!10 Add operator adapting in ME for SparseApplyFtrlD

Merge pull request !10 from zhangzheng/SparseApplyFtrlD

mindspore/ccsrc/transform/convert.cc

@@ -171,6 +171,7 @@ const char kNameAbsGrad[] = "AbsGrad";
 const char kNameBinaryCrossEntropy[] = "BinaryCrossEntropy";
 const char kNameBinaryCrossEntropyGrad[] = "BinaryCrossEntropyGrad";
 const char kNameSparseApplyAdagrad[] = "SparseApplyAdagrad";
+const char kNameSparseApplyFtrlD[] = "SparseApplyFtrlD";
 const char kNameSpaceToDepth[] = "SpaceToDepth";
 const char kNameDepthToSpace[] = "DepthToSpace";
 const char kNameSign[] = "Sign";
@@ -353,6 +354,7 @@ std::unordered_map<std::string, OpAdapterDescPtr> &DfGraphConvertor::get_adpt_ma
     {string(kNameBinaryCrossEntropy), ADPT_DESC(BinaryCrossEntropy)},
     {string(kNameBinaryCrossEntropyGrad), ADPT_DESC(BinaryCrossEntropyGrad)},
     {string(kNameSparseApplyAdagrad), ADPT_DESC(SparseApplyAdagradD)},
+    {string(kNameSparseApplyFtrlD), ADPT_DESC(SparseApplyFtrlD)},
     {string(kNameSpaceToDepth), ADPT_DESC(SpaceToDepth)},
     {string(kNameDepthToSpace), ADPT_DESC(DepthToSpace)},
     {string(kNameSign), ADPT_DESC(Sign)},

mindspore/ccsrc/transform/op_declare.cc

@@ -1120,6 +1120,19 @@ ATTR_MAP(SparseApplyAdagradD) = {{"lr", ATTR_DESC(lr, AnyTraits<float>())},
                                  {"use_locking", ATTR_DESC(use_locking, AnyTraits<bool>())}};
 OUTPUT_MAP(SparseApplyAdagradD) = {{0, OUTPUT_DESC(var)}};
 
+// SparseApplyFtrlD
+INPUT_MAP(SparseApplyFtrlD) = {{1, INPUT_DESC(var)},
+                               {2, INPUT_DESC(accum)},
+                               {3, INPUT_DESC(linear)},
+                               {4, INPUT_DESC(grad)},
+                               {5, INPUT_DESC(indices)}};
+ATTR_MAP(SparseApplyFtrlD) = {{"use_locking", ATTR_DESC(use_locking, AnyTraits<bool>())},
+                              {"lr", ATTR_DESC(lr, AnyTraits<float>())},
+                              {"l1", ATTR_DESC(l1, AnyTraits<float>())},
+                              {"l2", ATTR_DESC(l2, AnyTraits<float>())},
+                              {"lr_power", ATTR_DESC(lr_power, AnyTraits<float>())}};
+OUTPUT_MAP(SparseApplyFtrlD) = {{0, OUTPUT_DESC(var)}};
+
 // SpaceToDepth
 INPUT_MAP(SpaceToDepth) = {{1, INPUT_DESC(x)}};
 ATTR_MAP(SpaceToDepth) = {{"block_size", ATTR_DESC(block_size, AnyTraits<int64_t>())}};

mindspore/ccsrc/transform/op_declare.h

@@ -435,6 +435,8 @@ DECLARE_OP_ADAPTER(Round)
 DECLARE_OP_USE_OUTPUT(Round)
 DECLARE_OP_ADAPTER(ApplyFtrl)
 DECLARE_OP_USE_OUTPUT(ApplyFtrl)
+DECLARE_OP_ADAPTER(SparseApplyFtrlD)
+DECLARE_OP_USE_OUTPUT(SparseApplyFtrlD)
 #ifdef ENABLE_GE
 DECLARE_OP_ADAPTER(Print)
 DECLARE_OP_USE_DYN_INPUT(Print)

mindspore/ops/operations/__init__.py

@@ -65,7 +65,7 @@ from .nn_ops import (LSTM, SGD, Adam, ApplyMomentum, BatchNorm,
                      SmoothL1Loss, Softmax,
                      SoftmaxCrossEntropyWithLogits, ROIAlign,
                      SparseSoftmaxCrossEntropyWithLogits, Tanh,
-                     TopK, BinaryCrossEntropy, SparseApplyAdagrad, LARSUpdate, ApplyFtrl)
+                     TopK, BinaryCrossEntropy, SparseApplyAdagrad, LARSUpdate, ApplyFtrl, SparseApplyFtrlD)
 from .other_ops import Assign, IOU, BoundingBoxDecode, BoundingBoxEncode, CheckValid, MakeRefKey
 
 
@@ -217,6 +217,7 @@ __all__ = [
     "Abs",
     "BinaryCrossEntropy",
     "SparseApplyAdagrad",
+    "SparseApplyFtrlD",
     "SpaceToDepth",
     "DepthToSpace",
     "Conv2DBackpropInput",

mindspore/ops/operations/nn_ops.py

@@ -2141,6 +2141,80 @@ class SparseApplyAdagrad(PrimitiveWithInfer):
         return var_type
 
 
+class SparseApplyFtrlD(PrimitiveWithInfer):
+    r"""
+    Conduct experiment on updating on parameters related to FTRL optimization algorithm.
+
+    .. math ::
+            \text{accum} = \text{grad} * \text{grad}
+
+    .. math ::
+            \text{linear} += \text{grad} + (\text{accum} ^ {\text{-lr_power}} -
+            \frac{\text{accum} ^ \text{-lr_power}}{\text{lr}} * \text{var})
+
+    .. math ::
+            \text{quadratic} = {\text{1.0}/({\text{accum}^\text{lr_power} * \text{lr}}) + 2*\text{l2}
+
+    .. math ::
+            \text{var} = {\text{sign}({linear}) * \text{l1} - \text{linear}})/{ quadratic }
+            if \vert linear \vert > l1 \ else \ 0.0
+
+    Args:
+        lr (float): Learning rate.
+        l1 (float): temp value NO.1.
+        l2 (float): temp value No.2.
+        lr_power (float): temp value used as power number.
+        use_locking (bool): If true, updating the var and accum tensors will be protected. Default: False.
+
+    Inputs:
+       - **var** (Tensor) - Variable to be update. The type must be float32.
+       - **accum** (Tensor) - Accum to be update. The shape must be the same as `var`'s shape,
+         the type must be float32.
+       - **linear** (Tensor) - Linear to be update. The shape must be the same as `var`'s shape,
+         the type must be float32.
+       - **grad** (Tensor) - Gradient. The shape must be the same as `var`'s shape,
+         the type must be float32.
+       - **indices** (Tensor) - A vector of indices into the first dimension of 'var' and 'accum',
+         the shape of `indices` must be the same as `grad` in first dimension, the type must be int32.
+
+    Output:
+        Tensors, has the same shape and type as `var`.
+
+    """
+
+    @prim_attr_register
+    def __init__(self, lr, l1, l2, lr_power, use_locking=False):
+        """init SparseApplyFtrlD"""
+        self.lr = validator.check_type("lr", lr, [float])
+        self.l1 = validator.check_type("l1", l1, [float])
+        self.l2 = validator.check_type("l2", l2, [float])
+        self.lr_power = validator.check_type("lr_power", lr_power, [float])
+        self.use_locking = validator.check_type("use_locking", use_locaking, [bool])
+
+    def infer_shape(self, var_shape, accum_shape, linear_shape, grad_shape, indices_shape):
+        validator.check_param_equal('var shape', var_shape, 'accum shape', accum_shape)
+        validator.check_param_equal('len of var shape', len(var_shape), 'len of grad shape', len(grad_shape))
+        validator.check_param_equal('len of var shape', len(var_shape), 'len of linear shape', len(linear_shape))
+        if len(var_shape) > 1:
+            validator.check_param_equal('var_shape', var_shape[1:], 'grad_shape', grad_shape[1:])
+            validator.check_param_equal('var_shape', var_shape[1:], 'linear_shape', linear_shape[1:])
+        validator.check_integer("len of indices shape", len(indices_shape), 1, Rel.EQ)
+        validator.check('the first dimension of grad', grad_shape[0],
+                        'the shape of indices', indices_shape[0], Rel.EQ)
+
+        return var_shape
+
+    def infer_dtype(self, var_type, accum_type, linear_type, grad_type, indices_type):
+        validator.check_subclass("var_type", var_type, mstype.tensor)
+        validator.check_subclass("accum_type", accum_type, mstype.tensor)
+        validator.check_subclass("linear_type", linear_type, mstype.tensor)
+        validator.check_subclass("grad_type", grad_type, mstype.tensor)
+        validator.check_subclass("indices_type", indices_type, mstype.tensor)
+        validator.check_subclass('indices_type', indices_type, [mstype.int32])
+
+        return var_type
+
+
 class LARSUpdate(PrimitiveWithInfer):
     """
     Conduct lars (layer-wise adaptive rate scaling) update on the square sum of gradient.
@@ -2244,4 +2318,4 @@ class ApplyFtrl(PrimitiveWithInfer):
         validator.check_typename("l1", l1_type,[mstype.float16, mstype.float32])
         validator.check_typename("l2", l2_type,[mstype.float16, mstype.float32])
         validator.check_typename("lr_power", lr_power_type,[mstype.float16, mstype.float32])
-        return var_type
+        return var_type

tests/ut/python/ops/test_ops.py

@@ -749,6 +749,11 @@ test_case_nn_ops = [
         'desc_inputs': [[3, 3], [3, 3], [3, 3], Tensor(np.ones((3,), np.int32))],
         'desc_bprop': [3, 3],
         'skip': ['backward']}),
+    ('SparseApplyFtrlD', {
+        'block': P.SparseApplyFtrlD(0.1, 0.1, 0.1, -0.1),
+        'desc_inputs': [[3, 3], [3, 3], [3, 3], [3, 3], Tensor(2*np.ones((3,), np.int32))],
+        'desc_bprop': [3, 3],
+        'skip': ['backward']}),
     ('Flatten_1', {
         'block': NetForFlatten(),
         'desc_inputs': [Tensor(np.ones([2, 3, 4]).astype(np.int32)), Tensor(np.ones([2, 12]).astype(np.int32))],