tuned bvlc_googlenet/lars*

NVIDIA · Nov 2, 2017 · c1e595d · c1e595d
1 parent 759e155
commit c1e595d
Show file tree

Hide file tree

Showing 4 changed files with 82 additions and 77 deletions.
diff --git a/common_plot.py b/common_plot.py
@@ -41,7 +41,7 @@ def get_train_loss(log):
 def get_epochs(log):
     gpus = re.findall(r' GPU (\d*):', log)
     num_gpus = len(gpus)
-    print num_gpus
+    #print num_gpus
     max_iter = re.findall(r'max_iter: (\d*)', log)
     iter_size = re.findall(r'iter_size: (\d*)', log)
     batch_size = re.findall(r'batch_size: (\d*)',log)

diff --git a/models/bvlc_googlenet/solver_lars.prototxt b/models/bvlc_googlenet/solver_lars.prototxt
@@ -1,25 +1,30 @@
 net: "models/bvlc_googlenet/train_val_lars.prototxt"
 
 test_iter:     1562         # 50000/32
-test_interval: 5000    
+test_interval: 150    
 test_initialization: false
 
-display: 500
+display: 50
 
-max_iter: 500000 	# 100 epoch
-#iter_size: 4
+max_iter: 15600 	# 100 epoch
+iter_size: 8
 
-base_lr:   4.00 	# B=256
+base_lr:   6.00 	# B=256
 
 local_lr_auto: true
 local_gw_ratio: 0.001
 
-#rampup_lr: 1.0
-#rampup_interval: 10000
+rampup_lr: 0.2
+rampup_interval: 2000
 
 lr_policy: "poly"
 power: 2.0
+
 momentum: 0.9
+#momentum: 0.8
+#momentum_policy: "poly"
+#max_momentum: 0.92
+
 weight_decay: 1e-4
 
 snapshot: 1000000

diff --git a/models/bvlc_googlenet/train_googlenet_lars.sh b/models/bvlc_googlenet/train_googlenet_lars.sh
@@ -2,4 +2,4 @@
 
 ./build/tools/caffe train \
     --solver=models/bvlc_googlenet/solver_lars.prototxt -gpu=all \
-    2>&1 | tee models/bvlc_googlenet/logs/googlenet_lars_fp16_b256_lr4.log
+    2>&1 | tee models/bvlc_googlenet/logs/googlenet_lars_fp16_b8K_lr6.log