!2668 support weight decay for sparse optimizer

Merge pull request !2668 from wangnan39/support_weight_decay_for_sparse_optimizer

mindspore/nn/optim/adam.py

@@ -164,7 +164,7 @@ class Adam(Optimizer):
 
         The sparse strategy is applied while the SparseGatherV2 operator being used for forward network and the
         `sparse_grad` of `Parameter` being set. The sparse feature is under continuous development. The sparse
-        behavior is currently performed on the CPU, weight decay is not supported.
+        behavior is currently performed on the CPU.
 
     Args:
         params (Union[list[Parameter], list[dict]]): When the `params` is a list of `Parameter` which will be updated,

mindspore/nn/optim/ftrl.py

@@ -73,7 +73,7 @@ class FTRL(Optimizer):
     Note:
         The sparse strategy is applied while the SparseGatherV2 operator being used for forward network and the
         `sparse_grad` of `Parameter` being set. The sparse feature is under continuous development. The sparse
-        behavior is currently performed on the CPU, weight decay is not supported.
+        behavior is currently performed on the CPU.
 
     Args:
         params (list[Parameter]): A list of parameter, which will be updated. The element in `params`
@@ -124,7 +124,7 @@ class FTRL(Optimizer):
         linear = self.linear
         lr = self.learning_rate
         if self.weight_decay > 0.0:
-            grads = self.hyper_map(F.partial(_apply_decay, self.weight_decay), self.decay_tf, params, grads)
+            grads = self.map_(F.partial(_apply_decay, self.weight_decay), self.decay_tf, params, grads)
 
         grads = self.scale_grad(grads)
         success = self.map_(F.partial(_ftrl_opt, self.opt, self.sparse_opt, lr, self.l1, self.l2, self.lr_power),

mindspore/nn/optim/lazyadam.py

@@ -94,8 +94,7 @@ class LazyAdam(Optimizer):
         The sparse strategy is applied while the SparseGatherV2 operator being used for forward network and the
         `sparse_grad` of `Parameter` being set. The sparse behavior, to be notice, is not equivalent to the
         original Adam algorithm, as only the current indices parames will be updated. The sparse feature is under
-        continuous development. The sparse behavior is currently performed on the CPU, weight decay is
-        not supported.
+        continuous development. The sparse behavior is currently performed on the CPU.
 
     Args:
         params (Union[list[Parameter], list[dict]]): When the `params` is a list of `Parameter` which will be updated,

mindspore/nn/optim/optimizer.py

@@ -195,12 +195,12 @@ class Optimizer(Cell):
         params = self.parameters
         if self.is_group:
             if self.exec_weight_decay:
-                gradients = self.hyper_map(F.partial(_apply_decay), self.weight_decay, self.decay_flags,
-                                           params, gradients)
+                gradients = self.map_(F.partial(_apply_decay), self.weight_decay, self.decay_flags,
+                                      params, gradients)
         else:
             if self.weight_decay > 0:
-                gradients = self.hyper_map(F.partial(_apply_decay, self.weight_decay), self.decay_flags,
-                                           params, gradients)
+                gradients = self.map_(F.partial(_apply_decay, self.weight_decay), self.decay_flags,
+                                      params, gradients)
 
         return gradients
 
@@ -479,10 +479,20 @@ class Optimizer(Cell):
 
 
 op_add = P.AddN()
+op_gather = P.GatherV2()
 
 _apply_decay = C.MultitypeFuncGraph("apply_decay")
 
 
+@_apply_decay.register("Number", "Bool", "Tensor", "Tuple")
+def _tensor_apply_decay_with_sparse(weight_decay, if_apply, weight, gradient):
+    """Get grad with weight_decay."""
+    if if_apply:
+        weight = op_gather(weight, gradient[0], 0)
+        return gradient[0], op_add((weight * weight_decay, gradient[1])), gradient[2]
+    return gradient
+
+
 @_apply_decay.register("Number", "Bool", "Tensor", "Tensor")
 def _tensor_apply_decay(weight_decay, if_apply, weight, gradient):
     """Get grad with weight_decay."""

mindspore/nn/optim/proximal_ada_grad.py

@@ -60,7 +60,7 @@ class ProximalAdagrad(Optimizer):
     Note:
         The sparse strategy is applied while the SparseGatherV2 operator being used for forward network and the
         `sparse_grad` of `Parameter` being set as True. The sparse feature is under continuous development. The sparse
-        behavior is currently performed on the CPU, weight decay is not supported.
+        behavior is currently performed on the CPU.
 
     Args:
         params (list[Parameter]): A list of parameter, which will be updated. The element in `params`

tests/ut/python/nn/optim/test_adam.py

@@ -107,7 +107,7 @@ def test_sparse_adam_compile():
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = Adam(net.trainable_params(), learning_rate=0.1, loss_scale=1024.0)
+    optimizer = Adam(net.trainable_params(), learning_rate=0.1, loss_scale=1024.0, weight_decay=0.9)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)
 

tests/ut/python/nn/optim/test_ftrl.py

@@ -71,6 +71,6 @@ def test_spares_ftrl_compile():
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = FTRL(net.trainable_params(), loss_scale=2.0)
+    optimizer = FTRL(net.trainable_params(), weight_decay=0.9, loss_scale=2.0)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)

tests/ut/python/nn/optim/test_lazyadam.py

@@ -75,7 +75,7 @@ def test_spares_lazy_adam_compile():
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1, loss_scale=2.0)
+    optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1, weight_decay=0.9, loss_scale=2.0)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)
 

tests/ut/python/nn/optim/test_proximal_ada_grad.py

@@ -57,7 +57,7 @@ def test_proximal_ada_grad():
     net = Net()
     net.set_train()
     loss = nn.SoftmaxCrossEntropyWithLogits()
-    optimizer = ProximalAdagrad(net.trainable_params())
+    optimizer = ProximalAdagrad(net.trainable_params(), weight_decay=0.9, loss_scale=1024.0)
     net_with_loss = WithLossCell(net, loss)
     train_network = TrainOneStepCell(net_with_loss, optimizer)
     _executor.compile(train_network, inputs, label)
@@ -70,6 +70,6 @@ def test_spares_proximal_ada_grad_compile():
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = ProximalAdagrad(net.trainable_params(), loss_scale=2.0)
+    optimizer = ProximalAdagrad(net.trainable_params(), weight_decay=0.9, loss_scale=1024.0)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)

tests/ut/python/nn/optim/test_rmsprop.py

@@ -57,7 +57,7 @@ def test_rmsprop_compile():
 def test_rmsprop_e():
     net = Net()
     with pytest.raises(ValueError):
-        RMSProp(net.get_parameters(), momentum=-0.1, learning_rate=0.1)
+        RMSProp(net.get_parameters(), momentum=-0.1, learning_rate=0.1, weight_decay=0.9)
 
     with pytest.raises(TypeError):
-        RMSProp(net.get_parameters(), momentum=1, learning_rate=0.1)
+        RMSProp(net.get_parameters(), momentum=1, learning_rate=0.1, weight_decay=0.9)