gpu add test for amp

mindspore/ccsrc/kernel/gpu/arrays/slice_gpu_kernel.h

@@ -129,10 +129,10 @@ class SliceGpuFwdKernel : public GpuKernel {
     }
     begin_ = GetAttr<std::vector<int>>(kernel_node, "begin");
     for (size_t i = 0; i < input_shape.size(); i++) {
-      if ((begin_[i] > 0 && (begin_[i] >= SizeToInt(input_shape[i]))) ||
+      if ((begin_[i] > 0 && (begin_[i] > SizeToInt(input_shape[i]))) ||
           (begin_[i] < 0 && (std::abs(begin_[i]) > SizeToInt(input_shape[i])))) {
-        MS_LOG(ERROR) << "Error input, out of bounds " << input_shape[i] << " in axis " << i << ".";
-        return false;
+        MS_LOG(INFO) << "Input out of bounds " << input_shape[i] << " in axis " << i << ".";
+        begin_[i] = 0;
       }
     }
     return true;

tests/st/networks/test_gpu_resnet.py

@@ -32,9 +32,7 @@ from mindspore.nn.optim import Momentum
 from mindspore.ops import operations as P
 from mindspore.nn import TrainOneStepCell, WithLossCell
 from mindspore.nn import Dense
-from mindspore.common.initializer import initializer
-
-context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
+from mindspore import amp
 
 
 def random_normal_init(shape, mean=0.0, stddev=0.01, seed=None):
@@ -326,6 +324,7 @@ def resnet50(num_classes):
 @pytest.mark.platform_x86_gpu_training
 @pytest.mark.env_onecard
 def test_trainTensor(num_classes=10, epoch=8, batch_size=1):
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
     net = resnet50(num_classes)
     lr = 0.1
     momentum = 0.9
@@ -341,3 +340,26 @@ def test_trainTensor(num_classes=10, epoch=8, batch_size=1):
         loss = train_network(data, label)
         losses.append(loss)
     assert(losses[-1].asnumpy() < 1)
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_trainTensor_amp(num_classes=10, epoch=18, batch_size=16):
+    context.set_context(mode=context.GRAPH_MODE, device_target="GPU", enable_mem_reuse=False,
+                        enable_dynamic_memory=False)
+    net = resnet50(num_classes)
+    lr = 0.1
+    momentum = 0.9
+    optimizer = Momentum(filter(lambda x: x.requires_grad, net.get_parameters()), lr, momentum)
+    criterion = nn.SoftmaxCrossEntropyWithLogits(is_grad=False, sparse=True)
+    train_network = amp.build_train_network(net, optimizer, criterion, level="O2")
+    train_network.set_train()
+    losses = []
+    for i in range(0, epoch):
+        data = Tensor(np.ones([batch_size, 3, 224, 224]).astype(np.float32) * 0.01)
+        label = Tensor(np.ones([batch_size]).astype(np.int32))
+        loss = train_network(data, label)
+        losses.append(loss)
+    assert(losses[-1][0].asnumpy() < 1)
+    assert(losses[-1][1].asnumpy() == False)
+    assert(losses[-1][2].asnumpy() > 1)