transpose wrapper

rl_portfolio_management/wrappers/transpose_history.py

@@ -1,16 +1,31 @@
-import gym.wrappers
+import gym.spaces
 import numpy as np
 
 
 class TransposeHistory(gym.Wrapper):
     """Transpose history."""
 
+    def __init__(self, env, axes=(2, 1, 0)):
+        super().__init__(env)
+        self.axes = axes
+
+        hist_space = self.observation_space.spaces["history"]
+        hist_shape = hist_space.shape
+        self.observation_space = gym.spaces.Dict({
+            'history': gym.spaces.Box(
+                hist_space.low.min(),
+                hist_space.high.max(),
+                (hist_shape[axes[0]], hist_shape[axes[1]], hist_shape[axes[2]])
+            ),
+            'weights': self.observation_space.spaces["weights"]
+        })
+
     def step(self, action):
         state, reward, done, info = self.env.step(action)
-        state["history"] = np.transpose(state["history"], (2, 1, 0))
+        state["history"] = np.transpose(state["history"], self.axes)
         return state, reward, done, info
 
     def reset(self):
         state = self.env.reset()
-        state["history"] = np.transpose(state["history"], (2, 1, 0))
+        state["history"] = np.transpose(state["history"], self.axes)
         return state

test/test_wrappers.py

@@ -33,6 +33,7 @@ def test_transpose():
     env = gym.make("CryptoPortfolioEIIE-v0")
     env = TransposeHistory(env)
     obs = env.reset()
+    env.observation_space.contains(obs)
     assert obs["history"].shape == transposed_shape
     # should be no problem with actions that don't sum to one
     action = env.action_space.sample()