Merge pull request #15100 from jacquesqiao/fix-dist-sparse-decay

fix dist sparse l2 decay

paddle/fluid/operators/math/selected_rows_functor.cc

@@ -195,6 +195,10 @@ struct SelectedRowsAddToTensor<platform::CPUDeviceContext, T> {
   void operator()(const platform::CPUDeviceContext& context,
                   const framework::SelectedRows& input1,
                   framework::Tensor* input2) {
+    if (UNLIKELY(input1.rows().size() == 0)) {
+      LOG(WARNING) << "input selected rows is empty!";
+      return;
+    }
     auto in1_height = input1.height();
     auto in2_dims = input2->dims();
     PADDLE_ENFORCE_EQ(in1_height, in2_dims[0]);

paddle/fluid/operators/sum_op.cc

@@ -41,7 +41,9 @@ class SumOp : public framework::OperatorWithKernel {
       return;  // skip runtime infershape when is tensor array;
     }
 
+    auto x_var_types = ctx->GetInputsVarType("X");
     auto x_dims = ctx->GetInputsDim("X");
+
     size_t N = x_dims.size();
     PADDLE_ENFORCE_GT(N, 0, "Input tensors count should > 0.");
     if (N == 1) {
@@ -49,7 +51,13 @@ class SumOp : public framework::OperatorWithKernel {
     }
 
     framework::DDim in_dim({0});
-    for (auto& x_dim : x_dims) {
+    for (size_t i = 0; i < x_dims.size(); ++i) {
+      auto& x_dim = x_dims[i];
+      // x_dim.size() == 1 means the real dim of selected rows is [0]
+      if (x_var_types[i] == framework::proto::VarType::SELECTED_ROWS &&
+          x_dim.size() == 1) {
+        continue;
+      }
       if (framework::product(x_dim) == 0) {
         continue;
       }

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

@@ -31,6 +31,7 @@ fluid.default_main_program().random_seed = 1
 
 class TestDistCTR2x2(TestDistRunnerBase):
     def get_model(self, batch_size=2):
+
         dnn_input_dim, lr_input_dim = dist_ctr_reader.load_data_meta()
         """ network definition """
         dnn_data = fluid.layers.data(
@@ -97,7 +98,14 @@ class TestDistCTR2x2(TestDistRunnerBase):
 
         inference_program = paddle.fluid.default_main_program().clone()
 
-        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.0001)
+        regularization = None
+        use_l2_decay = bool(os.getenv('USE_L2_DECAY', 0))
+        if use_l2_decay:
+            regularization = fluid.regularizer.L2DecayRegularizer(
+                regularization_coeff=1e-1)
+
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.0001,
+                                            regularization=regularization)
         sgd_optimizer.minimize(avg_cost)
 
         dataset = dist_ctr_reader.Dataset()

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

@@ -235,7 +235,6 @@ class DistSeResneXt2x2(TestDistRunnerBase):
 
         bd = [step * e for e in epochs]
         base_lr = 0.1
-        lr = []
         lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)]
 
         optimizer = fluid.optimizer.Momentum(

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

@@ -18,7 +18,6 @@ import unittest
 from test_dist_base import TestDistBase
 
 
-# FIXME(tangwei): sum op can not handle when inputs is empty.
 class TestDistCTR2x2(TestDistBase):
     def _setup_config(self):
         self._sync_mode = True
@@ -28,5 +27,19 @@ class TestDistCTR2x2(TestDistBase):
         self.check_with_place("dist_ctr.py", delta=1e-7, check_error_log=False)
 
 
+class TestDistCTRWithL2Decay2x2(TestDistBase):
+    def _setup_config(self):
+        self._sync_mode = True
+        self._enforce_place = "CPU"
+
+    def test_dist_ctr(self):
+        need_envs = {"USE_L2_DECAY": "1"}
+        self.check_with_place(
+            "dist_ctr.py",
+            delta=1e-7,
+            check_error_log=False,
+            need_envs=need_envs)
+
+
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -752,12 +752,6 @@ class DistributeTranspiler(object):
             elif op not in lr_ops:
                 self._append_pserver_non_opt_ops(block, op)
 
-        def __op_have_grad_input__(op):
-            for varname in op.input_arg_names:
-                if varname.find("@GRAD") >= 0:
-                    return varname
-            return ""
-
         def __clone_lr_op_sub_block__(op, program, lr_block):
             if not op.has_attr('sub_block'):
                 return
@@ -808,7 +802,7 @@ class DistributeTranspiler(object):
             merged_var = None
             for _, op in enumerate(self.optimize_ops):
                 # find the origin grad var before clipping/L2Decay,
-                # merged_var should be the input var name of L2Decaybuil
+                # merged_var should be the input var name of L2Decay
                 grad_varname_for_block = op.attr(OP_ROLE_VAR_ATTR_NAME)[1]
                 if op.attr(OP_ROLE_VAR_ATTR_NAME)[
                         0] == optimize_target_param_name:
@@ -1684,7 +1678,16 @@ class DistributeTranspiler(object):
                 if self.config.enable_dc_asgd:
                     new_inputs[key] = dc
                 else:
-                    new_inputs[key] = merged_var
+                    # Note!! This is for l2decay on sparse gradient, because it will create a new tensor for
+                    # decayed gradient but not inplace modify the origin one
+                    origin_grad_name = opt_op.input(key)[0]
+                    if core.kNewGradSuffix(
+                    ) in origin_grad_name and pserver_block.has_var(
+                            origin_grad_name):
+                        new_grad = pserver_block.var(origin_grad_name)
+                        new_inputs[key] = new_grad
+                    else:
+                        new_inputs[key] = merged_var
             elif key == "Param":
                 param_block = _get_param_block(opt_op)
                 if not param_block: