!2071 optimizer support loss scale for sparse situation

Merge pull request !2071 from wangnan39/support_loss_scale_for_sparse_optimizer

mindspore/nn/optim/adam.py

@@ -234,8 +234,6 @@ class Adam(Optimizer):
         _check_param_value(beta1, beta2, eps, weight_decay, self.cls_name)
         validator.check_value_type("use_locking", use_locking, [bool], self.cls_name)
         validator.check_value_type("use_nesterov", use_nesterov, [bool], self.cls_name)
-        validator.check_value_type("loss_scale", loss_scale, [float], self.cls_name)
-        validator.check_number_range("loss_scale", loss_scale, 0.0, float("inf"), Rel.INC_LEFT, self.cls_name)
 
         self.beta1 = Tensor(beta1, mstype.float32)
         self.beta2 = Tensor(beta2, mstype.float32)
@@ -247,9 +245,8 @@ class Adam(Optimizer):
         self.moment2 = self.parameters.clone(prefix="moment2", init='zeros')
 
         self.hyper_map = C.HyperMap()
-        self.map_ = C.Map()
         self.opt = P.Adam(use_locking, use_nesterov)
-        self.sparse_opt = P.SparseApplyAdam()
+        self.sparse_opt = P.SparseApplyAdam(use_locking, use_nesterov)
 
     def construct(self, gradients):
         params = self.parameters

mindspore/nn/optim/ftrl.py

@@ -41,15 +41,11 @@ def _tensor_run_opt(opt, spars_opt, learning_rate, l1, l2, lr_power, linear, gra
     return success
 
 
-def _check_param(initial_accum, learning_rate, lr_power, l1, l2, use_locking, loss_scale=1.0, weight_decay=0.0,
-                 prim_name=None):
+def _check_param(initial_accum, lr_power, l1, l2, use_locking, weight_decay=0.0, prim_name=None):
     """Check param."""
     validator.check_value_type("initial_accum", initial_accum, [float], prim_name)
     validator.check_number("initial_accum", initial_accum, 0.0, Rel.GE, prim_name)
 
-    validator.check_value_type("learning_rate", learning_rate, [float], prim_name)
-    validator.check_number("learning_rate", learning_rate, 0.0, Rel.GT, prim_name)
-
     validator.check_value_type("lr_power", lr_power, [float], prim_name)
     validator.check_number("lr_power", lr_power, 0.0, Rel.LE, prim_name)
 
@@ -61,9 +57,6 @@ def _check_param(initial_accum, learning_rate, lr_power, l1, l2, use_locking, lo
 
     validator.check_value_type("use_locking", use_locking, [bool], prim_name)
 
-    validator.check_value_type("loss_scale", loss_scale, [float], prim_name)
-    validator.check_number("loss_scale", loss_scale, 1.0, Rel.GE, prim_name)
-
     validator.check_value_type("weight_decay", weight_decay, [float], prim_name)
     validator.check_number("weight_decay", weight_decay, 0.0, Rel.GE, prim_name)
 
@@ -110,21 +103,18 @@ class FTRL(Optimizer):
     """
     def __init__(self, params, initial_accum=0.1, learning_rate=0.001, lr_power=-0.5, l1=0.0, l2=0.0,
                  use_locking=False, loss_scale=1.0, weight_decay=0.0):
-        super(FTRL, self).__init__(learning_rate, params)
+        super(FTRL, self).__init__(learning_rate, params, loss_scale=loss_scale)
         if self.is_group:
             raise RuntimeError(f"The {self.cls_name} optimizer cannot support group setting.")
-        _check_param(initial_accum, learning_rate, lr_power, l1, l2, use_locking, loss_scale, weight_decay,
-                     self.cls_name)
+        _check_param(initial_accum, lr_power, l1, l2, use_locking, weight_decay, self.cls_name)
         self.moments = self.parameters.clone(prefix="moments", init=initial_accum)
         self.linear = self.parameters.clone(prefix="linear", init='zeros')
         self.l1 = l1
         self.l2 = l2
         self.lr_power = lr_power
-        self.reciprocal_scale = 1.0 / loss_scale
         self.weight_decay = weight_decay
         self.decay_tf = tuple((lambda: True)() for x in self.parameters)
         self.hyper_map = C.HyperMap()
-        self.map_ = C.Map()
         self.opt = P.ApplyFtrl(use_locking=use_locking)
         self.sparse_opt = P.SparseApplyFtrl(learning_rate, l1, l2, lr_power, use_locking=use_locking)
 
@@ -132,11 +122,11 @@ class FTRL(Optimizer):
         params = self.parameters
         moments = self.moments
         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)
-        if self.reciprocal_scale != 1.0:
-            grads = self.hyper_map(F.partial(grad_scale, self.reciprocal_scale), grads)
-        lr = self.learning_rate
+
+        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),
                             linear, grads, params, moments)
         return success

mindspore/nn/optim/lazyadam.py

@@ -164,8 +164,6 @@ class LazyAdam(Optimizer):
         _check_param_value(beta1, beta2, eps, weight_decay, self.cls_name)
         validator.check_value_type("use_locking", use_locking, [bool], self.cls_name)
         validator.check_value_type("use_nesterov", use_nesterov, [bool], self.cls_name)
-        validator.check_value_type("loss_scale", loss_scale, [float], self.cls_name)
-        validator.check_number_range("loss_scale", loss_scale, 1.0, float("inf"), Rel.INC_LEFT, self.cls_name)
 
         self.beta1 = Tensor(beta1, mstype.float32)
         self.beta2 = Tensor(beta2, mstype.float32)
@@ -179,7 +177,6 @@ class LazyAdam(Optimizer):
         self.moment2 = self.parameters.clone(prefix="moment2", init='zeros')
 
         self.hyper_map = C.HyperMap()
-        self.map_ = C.Map()
         self.opt = P.Adam(use_locking, use_nesterov)
         self.sparse_opt = P.SparseApplyLazyAdam(use_locking, use_nesterov)
 

mindspore/nn/optim/optimizer.py

@@ -153,6 +153,7 @@ class Optimizer(Cell):
         self.reciprocal_scale = 1.0 / loss_scale
         self.exec_weight_decay = any(self.decay_flags)
         self.param_length = len(self.parameters)
+        self.map_ = C.Map()
 
     def decay_weight(self, gradients):
         """
@@ -195,7 +196,7 @@ class Optimizer(Cell):
 
         """
         if self.reciprocal_scale != 1.0:
-            gradients = self.hyper_map(F.partial(grad_scale, self.reciprocal_scale), gradients)
+            gradients = self.map_(F.partial(grad_scale, self.reciprocal_scale), gradients)
 
         return gradients
 
@@ -409,3 +410,11 @@ def tensor_grad_scale(scale, grad):
     if scale == 1.0:
         return grad
     return grad * scale
+
+
+@grad_scale.register("Number", "Tuple")
+def tensor_grad_scale_with_sparse(scale, grad):
+    """Get grad with scale."""
+    if scale == 1.0:
+        return grad
+    return grad[0], grad[1] * scale, grad[2]

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

@@ -18,7 +18,6 @@ import pytest
 
 import mindspore.nn as nn
 from mindspore import Tensor, Parameter
-import mindspore.common.dtype as mstype
 from mindspore.common.api import _executor
 from mindspore.nn import TrainOneStepCell, WithLossCell
 from mindspore.nn.optim import Adam, AdamWeightDecay, AdamWeightDecayDynamicLR
@@ -100,14 +99,14 @@ def test_adam_compile():
     _executor.compile(train_network, inputs, label)
 
 
-def test_spares_adam_compile():
+def test_sparse_adam_compile():
     """ test_sparse_adam_compile """
     indices = Tensor(np.array([0, 1]).astype(np.int32))
     label = Tensor(np.zeros([2, 1, 2]).astype(np.float32))
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = Adam(net.trainable_params(), learning_rate=0.1)
+    optimizer = Adam(net.trainable_params(), learning_rate=0.1, loss_scale=1024.0)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)
 
@@ -149,34 +148,3 @@ def test_adam_mindspore_with_empty_params():
     net = nn.Flatten()
     with pytest.raises(ValueError, match=r"Optimizer got an empty parameter list"):
         AdamWeightDecay(net.get_parameters())
-
-
-class TestSparseOps(nn.Cell):
-    """Define sparse operator"""
-    def __init__(self, sparse_opt):
-        super(TestSparseOps, self).__init__()
-        self.sparse_apply_adam = sparse_opt
-        self.var = Parameter(Tensor(np.ones([3, 3, 3]).astype(np.float32)), name="var")
-        self.m = Parameter(Tensor(np.ones([3, 3, 3]).astype(np.float32)), name="m")
-        self.v = Parameter(Tensor(np.ones([3, 3, 3]).astype(np.float32)), name="v")
-
-    def construct(self, beta1_power, beta2_power, lr, beta1, beta2, epsilon, grad, indices):
-        out = self.sparse_apply_adam(self.var, self.m, self.v, beta1_power, beta2_power, lr, beta1, beta2, epsilon,
-                                     grad, indices)
-        return out
-
-
-def test_sparse_adam():
-    """test sparse operator"""
-    gradient = Tensor(np.random.rand(3, 3, 3).astype(np.float32))
-    indices = Tensor([0, 1, 2], mstype.int32)
-    net = TestSparseOps(P.SparseApplyAdam())
-    _executor.compile(net, 0.9, 0.999, 0.001, 0.9, 0.999, 1e-8, gradient, indices)
-
-
-def test_sparse_lazy_adam():
-    """test sparse operator"""
-    gradient = Tensor(np.random.rand(3, 3, 3).astype(np.float32))
-    indices = Tensor([0, 1, 2], mstype.int32)
-    net = TestSparseOps(P.SparseApplyLazyAdam())
-    _executor.compile(net, 0.9, 0.999, 0.001, 0.9, 0.999, 1e-8, gradient, indices)

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

@@ -57,7 +57,7 @@ def test_ftrl():
     net = Net()
     net.set_train()
     loss = nn.SoftmaxCrossEntropyWithLogits()
-    optimizer = FTRL(net.trainable_params())
+    optimizer = FTRL(net.trainable_params(), weight_decay=0.9, loss_scale=2.0)
     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_ftrl_compile():
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = FTRL(net.trainable_params())
+    optimizer = FTRL(net.trainable_params(), loss_scale=2.0)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)

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

@@ -60,7 +60,7 @@ def test_lazy_adam_compile():
     net.set_train()
 
     loss = nn.SoftmaxCrossEntropyWithLogits()
-    optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1, weight_decay=0.9)
+    optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1, weight_decay=0.9, loss_scale=2.0)
 
     net_with_loss = WithLossCell(net, loss)
     train_network = TrainOneStepCell(net_with_loss, optimizer)
@@ -74,7 +74,7 @@ def test_spares_lazy_adam_compile():
     net = NetWithSparseGatherV2()
     net.set_train()
 
-    optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1)
+    optimizer = LazyAdam(net.trainable_params(), learning_rate=0.1, loss_scale=2.0)
     train_network = TrainOneStepCell(net, optimizer)
     _executor.compile(train_network, indices, label)