access2rohit commented on a change in pull request #16715: Lamb optimizer update
URL: https://github.com/apache/incubator-mxnet/pull/16715#discussion_r346107327
##########
File path: tests/python/unittest/test_optimizer.py
##########
@@ -425,6 +425,77 @@ def test_nag():
continue
compare_optimizer(opt1(**kwarg), opt2(**kwarg), shape, dtype,
rtol=1e-3, atol=1e-4)
+
+# LAMB optimizer
+class PyLAMB(mx.optimizer.Optimizer):
+ """
+ Python reference implementation of LAMB optimizer.
+ """
+ def __init__(self, learning_rate=0.001, beta1=0.9, beta2=0.999,
epsilon=1e-6,
+ lower_bound=1e-3, upper_bound=10.0, bias_correction=False,
**kwargs):
+ super(PyLAMB, self).__init__(learning_rate=learning_rate, **kwargs)
+ self.beta1 = beta1
+ self.beta2 = beta2
+ self.epsilon = epsilon
+ self.lower_bound = lower_bound
+ self.upper_bound = upper_bound
+ self.bias_correction = bias_correction
+
+ def create_state(self, index, weight):
+ stype = weight.stype
+ return (mx.nd.zeros(weight.shape, weight.context, dtype=weight.dtype,
stype=stype),
+ mx.nd.zeros(weight.shape, weight.context, dtype=weight.dtype,
stype=stype))
+
+ def update(self, index, weight, grad, state):
+ self._update_count(index)
+ lr = self._get_lr(index)
+ wd = self._get_wd(index)
+ t = self._index_update_count[index]
+
+ grad *= self.rescale_grad
+ if self.clip_gradient is not None:
+ grad = mx.nd.clip(grad, -self.clip_gradient, self.clip_gradient)
+
+ mean, var = state
+ mean[:] = self.beta1 * mean + (1. - self.beta1) * grad
+ var[:] = self.beta2 * var + (1. - self.beta2) * mx.nd.square(grad)
+
+
+ r1 = weight.norm()
+ if not self.bias_correction:
+ r1 = mx.nd.minimum(mx.nd.maximum(r1, self.lower_bound),
self.upper_bound)
+ g = mean / (mx.nd.sqrt(var) + self.epsilon) + wd * weight
+
+ else:
+ mean_hat = mean / (1. - mx.nd.power(self.beta1, t))
+ var_hat = var / (1. - mx.nd.power(self.beta2, t))
+ g = mean_hat / mx.nd.sqrt(var_hat + self.epsilon) + wd * weight
+
+ r2 = g.norm()
+
+ # calculate lamb_trust_ratio
+ r = 1. if r1 == 0. or r2 == 0. else r1 / r2
+ lr *= r
+ # update weight
+ weight[:] -= lr * g
+
+ def update_multi_precision(self, index, weight, grad, state):
+ self.update(index, weight, grad, state)
+
+@with_seed()
+def test_lamb():
+ opt1 = PyLAMB
+ opt2 = mx.optimizer.LAMB
+ shape = (3, 4, 5)
+ cg_options = [{}, {'clip_gradient': 0.4}, {'clip_gradient': 0.5}]
+ rg_options = [{}, {'rescale_grad': 0.14}, {'rescale_grad': 0.8}]
+ wd_options = [{}, {'wd': 0.03}, {'wd': 0.05}, {'wd': 0.07}]
+ bc_options = [{}, {'bias_correction': False}, {'bias_correction': True}]
Review comment:
done
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org
With regards,
Apache Git Services
