[Cherry pick] instance_norm, gradients and batch_norm (#21301)

* Fix gradients (#20857)

* fix_gradients

* fix_gradients, test=develop

* fix instance norm (#21042)

* fix instance norm

* update unitest,test=develop

* fix_bn

* revert unittest,test=develop

paddle/fluid/operators/batch_norm_op.cu

@@ -418,7 +418,7 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
           }
         } else {
           if (d_x) {
-            BNBackwardData<T, block, framework::DataLayout::kNCHW><<<
+            BNBackwardData<T, block, framework::DataLayout::kNHWC><<<
                 grid2, block, 0, dev_ctx.stream()>>>(
                 d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
                 saved_mean_data, x->data<T>(), saved_var_data, C, N, H * W * D,

paddle/fluid/operators/instance_norm_op.cu

@@ -328,7 +328,7 @@ class InstanceNormGradKernel<platform::CUDADeviceContext, T>
           epsilon, saved_mean_data, saved_var_data));
     } else {
       if (d_x) {
-        GradComputeDX<T, block><<<grid, block, 0, dev_ctx.stream()>>>(
+        GradComputeDX<T, block><<<NxC, block, 0, dev_ctx.stream()>>>(
             d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
             saved_mean_data, x->data<T>(), saved_var_data, C, H * W * D,
             d_x->data<T>());

python/paddle/fluid/backward.py

@@ -1247,9 +1247,13 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
         target = targets[i]
         if grad is None:
             grad_name = _append_grad_suffix_(target.name)
-            target_shape = paddle.fluid.layers.shape(target)
+            target_shape = target.name + '_shape'
+            block.desc.append_op().copy_from(
+                _create_op_desc_("shape", {'Input': [target.name]},
+                                 {"Out": [target_shape]}, {}))
+            input_grad_names_set.add(target_shape)
             op_desc = _create_op_desc_("fill_constant",
-                                       {"ShapeTensor": [target_shape.name]},
+                                       {"ShapeTensor": [target_shape]},
                                        {"Out": [grad_name]}, {
                                            "shape": target.shape,
                                            "value": 1.0,

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

@@ -79,7 +79,7 @@ class TestInstanceNormOpTraining(unittest.TestCase):
         self.init_test_case()
 
     def init_test_case(self):
-        self.use_global_stats = False
+        self.shape = [2, 3, 4, 5]
         self.no_grad_set = set()
         self.fetch_list = [
             'y', 'saved_mean', 'saved_variance', 'x@GRAD', 'scale@GRAD',
@@ -181,12 +181,19 @@ class TestInstanceNormOpTraining(unittest.TestCase):
                 "instance_norm"):
             places.append(core.CUDAPlace(0))
         for place in places:
-            test_with_place(place, [2, 3, 4, 5])
+            test_with_place(place, self.shape)
 
 
 class TestInstanceNormOpTrainingCase1(TestInstanceNormOpTraining):
     def init_test_case(self):
-        self.use_global_stats = False
+        self.shape = [2, 3, 4, 5]
+        self.no_grad_set = set(['scale@GRAD', 'bias@GRAD'])
+        self.fetch_list = ['y', 'saved_mean', 'saved_variance', 'x@GRAD']
+
+
+class TestInstanceNormOpTrainingCase2(TestInstanceNormOpTraining):
+    def init_test_case(self):
+        self.shape = [20, 50, 4, 5]
         self.no_grad_set = set(['scale@GRAD', 'bias@GRAD'])
         self.fetch_list = ['y', 'saved_mean', 'saved_variance', 'x@GRAD']