Parallelize CI tests

Summary: Pull Request resolved: facebookresearch#174 Reviewed By: czxttkl Differential Revision: D17956417 fbshipit-source-id: 9ba6059fcec6b463aa96b914bded3e8cbec87b27
davidsc-unity3d · Oct 17, 2019 · adfa6ab · adfa6ab
1 parent 6dd81da
commit adfa6ab
Show file tree

Hide file tree

Showing 13 changed files with 45 additions and 255 deletions.
diff --git a/.circleci/config.yml b/.circleci/config.yml
@@ -28,6 +28,9 @@ binary_common: &binary_common
     resource_class_str:
       description: "Resource class to use"
       type: string
+    test_script:
+      description: "Path to test script"
+      type: string
   environment:
     DOCKER_IMAGE: << parameters.docker_image >>
 
@@ -39,16 +42,18 @@ jobs:
     resource_class: << parameters.resource_class_str >>
     steps:
       - checkout_merge
-      - run: .jenkins/build.sh
+      - run: << parameters.test_script >>
 
 workflows:
   build:
     jobs:
       - binary_linux_conda:
-          name: binary_linux_conda_py3.7_cpu
+          name: linux_conda_py3.7_end_to_end
           docker_image: kittipatv/reagent:cpu
+          test_script: scripts/ci/run_end_to_end_test.sh
           resource_class_str: large
       - binary_linux_conda:
-          name: binary_linux_conda_py3.7_cu92
+          name: linux_conda_py3.7_pytest
           docker_image: kittipatv/reagent:cuda
+          test_script: scripts/ci/run_python_unittest.sh
           resource_class_str: large
diff --git a/.jenkins/README.md b/.jenkins/README.md
diff --git a/docker/jenkins.Dockerfile b/docker/jenkins.Dockerfile
diff --git a/docker/jenkins/README.md b/docker/jenkins/README.md
diff --git a/docker/jenkins/add_jenkins_user.sh b/docker/jenkins/add_jenkins_user.sh
diff --git a/docker/jenkins/build.sh b/docker/jenkins/build.sh
diff --git a/ml/rl/parameters.py b/ml/rl/parameters.py
@@ -286,7 +286,7 @@ class CEMParameters(BaseDataClass):
 @dataclass
 class OpenAiRunDetails(BaseDataClass):
     solved_reward_threshold: Optional[int] = None
-    max_episodes_to_run_after_solved: int = 0
+    max_episodes_to_run_after_solved: Optional[int] = None
     stop_training_after_solved: bool = False
     num_episodes: int = 301
     max_steps: Optional[int] = None
@@ -327,6 +327,7 @@ class OpenAiGymParameters(BaseDataClass):
     actor_training: Optional[FeedForwardParameters] = None
     cem: Optional[CEMParameters] = None
     mdnrnn: Optional[MDNRNNParameters] = None
+    evaluation: EvaluationParameters = EvaluationParameters()
 
 
 #################################################

diff --git a/ml/rl/test/gym/discrete_dqn_cartpole_small_v0.json b/ml/rl/test/gym/discrete_dqn_cartpole_small_v0.json
@@ -6,7 +6,6 @@
   "rl": {
     "gamma": 0.99,
     "target_update_rate": 0.1,
-
     "maxq_learning": true,
     "epsilon": 0.05,
     "temperature": 0.35,
@@ -28,20 +27,25 @@
       "relu",
       "linear"
     ],
-    "minibatch_size": 1024,
-    "learning_rate": 0.001,
+    "minibatch_size": 256,
+    "learning_rate": 0.005,
     "optimizer": "ADAM",
     "lr_decay": 0.999
   },
+  "evaluation": {
+    "calc_cpe_in_training": false
+  },
   "run_details": {
-    "num_episodes": 150,
+    "num_episodes": 70,
     "max_steps": 200,
-    "train_every_ts": 1,
+    "stop_training_after_solved": true,
+    "solved_reward_threshold": 200,
+    "train_every_ts": 2,
     "train_after_ts": 1,
-    "test_every_ts": 2000,
+    "test_every_ts": 400,
     "test_after_ts": 1,
     "num_train_batches": 1,
-    "avg_over_num_episodes": 100,
-    "offline_train_epochs": 30
+    "avg_over_num_episodes": 25,
+    "offline_train_epochs": 10
   }
 }
diff --git a/ml/rl/test/gym/run_gym.py b/ml/rl/test/gym/run_gym.py
@@ -147,9 +147,6 @@ def train(
     run_details: OpenAiRunDetails,
     save_timesteps_to_dataset=None,
     start_saving_from_score=None,
-    solved_reward_threshold=None,
-    max_episodes_to_run_after_solved=None,
-    stop_training_after_solved=False,
     bcq_imitator_hyperparams=None,
     reward_shape_func=None,
 ):
@@ -193,9 +190,9 @@ def train(
             run_details.render,
             save_timesteps_to_dataset,
             start_saving_from_score,
-            solved_reward_threshold,
-            max_episodes_to_run_after_solved,
-            stop_training_after_solved,
+            run_details.solved_reward_threshold,
+            run_details.max_episodes_to_run_after_solved,
+            run_details.stop_training_after_solved,
             reward_shape_func,
         )
 
@@ -515,7 +512,7 @@ def train_gym_online_rl(
 
                 if (
                     solved_reward_threshold is not None
-                    and best_episode_score_seen > solved_reward_threshold
+                    and best_episode_score_seen >= solved_reward_threshold
                 ):
                     solved = True
 
@@ -782,6 +779,7 @@ def create_trainer(params: OpenAiGymParameters, env: OpenAIGymEnvironment):
             rl=rl_parameters,
             training=training_parameters,
             rainbow=params.rainbow,
+            evaluation=params.evaluation,
         )
         trainer = create_dqn_trainer_from_params(
             discrete_trainer_params, env.normalization, use_gpu

diff --git a/ml/rl/workflow/dqn_workflow.py b/ml/rl/workflow/dqn_workflow.py
@@ -177,6 +177,10 @@ def main(params):
             "Horizon is configured to use all GPUs but your platform doesn't support torch.distributed & torch.cuda!"
         )
         params["use_all_avail_gpus"] = False
+    if params["use_gpu"] and not torch.cuda.is_available():
+        logger.info("GPU requested but not available")
+        params["use_gpu"] = False
+
     if params["use_all_avail_gpus"]:
         params["num_processes_per_node"] = max(1, torch.cuda.device_count())
         multiprocessing.spawn(

diff --git a/ml/rl/workflow/sample_configs/discrete_action/dqn_small.json b/ml/rl/workflow/sample_configs/discrete_action/dqn_small.json
@@ -22,7 +22,6 @@
   "rl": {
     "gamma": 0.99,
     "target_update_rate": 0.2,
-
     "maxq_learning": true,
     "epsilon": 0.2,
     "temperature": 0.35,