[GitHub] ZiyueHuang commented on a change in pull request #9262: FTML optimizer implementation

[GitBox <gi...@apache.org>]

ZiyueHuang commented on a change in pull request #9262: FTML optimizer implementation
URL: https://github.com/apache/incubator-mxnet/pull/9262#discussion_r159150225
 
 

 ##########
 File path: tests/python/unittest/test_optimizer.py
 ##########
 @@ -333,6 +333,75 @@ def test_sparse_sgd():
                             compare_optimizer(opt1(**kwarg), opt2(**kwarg), shape, dtype,
                                               w_stype='row_sparse', g_stype='row_sparse')
 
+
+# FTML
+
+class PyFTML(mx.optimizer.Optimizer):
+    """python reference implemenation of FTML"""
+    def __init__(self, beta1=0.6, beta2=0.999, epsilon=1e-8, **kwargs):
+        super(PyFTML, self).__init__(**kwargs)
+        self.beta1 = beta1
+        self.beta2 = beta2
+        self.epsilon = epsilon
+
+    def create_state(self, index, weight):
+        return (mx.nd.zeros(weight.shape, weight.context, dtype=weight.dtype), # d_0
+                mx.nd.zeros(weight.shape, weight.context, dtype=weight.dtype), # v_0
+                mx.nd.zeros(weight.shape, weight.context, dtype=weight.dtype)) # z_0
+
+    def update(self, index, weight, grad, state):
+        assert(isinstance(weight, mx.nd. NDArray))
+        assert(isinstance(grad, mx.nd.NDArray))
+        self._update_count(index)
+        lr = self._get_lr(index)
+        wd = self._get_wd(index)
+        t = self._index_update_count[index]
+
+        grad = grad * self.rescale_grad
+        if self.clip_gradient is not None:
+            grad = mx.nd.clip(grad, -self.clip_gradient, self.clip_gradient)
+        grad += wd * weight 
 
 Review comment:
   It seems that L2 term is out of clip in other optimizers, such as sgd in https://github.com/apache/incubator-mxnet/blob/master/src/operator/optimizer_op-inl.h#L76-L78.

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services