   include(CheckCXXCompilerFlag)
   check_cxx_compiler_flag("-std=c++11"    SUPPORT_CXX11)
   if ("${CMAKE_BUILD_TYPE}" STREQUAL "Debug")
-    add_compile_options(-Wall -fPIC -std=c++11)
+    add_compile_options(-O0 -Wall -fPIC -std=c++11)
   else()
     set(CMAKE_C_FLAGS "-O2 -Wall -fPIC ${CMAKE_C_FLAGS}")
     set(CMAKE_CXX_FLAGS "-O2 -Wall -fPIC -std=c++11 ${CMAKE_CXX_FLAGS}")

python/tvm/relay/op/_tensor.py

@@ -5,17 +5,21 @@ import topi
 from .op import register_compute, register_schedule, register_pattern
 from .op import register_gradient
 from .op import schedule_injective, OpPattern
+from .transform import collapse_sum_like
+from .tensor import negative
+
 
 def add_grad(orig, grad):
-    from tvm.relay import op
-    return [op.broadcast_to_like(grad, orig.args[0]), op.broadcast_to_like(grad, orig.args[1])]
+    return [collapse_sum_like(grad, orig.args[0]), collapse_sum_like(grad, orig.args[1])]
+
 
 register_gradient("add", add_grad)
 
+
 def subtract_grad(orig, grad):
-    from tvm.relay import op
-    return [op.broadcast_to_like(grad, orig.args[0]),
-            op.broadcast_to_like(op.negative(grad), orig.args[1])]
+    return [collapse_sum_like(grad, orig.args[0]),
+            collapse_sum_like(negative(grad), orig.args[1])]
+
 
 register_gradient("subtract", subtract_grad)
 

tests/python/relay/test_ad.py

@@ -69,8 +69,64 @@ def test_sub():
     np.testing.assert_allclose(grad.asnumpy(), np.zeros_like(x.asnumpy()))
 
 
+def test_broadcast_add():
+    shape1 = (3, 4, 1)
+    shape2 = (1, 5)
+    dtype = 'float32'
+    x_nd = rand(dtype, *shape1)
+    y_nd = rand(dtype, *shape2)
+    x_np = x_nd.asnumpy()
+    y_np = y_nd.asnumpy()
+    expected_forward = x_np + y_np
+    t1 = relay.TensorType(shape1, dtype)
+    t2 = relay.TensorType(shape2, dtype)
+    x = relay.var("x", t1)
+    y = relay.var("y", t2)
+    func = relay.Function([x, y], x + y)
+    full_func = relay.ir_pass.infer_type(gradient(func))
+    assert full_func.checked_type == relay.FuncType([t1, t2],
+                                                    relay.TupleType([relay.TensorType(expected_forward.shape, dtype),
+                                                                     relay.TupleType([t1, t2])]))
+    ex = create_executor()
+    forward, (grad_x, grad_y) = ex.evaluate(full_func)(x_nd, y_nd)
+    np.testing.assert_allclose(forward.asnumpy(), expected_forward)
+    np.testing.assert_allclose(grad_x.asnumpy(),
+                               np.ones_like(expected_forward).sum(axis=2, keepdims=True))
+    np.testing.assert_allclose(grad_y.asnumpy(),
+                               np.ones_like(expected_forward).sum(axis=(0, 1), keepdims=True).squeeze(axis=0))
+
+
+def test_broadcast_subtract():
+    shape1 = (3, 4, 1)
+    shape2 = (1, 5)
+    dtype = 'float32'
+    x_nd = rand(dtype, *shape1)
+    y_nd = rand(dtype, *shape2)
+    x_np = x_nd.asnumpy()
+    y_np = y_nd.asnumpy()
+    expected_forward = x_np - y_np
+    t1 = relay.TensorType(shape1, dtype)
+    t2 = relay.TensorType(shape2, dtype)
+    x = relay.var("x", t1)
+    y = relay.var("y", t2)
+    func = relay.Function([x, y], x - y)
+    full_func = relay.ir_pass.infer_type(gradient(func))
+    assert full_func.checked_type == relay.FuncType([t1, t2],
+                                                    relay.TupleType([relay.TensorType(expected_forward.shape, dtype),
+                                                                     relay.TupleType([t1, t2])]))
+    ex = create_executor()
+    forward, (grad_x, grad_y) = ex.evaluate(full_func)(x_nd, y_nd)
+    np.testing.assert_allclose(forward.asnumpy(), expected_forward)
+    np.testing.assert_allclose(grad_x.asnumpy(),
+                               np.ones_like(expected_forward).sum(axis=2, keepdims=True))
+    np.testing.assert_allclose(grad_y.asnumpy(),
+                               -np.ones_like(expected_forward).sum(axis=(0, 1), keepdims=True).squeeze(axis=0))
+
+
 if __name__ == "__main__":
     test_id()
     test_add()
     test_temp_add()
     test_sub()
+    test_broadcast_add()
+    test_broadcast_subtract()