!9967 revert dropout mindir

From: @jjfeing
Reviewed-by: @kisnwang,@zhoufeng54
Signed-off-by: @zhoufeng54

mindspore/ccsrc/backend/session/ascend_session.cc

@@ -439,17 +439,11 @@ void AscendSession::UnifyMindIR(const KernelGraphPtr &graph) {
   }
   auto optimizer = std::make_shared<opt::GraphOptimizer>();
   auto unify_mindir_pm = std::make_shared<opt::PassManager>("unify_mindir_pm");
-  unify_mindir_pm->AddPass(std::make_shared<opt::DropoutGradUnifyMindIR>());
-  unify_mindir_pm->AddPass(std::make_shared<opt::DropoutUnifyMindIR>());
-  unify_mindir_pm->AddPass(std::make_shared<opt::MaxPool2MaxPoolWithArgmax>());
   unify_mindir_pm->AddPass(std::make_shared<opt::MaxPoolWithArgmaxUnifyMindIR>());
   unify_mindir_pm->AddPass(std::make_shared<opt::MaxPoolGradWithArgmaxUnifyMindIR>());
   unify_mindir_pm->AddPass(std::make_shared<opt::Conv2DUnifyMindIR>());
   unify_mindir_pm->AddPass(std::make_shared<opt::Conv2DBackpropInputUnifyMindIR>());
   unify_mindir_pm->AddPass(std::make_shared<opt::Conv2DBackpropFilterUnifyMindIR>());
-  unify_mindir_pm->AddPass(std::make_shared<opt::SparseSoftmaxCrossEntropyWithLogitsUnifyMindIR>());
-  unify_mindir_pm->AddPass(std::make_shared<opt::GradSparseSoftmaxCrossEntropyWithLogitsUnifyMindIR>());
-  unify_mindir_pm->AddPass(std::make_shared<opt::GradSparseSoftmaxCrossEntropyWithLogitsUnifyMindIRV2>());
 
   optimizer->AddPassManager(unify_mindir_pm);
   (void)optimizer->Optimize(graph);

mindspore/nn/layer/basic.py

@@ -141,21 +141,37 @@ class Dropout(Cell):
             raise ValueError("dropout probability should be a number in range (0, 1], but got {}".format(keep_prob))
         Validator.check_subclass("dtype", dtype, mstype.number_type, self.cls_name)
         Validator.check_value_type('keep_prob', keep_prob, [float], self.cls_name)
+        self.keep_prob = keep_prob
         seed0, seed1 = _get_graph_seed(0, "dropout")
         self.seed0 = seed0
         self.seed1 = seed1
-        self.keep_prob = keep_prob
-        self.dropout = P.Dropout(keep_prob, seed0, seed1)
+        self.dtype = dtype
+        self.get_shape = P.Shape()
+        self.dropout_gen_mask = P.DropoutGenMask(Seed0=self.seed0, Seed1=self.seed1)
+        self.dropout_do_mask = P.DropoutDoMask()
+        self.cast = P.Cast()
+        self.is_ascend = context.get_context('device_target') in ["Ascend"]
+        self.dropout = P.Dropout(keep_prob)
 
     def construct(self, x):
         if not self.training:
             return x
 
+        if not self.is_ascend:
+            out, _ = self.dropout(x)
+            return out
+
         if self.keep_prob == 1:
             return x
 
-        out, _ = self.dropout(x)
-        return out
+        shape = self.get_shape(x)
+        dtype = P.DType()(x)
+        if _is_float_dtype(dtype):
+            keep_prob = self.cast(self.keep_prob, dtype)
+        else:
+            keep_prob = self.cast(self.keep_prob, mstype.float16)
+        output = self.dropout_gen_mask(shape, keep_prob)
+        return self.dropout_do_mask(x, output, keep_prob)
 
     def extend_repr(self):
         return 'keep_prob={}, dtype={}'.format(self.keep_prob, self.dtype)

mindspore/nn/layer/pooling.py

@@ -124,9 +124,16 @@ class MaxPool2d(_PoolNd):
                                   strides=self.stride,
                                   padding=self.pad_mode,
                                   data_format=self.format)
+        self.max_pool_with_arg_max = P.MaxPoolWithArgmax(ksize=self.kernel_size,
+                                                         strides=self.stride,
+                                                         padding=self.pad_mode)
+        self.is_tbe = context.get_context("device_target") == "Ascend"
 
     def construct(self, x):
-        out = self.max_pool(x)
+        if self.is_tbe and self.training:
+            out = self.max_pool_with_arg_max(x)[0]
+        else:
+            out = self.max_pool(x)
         return out
 
 
@@ -191,15 +198,22 @@ class MaxPool1d(_PoolNd):
         self.max_pool = P.MaxPool(ksize=self.kernel_size,
                                   strides=self.stride,
                                   padding=self.pad_mode)
+        self.max_pool_with_arg_max = P.MaxPoolWithArgmax(ksize=self.kernel_size,
+                                                         strides=self.stride,
+                                                         padding=self.pad_mode)
         self.shape = F.shape
         self.reduce_mean = P.ReduceMean(keep_dims=True)
         self.expand = P.ExpandDims()
         self.squeeze = P.Squeeze(2)
+        self.is_tbe = context.get_context("device_target") == "Ascend"
 
     def construct(self, x):
         _shape_check(self.shape(x))
         x = self.expand(x, 2)
-        output = self.max_pool(x)
+        if self.is_tbe and self.training:
+            output = self.max_pool_with_arg_max(x)[0]
+        else:
+            output = self.max_pool(x)
         output = self.squeeze(output)
         return output
 

mindspore/nn/loss/loss.py

@@ -267,13 +267,15 @@ class SoftmaxCrossEntropyWithLogits(_Loss):
         self.on_value = Tensor(1.0, mstype.float32)
         self.off_value = Tensor(0., mstype.float32)
         self.is_cpugpu = context.get_context('device_target') in ["CPU", "GPU"]
-        self.sparse_softmax_cross_entropy = P.SparseSoftmaxCrossEntropyWithLogits()
+        if self.is_cpugpu:
+            self.sparse_softmax_cross_entropy = P.SparseSoftmaxCrossEntropyWithLogits()
 
     def construct(self, logits, labels):
+        if self.is_cpugpu and self.sparse and self.reduction == 'mean':
+            x = self.sparse_softmax_cross_entropy(logits, labels)
+            return x
+
         if self.sparse:
-            if self.reduction == 'mean':
-                x = self.sparse_softmax_cross_entropy(logits, labels)
-                return x
             labels = self.one_hot(labels, F.shape(logits)[-1], self.on_value, self.off_value)
         x = self.softmax_cross_entropy(logits, labels)[0]
         return self.get_loss(x)