edits

.gitignore

@@ -1,4 +1,5 @@
 tmp/
 notebooks/
 __pycache__/
-.ipynb_checkpoints/
+.ipynb_checkpoints/
+scripts/run_*.sh

scripts/run_multiple.sh

@@ -4,12 +4,12 @@ gpuList=(1 2 3 4 5 6 7)
 (
 for i in ${gpuList[@]}; do 
    CUDA_VISIBLE_DEVICES=$i python train.py \
-    --domain_name reacher \
-    --task_name easy \
+    --domain_name ball_in_cup \
+    --task_name catch \
     --encoder_type pixel \
     --action_repeat 4 \
     --save_tb --pre_transform_image_size 100 --image_size 84 \
-    --work_dir ./tmp/curl/reacher_easy \
+    --work_dir ./tmp/curl/ball_in_cup \
     --agent curl_sac --frame_stack 3 \
     --seed -1 --critic_lr 1e-3 --actor_lr 1e-3 --eval_freq 20000 --batch_size 128 --num_train_steps 1000000 &
 done

test_run.py

@@ -0,0 +1,24 @@
+import dmc2gym
+import time
+import tqdm
+
+env = dmc2gym.make(
+        domain_name='reacher',
+        task_name='easy',
+        seed=0,
+        visualize_reward=False,
+        from_pixels=True,
+        height=84,
+        width=84,
+        frame_skip=3
+    )
+
+t = time.time()
+obs = env.reset()
+for i in tqdm.tqdm(range(1000)):
+    a = env.action_space.sample()
+    o,r,d,info = env.step(a)
+    if d:
+        break
+
+print('time',time.time()-t)

train.py

@@ -82,6 +82,7 @@ def evaluate(env, agent, video, num_episodes, L, step, args):
     all_ep_rewards = []
 
     def run_eval_loop(sample_stochastically=True):
+        start_time = time.time()
         prefix = 'stochastic_' if sample_stochastically else ''
         for i in range(num_episodes):
             obs = env.reset()
@@ -104,12 +105,13 @@ def evaluate(env, agent, video, num_episodes, L, step, args):
             L.log('eval/' + prefix + 'episode_reward', episode_reward, step)
             all_ep_rewards.append(episode_reward)
         
+        L.log('eval/' + prefix + 'eval_time', time.time()-start_time , step)
         mean_ep_reward = np.mean(all_ep_rewards)
         best_ep_reward = np.max(all_ep_rewards)
         L.log('eval/' + prefix + 'mean_episode_reward', mean_ep_reward, step)
         L.log('eval/' + prefix + 'best_episode_reward', best_ep_reward, step)
 
-    run_eval_loop(sample_stochastically=True)
+    #run_eval_loop(sample_stochastically=True)
     run_eval_loop(sample_stochastically=False)
     L.dump(step)