#51 started adjusting DDPG

experiments/issue51_new/stable_baselinesDDPG.py

@@ -2,6 +2,10 @@
 from os import makedirs
 from typing import List
 
+import torch as th
+import torch.nn as nn
+from stable_baselines3.common.torch_layers import BaseFeaturesExtractor
+
 import gym
 import numpy as np
 from stable_baselines3 import DDPG
@@ -110,13 +114,47 @@ def _on_step(self) -> bool:
 n_actions = env.action_space.shape[-1]
 action_noise = NormalActionNoise(mean=np.zeros(n_actions), sigma=0.1 * np.ones(n_actions))
 
-model = DDPG('MlpPolicy', env, verbose=1, tensorboard_log=f'{timestamp}/')
-checkpoint_on_event = CheckpointCallback(save_freq=100000, save_path=f'{timestamp}/checkpoints/')
+
+class CustomMPL(BaseFeaturesExtractor):
+
+    def __init__(self, observation_space: gym.spaces.Box, features_dim: int = 256):
+        super(CustomMPL, self).__init__(observation_space, features_dim)
+        # We assume CxHxW images (channels first)
+        # Re-ordering will be done by pre-preprocessing or wrapper
+        n_input_channels = observation_space.shape[0]
+        self.cnn = nn.Sequential(
+            nn.Linear(n_input_channels, 32),
+            nn.ReLU(),
+            nn.Linear(32, 64),
+            nn.ReLU(),
+        )
+
+        # Compute shape by doing one forward pass
+        with th.no_grad():
+            n_flatten = self.cnn(
+                th.as_tensor(observation_space.sample()[None]).float()
+            ).shape[1]
+
+        self.linear = nn.Sequential(nn.Linear(n_flatten, features_dim), nn.ReLU())
+
+    def forward(self, observations: th.Tensor) -> th.Tensor:
+        return self.linear(self.cnn(observations))
+
+
+policy_kwargs = dict(
+    features_extractor_class=CustomMPL,
+    features_extractor_kwargs=dict(features_dim=128, net_arch=[32, 32]),
+)
+
+# policy_kwargs = dict(net_arch=dict(pi=[5, 5], qf=[10, 10]))
+# policy_kwargs = dict( activation_fn=th.nn.LeakyReLU, net_arch=[32, 32])
+model = DDPG('MlpPolicy', env, verbose=1, tensorboard_log=f'{timestamp}/', policy_kwargs=policy_kwargs)
+checkpoint_on_event = CheckpointCallback(save_freq=10000, save_path=f'{timestamp}/checkpoints/')
 record_env = RecordEnvCallback()
-plot_callback = EveryNTimesteps(n_steps=50000, callback=record_env)
-model.learn(total_timesteps=500000, callback=[checkpoint_on_event, plot_callback])
+plot_callback = EveryNTimesteps(n_steps=10000, callback=record_env)
+model.learn(total_timesteps=50000, callback=[checkpoint_on_event, plot_callback])
 
-model.save('ddpg_CC')
+model.save('ddpg_CC2')
 
 del model  # remove to demonstrate saving and loading