remote min_row_size_to_use_multithread in adam interface test=develop

paddle/fluid/API.spec

@@ -418,7 +418,7 @@ paddle.fluid.optimizer.AdagradOptimizer.__init__ ArgSpec(args=['self', 'learning
 paddle.fluid.optimizer.AdagradOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.optimizer.AdagradOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdagradOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))
-paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode', 'min_row_size_to_use_multithread'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False, 0))
+paddle.fluid.optimizer.AdamOptimizer.__init__ ArgSpec(args=['self', 'learning_rate', 'beta1', 'beta2', 'epsilon', 'regularization', 'name', 'lazy_mode'], varargs=None, keywords=None, defaults=(0.001, 0.9, 0.999, 1e-08, None, None, False))
 paddle.fluid.optimizer.AdamOptimizer.apply_gradients ArgSpec(args=['self', 'params_grads'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.optimizer.AdamOptimizer.backward ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set', 'callbacks'], varargs=None, keywords=None, defaults=(None, None, None, None))
 paddle.fluid.optimizer.AdamOptimizer.minimize ArgSpec(args=['self', 'loss', 'startup_program', 'parameter_list', 'no_grad_set'], varargs=None, keywords=None, defaults=(None, None, None))

paddle/fluid/operators/optimizers/adam_op.cc

@@ -120,7 +120,7 @@ class AdamOpMaker : public framework::OpProtoAndCheckerMaker {
                      "min_row_size_to_use_multithread and "
                      "inner_op_parallelism is larger then 0, sparse update "
                      "will run in multithread mode")
-        .SetDefault(0);
+        .SetDefault(1000);
 
     AddComment(R"DOC(
 Adam Optimizer.

paddle/fluid/operators/optimizers/adam_op.h

@@ -494,16 +494,16 @@ class AdamOpKernel : public framework::OpKernel<T> {
                   << " min_row_size_to_use_multithread="
                   << min_row_size_to_use_multithread;
           if (FLAGS_inner_op_parallelism > 10) {
-            LOG(WARNING) << "FLAGS_inner_op_parallelism "
-                         << FLAGS_inner_op_parallelism << " is two large!";
+            VLOG(1) << "FLAGS_inner_op_parallelism "
+                    << FLAGS_inner_op_parallelism << " is two large!";
           }
           auto& grad_rows = grad_merge.rows();
           std::unordered_map<size_t, int> row_id_to_grad_row_offset;
           size_t param_row_count = param.numel() / row_numel;
           if (param_row_count < 1000) {
-            LOG(WARNING) << "param_row_count should be larger then 1000 to use "
-                            "multi thread, currently "
-                         << param_row_count;
+            VLOG(1) << "param_row_count should be larger then 1000 to use "
+                       "multi thread, currently "
+                    << param_row_count;
           }
           for (size_t i = 0; i < grad_rows.size(); ++i) {
             row_id_to_grad_row_offset[grad_rows[i]] = i;

python/paddle/fluid/optimizer.py

@@ -734,8 +734,6 @@ class AdamOptimizer(Optimizer):
         may be very slow. The lazy mode only update the element that has gradient is the current
         mini-batch, so it will be much more faster. But this mode has different semantics with the
         original Adam algorithm and may lead to different result.
-        min_row_size_to_use_multithread: if adam use sparse update and the param rows is very large,
-                        you can use FLAGS_inner_op_parallelism and this flag to enable multi thread optimize.
 
     Examples:
         .. code-block:: python
@@ -756,8 +754,7 @@ class AdamOptimizer(Optimizer):
                  epsilon=1e-8,
                  regularization=None,
                  name=None,
-                 lazy_mode=False,
-                 min_row_size_to_use_multithread=0):
+                 lazy_mode=False):
         assert learning_rate is not None
         assert beta1 is not None
         assert beta2 is not None
@@ -771,7 +768,6 @@ class AdamOptimizer(Optimizer):
         self._beta2 = beta2
         self._epsilon = epsilon
         self._lazy_mode = lazy_mode
-        self._min_row_size_to_use_multithread = min_row_size_to_use_multithread
 
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
@@ -826,9 +822,7 @@ class AdamOptimizer(Optimizer):
                 "beta1": self._beta1,
                 "beta2": self._beta2,
                 "epsilon": self._epsilon,
-                "lazy_mode": self._lazy_mode,
-                "min_row_size_to_use_multithread":
-                self._min_row_size_to_use_multithread
+                "lazy_mode": self._lazy_mode
             },
             stop_gradient=True)