Save before modify to autoaug version.

HaojieYuan · Aug 21, 2020 · 3275d13 · 3275d13
1 parent f6c86de
commit 3275d13
Show file tree

Hide file tree

Showing 14 changed files with 5,559 additions and 12 deletions.
diff --git a/branch_num_exp.sh b/branch_num_exp.sh
@@ -1,14 +1,14 @@
 srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=branch1 python -u random_search.py \
-     --opnum 1 --opdepth 1 --policybatch 5 --attackbatch 32 --datasplit 500 --log branch_1.log &
+     --opnum 1 --opdepth 1 --policybatch 3 --attackbatch 32 --datasplit 500 --log branch_1_new.log &
 
 srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=branch3 python -u random_search.py \
-     --opnum 3 --opdepth 1 --policybatch 5 --attackbatch 16 --datasplit 500 --log branch_3.log &
+     --opnum 3 --opdepth 1 --policybatch 3 --attackbatch 16 --datasplit 500 --log branch_3_new.log &
 
 srun -p VI_Face_1080TI -N 1  --gres=gpu:8 --job-name=branch5 python -u random_search.py \
-     --opnum 5 --opdepth 1 --policybatch 5 --attackbatch 8 --datasplit 500 --log branch_5.log &
+     --opnum 5 --opdepth 1 --policybatch 3 --attackbatch 8 --datasplit 500 --log branch_5_new.log &
 
 srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=branch7 python -u random_search.py \
-     --opnum 7 --opdepth 1 --policybatch 5 --attackbatch 8 --datasplit 500 --log branch_7.log &
+     --opnum 7 --opdepth 1 --policybatch 3 --attackbatch 8 --datasplit 500 --log branch_7_new.log &
 
 srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=branch9 python -u random_search.py \
-     --opnum 9 --opdepth 1 --policybatch 5 --attackbatch 4 --datasplit 500 --log branch_9.log &
+     --opnum 9 --opdepth 1 --policybatch 3 --attackbatch 4 --datasplit 500 --log branch_9_new.log &
diff --git a/log/branch_1.log b/log/branch_1.log
diff --git a/log/branch_1_new.log b/log/branch_1_new.log
diff --git a/log/branch_3_new.log b/log/branch_3_new.log
diff --git a/log/branch_5_new.log b/log/branch_5_new.log
diff --git a/log/branch_7_new.log b/log/branch_7_new.log
diff --git a/log/branch_9_new.log b/log/branch_9_new.log
diff --git a/log/op_2.log b/log/op_2.log
diff --git a/log/op_3.log b/log/op_3.log
diff --git a/log/op_4.log b/log/op_4.log
diff --git a/log/op_4_back.log b/log/op_4_back.log
diff --git a/log/op_5.log b/log/op_5.log
diff --git a/op_num_exp.sh b/op_num_exp.sh
@@ -0,0 +1,12 @@
+
+#srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=op3 python -u random_search.py \
+#     --opnum 3 --opdepth 2 --policybatch 3 --attackbatch 16 --datasplit 500 --log op_2.log &
+
+#srun -p VI_Face_1080TI -N 1  --gres=gpu:8 --job-name=op5 python -u random_search.py \
+#     --opnum 3 --opdepth 3 --policybatch 3 --attackbatch 16 --datasplit 500 --log op_3.log &
+
+srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=op7 python -u random_search.py \
+     --opnum 3 --opdepth 4 --policybatch 3 --attackbatch 16 --datasplit 500 --log op_4.log &
+
+#srun -p VI_Face_1080TI -N 1 --gres=gpu:8 --job-name=op9 python -u random_search.py \
+#     --opnum 3 --opdepth 5 --policybatch 3 --attackbatch 16 --datasplit 500 --log op_5.log &
diff --git a/random_search.py b/random_search.py
@@ -40,14 +40,14 @@
 # Search process hyper parameters
 interval = 0.5
 sample_batch = ARGS.policybatch
-lr = 5
-epoch = 300
+lr = 2
+epoch = 100
 
-RESUME = False
-resume = 210
-policy = None
-best_policy = None
-best_reward =  13.447103881835938
+RESUME = True
+resume = 43
+policy = [[9, [0.0, 9, 7.0], [3, 5.0, 9], [5.0, 0, 3.0], [8, 0, 7.0]], [9, [4.0, 9, 0], [1.0, 0, 0], [8, 7.0, 0], [5, 0, 9]], [9, [0, 9, 0], [5.0, 9.0, 2.0], [2.0, 2.0, 5.0], [7.0, 3.0, 5.0]]]
+best_policy = [[9, [2, 9, 8], [0.0, 8, 9], [6.0, 4, 6.0], [7.0, -0.0, 9]], [8.0, [2, 9, 2.0], [0, 3, -0.0], [8, 9, 0], [6, 2.0, 9]], [9, [0, 7.0, 0], [4, 8, 6.0], [2.0, 2.0, 4], [6, 0.0, 7.0]]]
+best_reward =  9.170172293980917
 
 
 
@@ -242,6 +242,7 @@ def single_epoch(policy, reward_getter, lr=0.1, sample_batch=10):
         best_reward = 0
         policy = random_policy()
         resume = -1
+        best_policy = policy.copy()
 
     reward_getter = RewardCal()
 
@@ -256,6 +257,10 @@ def single_epoch(policy, reward_getter, lr=0.1, sample_batch=10):
         policy, reward, epoch_best_policy, epoch_best_reward = single_epoch(policy, reward_getter,
                                                                             lr=lr, sample_batch=sample_batch)
 
+
+        # Eval models are randomrized, recalculate best policy.
+        best_reward = reward_getter.get_reward(best_policy)
+
         if reward > best_reward:
             best_reward = reward
             best_policy = policy