From 3f1169fec2da4d51f9988857b8dea202e1e02b9d Mon Sep 17 00:00:00 2001
From: WangXi <wangxi16@baidu.com>
Date: Wed, 4 Dec 2019 16:24:28 +0800
Subject: [PATCH] Fix dgc clip & rampup step, test=release/1.6 (#21519)

---
 paddle/fluid/operators/dgc_op.h                           | 7 ++++---
 python/paddle/fluid/optimizer.py                          | 4 ++--
 python/paddle/fluid/tests/unittests/test_dgc_optimizer.py | 2 ++
 3 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/paddle/fluid/operators/dgc_op.h b/paddle/fluid/operators/dgc_op.h
index 65aaf47472d..1736fc36f64 100644
--- a/paddle/fluid/operators/dgc_op.h
+++ b/paddle/fluid/operators/dgc_op.h
@@ -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);
diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py
index 168afd4a52c..8d66a810b0b 100644
--- a/python/paddle/fluid/optimizer.py
+++ b/python/paddle/fluid/optimizer.py
@@ -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()
 
diff --git a/python/paddle/fluid/tests/unittests/test_dgc_optimizer.py b/python/paddle/fluid/tests/unittests/test_dgc_optimizer.py
index 996b6ae6ea4..07dda4d5942 100644
--- a/python/paddle/fluid/tests/unittests/test_dgc_optimizer.py
+++ b/python/paddle/fluid/tests/unittests/test_dgc_optimizer.py
@@ -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")
-- 
GitLab