Add R2D1 agents (#248)

* Add R2D1 DQNAgent Change grandiosely used functions Fix zero-padding & torch contiguous Fix zero-padding & Change indices sampling function Change hyperparameters Remove redundant codes Add CNN compatibility to R2D1Agent Remove redundant code Implement rlpyt forward style Add previous_action & previous_reward GRU input structure Fix error Fix prev_action bug & Use make_one_hot function Fix error Update descriptions & move leading_dims functions to helper_functions.py Move valid_from_done from R2D1Loss to helper_functions.py Fix parameters r2d1_iqn loss & agent Fix GRUBrain compatible with c51 Add R2D1C51Loss Add r2d1_c51 configs Fix priority > 0 assert error Change parameters * Fix descriptions & Rename R2D1 to R2D1DQN Change parameters Add total_step to wandb log Add upndown env & configs Fix test score Fix test score sum to mean Add total step to recurrent dqn_agent Fix test log position Add framestack argument Remove upndown environment Fix no_framestack argument Add r2d1 resnet configs Delete lunarlander iqn & Fix R2D1C51 lunarlander config description Fix configs Change total_step count startpoint after warmup Chage test startpoint Fix epsilon decay Change r2d1 agent epsilon_decay Fix several issues commented * Rebase commit to a302479 * Rebase commit to 990c78a * Fix several issues commented * Delete PrioritizedRecurrentReplayBuffer * Resolved issues commented * Fix issues commented * Change R2D1Learner parent class & Add descriptions * Fix issues commented * Change descriptions * Delete unnecessary configs * Change descriptions due to the length limit * Fix several issues commented * Fix contiguous issues * Add __init__.py in recurrent * Rebase commit to 815a1ca * Use no grad tensor to select action * Modify documentation * Add torch.no_grad to select action in other class * Add R2D1DQN ResNet config * Fix R2D1DQN ResNet config * Merge recurrent_replay_buffer into replay_bufer * Add R2D1 on readme * Remove off-framestack explanation * Change r2d1 configs' framestack 1 to 4 Co-authored-by: khkim <kh.kim@medipixel.io>
medipixel · Jun 23, 2020 · 07743f6 · 07743f6
1 parent 815a1ca
commit 07743f6
Show file tree

Hide file tree

Showing 32 changed files with 1,651 additions and 78 deletions.
diff --git a/README.md b/README.md
@@ -63,6 +63,7 @@ This project follows the [all-contributors](https://github.com/all-contributors/
 7. [Rainbow DQN](https://github.com/medipixel/rl_algorithms/tree/master/rl_algorithms/dqn)
 8. [Rainbow IQN (without DuelingNet)](https://github.com/medipixel/rl_algorithms/tree/master/rl_algorithms/dqn) - DuelingNet [degrades performance](https://github.com/medipixel/rl_algorithms/pull/137)
 9. Rainbow IQN (with [ResNet](https://github.com/medipixel/rl_algorithms/blob/master/rl_algorithms/common/networks/backbones/resnet.py))
+10. [Recurrent Replay DQN (R2D1)](https://github.com/medipixel/rl_algorithms/tree/master/rl_algorithms/recurrent/dqn_agent.py)
 
 
 ## Performance
@@ -204,6 +205,7 @@ python <run-file> -h
     - Start rendering after the number of episodes.
 - `--load-from <save-file-path>`
     - Load the saved models and optimizers at the beginning.
+
 
 #### Show feature map with Grad-CAM
 You can show a feature map that the trained agent extract using **[Grad-CAM(Gradient-weighted Class Activation Mapping)](https://arxiv.org/pdf/1610.02391.pdf)**. Grad-CAM is a way of combining feature maps using the gradient signal, and produce a coarse localization map of the important regions in the image. You can use it by adding [Grad-CAM config](https://github.com/medipixel/rl_algorithms/blob/master/configs/pong_no_frameskip_v4/dqn.py#L39) and `--grad-cam` flag when you run. For example:
@@ -249,3 +251,4 @@ This won't be frequently updated.
 16. [W. Dabney et al., "Implicit Quantile Networks for Distributional Reinforcement Learning." arXiv preprint arXiv:1806.06923, 2018.](https://arxiv.org/pdf/1806.06923.pdf)
 17. [Ramprasaath R. Selvaraju et al., "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization." arXiv preprint arXiv:1610.02391, 2016.](https://arxiv.org/pdf/1610.02391.pdf)
 18. [Kaiming He et al., "Deep Residual Learning for Image Recognition." arXiv preprint arXiv:1512.03385, 2015.](https://arxiv.org/pdf/1512.03385)
+19. [Steven Kapturowski et al., "Recurrent Experience Replay in Distributed Reinforcement Learning." in International Conference on Learning Representations https://openreview.net/forum?id=r1lyTjAqYX, 2019.](https://openreview.net/forum?id=r1lyTjAqYX)
diff --git a/configs/lunarlander_v2/r2d1.py b/configs/lunarlander_v2/r2d1.py
@@ -0,0 +1,52 @@
+"""Config for R2D1 on LunarLander-v2.
+
+- Author: Euijin Jeong
+- Contact: euijin.jeong@medipixel.io
+"""
+from rl_algorithms.common.helper_functions import identity
+
+agent = dict(
+    type="R2D1Agent",
+    hyper_params=dict(
+        gamma=0.99,
+        tau=5e-3,
+        buffer_size=int(1e4),  # openai baselines: int(1e4)
+        batch_size=64,  # openai baselines: 32
+        update_starts_from=int(1e3),  # openai baselines: int(1e4)
+        multiple_update=1,  # multiple learning updates
+        train_freq=1,  # in openai baselines, train_freq = 4
+        gradient_clip=10.0,  # dueling: 10.0
+        n_step=3,
+        w_n_step=1.0,
+        w_q_reg=0.0,
+        per_alpha=0.6,  # openai baselines: 0.6
+        per_beta=0.4,
+        per_eps=1e-6,
+        # R2D1
+        sequence_size=32,
+        overlap_size=16,
+        loss_type=dict(type="R2D1C51Loss"),
+        # Epsilon Greedy
+        max_epsilon=1.0,
+        min_epsilon=0.01,  # openai baselines: 0.01
+        epsilon_decay=2e-5,  # openai baselines: 1e-7 / 1e-1
+    ),
+    learner_cfg=dict(
+        type="R2D1Learner",
+        backbone=dict(),
+        gru=dict(rnn_hidden_size=64, burn_in_step=16,),
+        head=dict(
+            type="C51DuelingMLP",
+            configs=dict(
+                hidden_sizes=[128, 64],
+                v_min=-300,
+                v_max=300,
+                atom_size=51,
+                output_activation=identity,
+                # NoisyNet
+                use_noisy_net=False,
+            ),
+        ),
+        optim_cfg=dict(lr_dqn=1e-4, weight_decay=1e-7, adam_eps=1e-8),
+    ),
+)
diff --git a/configs/pong_no_frameskip_v4/dqn.py b/configs/pong_no_frameskip_v4/dqn.py
@@ -1,4 +1,4 @@
-"""Config for DQN on Pong-No_FrameSkip-v4.
+"""Config for DQN(IQN) on Pong-No_FrameSkip-v4.
 
 - Author: Kyunghwan Kim
 - Contact: kh.kim@medipixel.io

diff --git a/configs/pong_no_frameskip_v4/dqn_resnet.py b/configs/pong_no_frameskip_v4/dqn_resnet.py
@@ -1,4 +1,4 @@
-"""Config for DQN on Pong-No_FrameSkip-v4.
+"""Config for DQN(IQN with ResNet) on Pong-No_FrameSkip-v4.
 
 - Author: Kyunghwan Kim
 - Contact: kh.kim@medipixel.io

diff --git a/configs/pong_no_frameskip_v4/r2d1.py b/configs/pong_no_frameskip_v4/r2d1.py
@@ -0,0 +1,65 @@
+"""Config for R2D1 on PongNoFrameskip-v4.
+
+- Author: Euijin Jeong
+- Contact: euijin.jeong@medipixel.io
+"""
+from rl_algorithms.common.helper_functions import identity
+
+agent = dict(
+    type="R2D1Agent",
+    hyper_params=dict(
+        gamma=0.99,
+        tau=5e-3,
+        buffer_size=int(4e3),  # openai baselines: int(1e4)
+        batch_size=32,  # openai baselines: 32
+        update_starts_from=int(4e3),  # openai baselines: int(1e4)
+        multiple_update=1,  # multiple learning updates
+        train_freq=4,  # in openai baselines, train_freq = 4
+        gradient_clip=10.0,  # dueling: 10.0
+        n_step=5,
+        w_n_step=1.0,
+        w_q_reg=0.0,
+        per_alpha=0.6,  # openai baselines: 0.6
+        per_beta=0.4,
+        per_eps=1e-6,
+        # R2D1
+        sequence_size=20,
+        overlap_size=10,
+        loss_type=dict(type="R2D1DQNLoss"),
+        # Epsilon Greedy
+        max_epsilon=1.0,
+        min_epsilon=0.01,  # openai baselines: 0.01
+        epsilon_decay=3e-6,  # openai baselines: 1e-7 / 1e-1
+        # grad_cam
+        grad_cam_layer_list=[
+            "backbone.cnn.cnn_0.cnn",
+            "backbone.cnn.cnn_1.cnn",
+            "backbone.cnn.cnn_2.cnn",
+        ],
+    ),
+    learner_cfg=dict(
+        type="R2D1Learner",
+        backbone=dict(
+            type="CNN",
+            configs=dict(
+                input_sizes=[4, 32, 64],
+                output_sizes=[32, 64, 64],
+                kernel_sizes=[8, 4, 3],
+                strides=[4, 2, 1],
+                paddings=[1, 0, 0],
+            ),
+        ),
+        gru=dict(rnn_hidden_size=512, burn_in_step=10,),
+        head=dict(
+            type="DuelingMLP",
+            configs=dict(
+                hidden_sizes=[512], use_noisy_net=False, output_activation=identity,
+            ),
+        ),
+        optim_cfg=dict(
+            lr_dqn=1e-4,  # dueling: 6.25e-5, openai baselines: 1e-4
+            weight_decay=0.0,  # this makes saturation in cnn weights
+            adam_eps=1e-8,  # rainbow: 1.5e-4, openai baselines: 1e-8
+        ),
+    ),
+)
diff --git a/configs/pong_no_frameskip_v4/r2d1_dqn_resnet.py b/configs/pong_no_frameskip_v4/r2d1_dqn_resnet.py
@@ -0,0 +1,69 @@
+"""Config for R2D1DQN on PongNoFrameskip-v4.
+- Author: Kyunghwan Kim, Euijin Jeong
+- Contact: kh.kim@medipixel.io, euijin.jeong@medipixel.io
+"""
+from rl_algorithms.common.helper_functions import identity
+
+agent = dict(
+    type="R2D1Agent",
+    hyper_params=dict(
+        gamma=0.99,
+        tau=5e-3,
+        buffer_size=int(4e3),  # openai baselines: int(1e4)
+        batch_size=16,  # openai baselines: 32
+        update_starts_from=int(4e3),  # openai baselines: int(1e4)
+        multiple_update=1,  # multiple learning updates
+        train_freq=4,  # in openai baselines, train_freq = 4
+        gradient_clip=10.0,  # dueling: 10.0
+        n_step=5,
+        w_n_step=1.0,
+        w_q_reg=0.0,
+        per_alpha=0.6,  # openai baselines: 0.6
+        per_beta=0.4,
+        per_eps=1e-6,
+        # R2D1
+        sequence_size=20,
+        overlap_size=10,
+        loss_type=dict(type="R2D1DQNLoss"),
+        # Epsilon Greedy
+        max_epsilon=1.0,
+        min_epsilon=0.01,  # openai baselines: 0.01
+        epsilon_decay=3e-6,  # openai baselines: 1e-7 / 1e-1
+        # grad_cam
+        grad_cam_layer_list=[
+            "backbone.layer1.0.conv2",
+            "backbone.layer2.0.shortcut.0",
+            "backbone.layer3.0.shortcut.0",
+            "backbone.layer4.0.shortcut.0",
+            "backbone.conv_out",
+        ],
+    ),
+    learner_cfg=dict(
+        type="R2D1Learner",
+        backbone=dict(
+            type="ResNet",
+            configs=dict(
+                use_bottleneck=False,
+                num_blocks=[1, 1, 1, 1],
+                block_output_sizes=[32, 32, 64, 64],
+                block_strides=[1, 2, 2, 2],
+                first_input_size=4,
+                first_output_size=32,
+                expansion=1,
+                channel_compression=4,  # compression ratio
+            ),
+        ),
+        gru=dict(rnn_hidden_size=512, burn_in_step=10,),
+        head=dict(
+            type="DuelingMLP",
+            configs=dict(
+                hidden_sizes=[512], use_noisy_net=False, output_activation=identity,
+            ),
+        ),
+        optim_cfg=dict(
+            lr_dqn=1e-4,  # dueling: 6.25e-5, openai baselines: 1e-4
+            weight_decay=0.0,  # this makes saturation in cnn weights
+            adam_eps=1e-8,  # rainbow: 1.5e-4, openai baselines: 1e-8
+        ),
+    ),
+)
diff --git a/rl_algorithms/__init__.py b/rl_algorithms/__init__.py
@@ -19,6 +19,9 @@
 from .fd.sac_learner import SACfDLearner
 from .ppo.agent import PPOAgent
 from .ppo.learner import PPOLearner
+from .recurrent.dqn_agent import R2D1Agent
+from .recurrent.learner import R2D1Learner
+from .recurrent.losses import R2D1C51Loss, R2D1DQNLoss, R2D1IQNLoss
 from .registry import build_agent, build_her
 from .sac.agent import SACAgent
 from .sac.learner import SACLearner
@@ -33,6 +36,7 @@
     "DQNAgent",
     "DDPGfDAgent",
     "DQfDAgent",
+    "R2D1Agent",
     "SACfDAgent",
     "PPOAgent",
     "SACAgent",
@@ -48,6 +52,7 @@
     "PPOLearner",
     "SACLearner",
     "TD3Learner",
+    "R2D1Learner",
     "LunarLanderContinuousHER",
     "ReacherHER",
     "build_agent",
@@ -57,4 +62,7 @@
     "IQNLoss",
     "C51Loss",
     "DQNLoss",
+    "R2D1IQNLoss",
+    "R2D1C51Loss",
+    "R2D1DQNLoss",
 ]
diff --git a/rl_algorithms/common/abstract/agent.py b/rl_algorithms/common/abstract/agent.py
@@ -51,6 +51,7 @@ def __init__(
         self.log_cfg = log_cfg
         self.log_cfg.env_name = env.spec.id if env.spec is not None else env.name
 
+        self.total_step = 0
         self.learner = None
 
         if isinstance(env.action_space, Discrete):
@@ -83,6 +84,7 @@ def set_wandb(self):
             name=f"{self.log_cfg.agent}/{self.log_cfg.curr_time}",
         )
         wandb.config.update(vars(self.args))
+        wandb.config.update(self.hyper_params)
         shutil.copy(self.args.cfg_path, os.path.join(wandb.run.dir, "config.py"))
 
     def interim_test(self):
@@ -122,6 +124,7 @@ def _test(self, interim_test: bool = False):
         else:
             test_num = self.args.episode_num
 
+        score_list = []
         for i_episode in range(test_num):
             state = self.env.reset()
             done = False
@@ -142,9 +145,15 @@ def _test(self, interim_test: bool = False):
             print(
                 "[INFO] test %d\tstep: %d\ttotal score: %d" % (i_episode, step, score)
             )
+            score_list.append(score)
 
-            if self.args.log:
-                wandb.log({"test score": score})
+        if self.args.log:
+            wandb.log(
+                {
+                    "test score": round(sum(score_list) / len(score_list), 2),
+                    "test total step": self.total_step,
+                }
+            )
 
     def test_with_gradcam(self):
         """Test agent with Grad-CAM."""