Performance

lib/data/features/transform.py

@@ -52,3 +52,42 @@ def difference(iterable: Iterable, inplace: bool = True, columns: List[str] = No
 
 def log_and_difference(iterable: Iterable, inplace: bool = True, columns: List[str] = None):
     return transform(iterable, inplace, columns, lambda t_iterable: np.log(t_iterable) - np.log(t_iterable).shift(1))
+
+
+class FastTransform:
+
+    MIN="min"
+    MAX="max"
+
+    def __init__(self):
+        self.cached_min_max = dict()
+
+    def max_min_normalize(self, iterable: Iterable, inplace: bool = True, columns: List[str] = None):
+        columns = iterable.columns if columns is None else columns
+
+        if iterable.empty:
+            return iterable
+
+        for column in columns:
+            cached_min = self.cached_min_max[FastTransform.MIN + column] if FastTransform.MIN + column in self.cached_min_max else np.min(iterable[column])
+            cached_max = self.cached_min_max[FastTransform.MAX + column] if FastTransform.MAX + column in self.cached_min_max else np.max(iterable[column])
+
+            new_min = min(cached_min, iterable.iloc[-1][column])
+            new_max = max(cached_max, iterable.iloc[-1][column])
+
+            iterable.iloc[-1][column] = (iterable.iloc[-1][column] - new_min) / (new_max - new_min)
+
+            self._update_cache(column, new_min, new_max)
+
+        return iterable
+
+    def _update_cache(self,column, min, max):
+        #don't want to save nan's as it will force recalculation on the next iteration
+        if not np.isnan(min):
+            self.cached_min_max[FastTransform.MIN + column] = min
+
+        if not np.isnan(max):
+            self.cached_min_max[FastTransform.MAX + column] = max
+
+    def _get_cached(self, column):
+        return self.cached_min_max[FastTransform.MIN + column], self.cached_min_max[FastTransform.MAX + column]

lib/env/TradingEnv.py

@@ -9,7 +9,7 @@
 from lib.env.render import TradingChart
 from lib.env.reward import BaseRewardStrategy, IncrementalProfit
 from lib.data.providers import BaseDataProvider
-from lib.data.features.transform import max_min_normalize, mean_normalize, log_and_difference, difference
+from lib.data.features.transform import max_min_normalize, log_and_difference, FastTransform, difference
 from lib.util.logger import init_logger
 
 
@@ -58,6 +58,10 @@ def __init__(self,
         self.observation_space = spaces.Box(low=0, high=1, shape=self.obs_shape, dtype=np.float16)
 
         self.observations = pd.DataFrame(None, columns=self.data_provider.columns)
+        self.fast_transform_observations = FastTransform()
+        self.fast_transform_account_history = FastTransform()
+
+        self.fast_log_observation = self.observations
 
     def _current_price(self, ohlcv_key: str = 'Close'):
         return float(self.current_ohlcv[ohlcv_key])
@@ -97,6 +101,38 @@ def _make_trade(self, action: int, current_price: float):
 
         return asset_bought, asset_sold, cost_of_asset, revenue_from_sold
 
+    def _next_observation(self):
+        self.current_ohlcv = self.data_provider.next_ohlcv()
+        self.observations = self.observations.append(self.current_ohlcv, ignore_index=True)
+
+        if self.stationarize_obs:
+            log_observation = log_and_difference(self.observations.tail(2), inplace=False)
+            self.fast_log_observation = self.fast_log_observation.append(log_observation.tail(1), ignore_index=True)
+            observations = self.fast_log_observation
+        else:
+            observations = self.observations
+
+        if self.normalize_obs:
+            observations = self.fast_transform_observations.max_min_normalize(observations)
+
+        obs = observations.values[-1]
+
+        if self.stationarize_obs:
+            scaled_history = log_and_difference(self.account_history, inplace=False)
+        else:
+            scaled_history = self.account_history
+
+        if self.normalize_obs:
+            scaled_history = self.fast_transform_account_history.max_min_normalize(self.account_history)
+
+        obs = np.insert(obs, len(obs), scaled_history.values[-1], axis=0)
+
+        obs = np.reshape(obs.astype('float16'), self.obs_shape)
+        obs[np.bitwise_not(np.isfinite(obs))] = 0
+
+        return obs
+
+
     def _take_action(self, action: int):
         current_price = self._current_price()
 
@@ -143,35 +179,6 @@ def _reward(self):
 
         return float(rewards[-1])
 
-    def _next_observation(self):
-        self.current_ohlcv = self.data_provider.next_ohlcv()
-        self.observations = self.observations.append(self.current_ohlcv, ignore_index=True)
-
-        if self.stationarize_obs:
-            observations = log_and_difference(self.observations, inplace=False)
-        else:
-            observations = self.observations
-
-        if self.normalize_obs:
-            observations = max_min_normalize(observations)
-
-        obs = observations.values[-1]
-
-        if self.stationarize_obs:
-            scaled_history = log_and_difference(self.account_history, inplace=False)
-        else:
-            scaled_history = self.account_history
-
-        if self.normalize_obs:
-            scaled_history = max_min_normalize(scaled_history, inplace=False)
-
-        obs = np.insert(obs, len(obs), scaled_history.values[-1], axis=0)
-
-        obs = np.reshape(obs.astype('float16'), self.obs_shape)
-        obs[np.bitwise_not(np.isfinite(obs))] = 0
-
-        return obs
-
     def reset(self):
         self.data_provider.reset_ohlcv_index()