fix SoftmaxOutput resource bug (apache#14302)

* fix SoftmaxOutput resource

* remove BackwardResource since SoftmaxOutput is not a legacy op

* add test_SoftmaxOutput_normalization

* igore_label=-1 when use_ignore=False

* retrigger CI

* add multi_output test for SoftmaxOutput

* rename test_SoftmaxOutput_normalization to test_softmax_output_normalization

* retrigger CI

* retrigger CI

* fix test bug

src/operator/softmax_output-inl.h

@@ -426,11 +426,6 @@ class SoftmaxOutputProp : public OperatorProperty {
     return {{in_data[softmaxout_enum::kData], out_data[softmaxout_enum::kOut]}};
   }
 
-  std::vector<ResourceRequest> BackwardResource(
-      const mxnet::ShapeVector &in_shape) const override {
-    return {ResourceRequest::kTempSpace};
-  }
-
   Operator* CreateOperator(Context ctx) const override {
     LOG(FATAL) << "Not Implemented.";
     return NULL;

src/operator/softmax_output.cc

@@ -262,6 +262,9 @@ NNVM_REGISTER_OP(_backward_SoftmaxOutput)
 .set_attr<nnvm::FInplaceOption>("FInplaceOption", [](const NodeAttrs& attrs){
   return std::vector<std::pair<int, int> >{{0, 0}};
 })
+.set_attr<FResourceRequest>("FResourceRequest", [](const NodeAttrs& n){
+  return std::vector<ResourceRequest>{ResourceRequest::kTempSpace};
+})
 .set_attr_parser(ParamParser<SoftmaxOutputParam>)
 .set_attr<FCompute>("FCompute<cpu>", SoftmaxOutputGradCompute<cpu>);
 }  // namespace op

tests/python/unittest/test_operator.py

@@ -6496,6 +6496,75 @@ def test_softmax():
     check_smoothed_softmax_grad(default_context())
 
 
+@with_seed()
+def test_softmax_output_normalization():
+    def _softmaxoutput_normalization(multi_output, use_ignore, normalization):
+        grad_scale = np.random.random()
+        batch_size = 8
+        num_labels = 6
+        H, W = 3, 3
+        ignore_label = np.random.randint(0, num_labels) if use_ignore else -1
+
+        if multi_output:
+            data_shape = (batch_size, num_labels, H, W)
+            label_shape = (batch_size, H, W)
+        else:
+            data_shape = (batch_size, num_labels)
+            label_shape = (batch_size, )
+
+        data = mx.nd.random.uniform(-1, 1, shape=data_shape)
+        label = mx.nd.random.randint(
+            0, num_labels, shape=label_shape).astype('float32')
+        data.attach_grad()
+
+        kwargs = dict(grad_scale=grad_scale,
+                      normalization=normalization, multi_output=multi_output)
+        if use_ignore:
+            kwargs.update(use_ignore=True, ignore_label=ignore_label)
+
+        with mx.autograd.record():
+            out = mx.nd.SoftmaxOutput(data=data, label=label, **kwargs)
+        out.backward(mx.nd.ones_like(data))
+
+        exp_data = mx.nd.exp(data)
+        softmax_data = exp_data / exp_data.sum(1, keepdims=True)
+        argmax_data = mx.nd.argmax(data, axis=1)
+
+        assert_almost_equal(out.asnumpy(), softmax_data.asnumpy())
+        one_hot_label = mx.nd.one_hot(label, num_labels)
+        if multi_output:
+            one_hot_label = one_hot_label.transpose((0, 3, 1, 2))
+        data_grad = softmax_data - one_hot_label
+
+        if use_ignore:
+            if multi_output:
+                data_grad *= (label !=
+                              ignore_label).reshape((batch_size, 1, H, W))
+            else:
+                data_grad *= (label != ignore_label).reshape((batch_size, 1))
+
+        valid_cnt = 1
+        if normalization == 'batch':
+            valid_cnt = batch_size
+        elif normalization == 'valid':
+            valid_cnt = mx.nd.maximum(1, (label != ignore_label).sum())
+        scale = grad_scale / valid_cnt
+
+        if multi_output:
+            if normalization != 'valid':
+                scale /= H * W
+
+        data_grad *= scale
+
+        assert_almost_equal(data.grad.asnumpy(), data_grad.asnumpy())
+
+    for multi_output in [False, True]:
+        for use_ignore in [False, True]:
+            for normalization in ['null', 'batch', 'valid']:
+                _softmaxoutput_normalization(
+                    multi_output, use_ignore, normalization)
+
+
 @with_seed()
 def test_slice():
     def test_slice_forward_backward(a, index):
@@ -6533,6 +6602,7 @@ def test_slice_forward_backward(a, index):
     slice_sym = mx.sym.slice(data, begin=[0, None], end=[1, None], step=[2, -1])
     check_numeric_gradient(slice_sym, [in_data])
 
+
 def test_slice_partial_infer():
     def check_slice_partial_infer(data, begin, end, step, expected_out_shape):
         out = mx.sym.slice(data, begin=begin, end=end, step=step)
@@ -6555,6 +6625,7 @@ def check_slice_axis_partial_infer(data, axis, begin, end, expected_out_shape):
     check_slice_axis_partial_infer(var1, 0, 0, 5, (5, 0))
     check_slice_axis_partial_infer(var1, 1, 0, 5, (10, 0))
 
+
 @with_seed()
 def test_float16_min_max():
     """Test for issue: https://github.com/apache/incubator-mxnet/issues/9007"""