提交 3f1169fe 编写于 作者: W WangXi 提交者: gongweibao

Fix dgc clip & rampup step, test=release/1.6 (#21519)

上级 0e63746b
......@@ -28,7 +28,7 @@ inline float get_period_sparcity(const std::vector<float>& sparsity,
size_t idx = static_cast<int>(cur_step * sparsity.size() / rampup_steps);
if (idx >= sparsity.size()) {
return 0.999;
idx = sparsity.size() - 1;
}
PADDLE_ENFORCE_LT(idx, sparsity.size());
......@@ -102,8 +102,9 @@ class DGCOpKernel : public framework::OpKernel<T> {
}
float ratio =
1 - get_period_sparcity(sparsity, static_cast<float>(*current_step),
rampup_step);
1 - get_period_sparcity(
sparsity, static_cast<float>(*current_step - rampup_begin_step),
rampup_step);
PADDLE_ENFORCE_GE(ratio, 0.0);
PADDLE_ENFORCE_LT(ratio, 1.0);
int k = static_cast<int>(g->numel() * ratio);
......
......@@ -947,6 +947,7 @@ class DGCMomentumOptimizer(Optimizer):
self._momentum = momentum
self._use_nesterov = bool(use_nesterov)
assert rampup_begin_step >= 0, "rampup_begin_step must >= 0"
self._rampup_begin_step = rampup_begin_step
self._rampup_step = rampup_step
self._sparsity = sparsity
......@@ -963,8 +964,7 @@ class DGCMomentumOptimizer(Optimizer):
self._local_grad_clip_norm = local_grad_clip_norm
self._num_trainers = num_trainers
self._clip_norm = local_grad_clip_norm / (num_trainers *
num_trainers)
self._clip_norm = local_grad_clip_norm * (num_trainers**-0.5)
self._get_dgc_regularization_param()
......
......@@ -67,6 +67,8 @@ class TestDGCMomentumOptimizer(unittest.TestCase):
learning_rate=learning_rate,
momentum=0.2,
rampup_begin_step=0,
local_grad_clip_norm=1.0,
num_trainers=2,
regularization=regularization)
mean_out = block.create_var(
dtype="float32", shape=[1], lod_level=0, name="mean.out")
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册