Match clamping between Pytorch and c2 preprocessing

Summary: title Reviewed By: kittipatv Differential Revision: D13006446 fbshipit-source-id: ad79caab4bfd05475d56d9565f4f891ececd503d
Yk1n0gEa · Nov 9, 2018 · 0e7fe9d · 0e7fe9d
1 parent 377786d
commit 0e7fe9d
Show file tree

Hide file tree

Showing 3 changed files with 22 additions and 11 deletions.
diff --git a/ml/rl/preprocessing/normalization.py b/ml/rl/preprocessing/normalization.py
@@ -34,6 +34,9 @@
 MINIMUM_SAMPLES_TO_IDENTIFY = 20
 DEFAULT_MAX_QUANTILE_SIZE = 20
 DEFAULT_NUM_SAMPLES = 100000
+MAX_FEATURE_VALUE = 6.0
+MIN_FEATURE_VALUE = MAX_FEATURE_VALUE * -1
+EPS = 1e-6
 
 
 class NumpyEncoder(json.JSONEncoder):

diff --git a/ml/rl/preprocessing/preprocessor.py b/ml/rl/preprocessing/preprocessor.py
@@ -9,14 +9,16 @@
 import torch.nn as nn
 from ml.rl import types as rlt
 from ml.rl.preprocessing.identify_types import ENUM, FEATURE_TYPES
-from ml.rl.preprocessing.normalization import MISSING_VALUE, NormalizationParameters
+from ml.rl.preprocessing.normalization import (
+    EPS,
+    MAX_FEATURE_VALUE,
+    MIN_FEATURE_VALUE,
+    MISSING_VALUE,
+    NormalizationParameters,
+)
 from torch.nn import Module, Parameter
 
 
-MAX_FEATURE_VALUE = 6
-MIN_FEATURE_VALUE = MAX_FEATURE_VALUE * -1
-EPS = 1e-6
-
 logger = logging.getLogger(__name__)
 
 
@@ -287,7 +289,7 @@ def _preprocess_CONTINUOUS(
         if not self.clamp:
             return continuous_output
         else:
-            return torch.clamp(continuous_output, -3.0, 3.0)
+            return torch.clamp(continuous_output, MIN_FEATURE_VALUE, MAX_FEATURE_VALUE)
 
     def _create_parameters_BOXCOX(
         self, begin_index: int, norm_params: List[NormalizationParameters]

diff --git a/ml/rl/preprocessing/preprocessor_net.py b/ml/rl/preprocessing/preprocessor_net.py
@@ -10,7 +10,13 @@
 from ml.rl.caffe_utils import C2
 from ml.rl.preprocessing import identify_types
 from ml.rl.preprocessing.identify_types import FEATURE_TYPES
-from ml.rl.preprocessing.normalization import MISSING_VALUE, NormalizationParameters
+from ml.rl.preprocessing.normalization import (
+    EPS,
+    MAX_FEATURE_VALUE,
+    MIN_FEATURE_VALUE,
+    MISSING_VALUE,
+    NormalizationParameters,
+)
 
 
 logger = logging.getLogger(__name__)
@@ -242,7 +248,7 @@ def preprocess_blob(self, blob, normalization_parameters):
             blob = C2.Sub(blob, means_blob, broadcast=1, axis=0)
             blob = C2.Div(blob, stddevs_blob, broadcast=1, axis=0)
             if self.clip_anomalies:
-                blob = C2.Clip(blob, min=-6.0, max=6.0)
+                blob = C2.Clip(blob, min=MIN_FEATURE_VALUE, max=MAX_FEATURE_VALUE)
         elif feature_type == identify_types.CONTINUOUS_ACTION:
             serving_min_value = np.array(
                 [norm.min_value for norm in normalization_parameters], dtype=np.float32
@@ -252,11 +258,11 @@ def preprocess_blob(self, blob, normalization_parameters):
             )
 
             training_min_value = (
-                np.ones(len(normalization_parameters), dtype=np.float32) * -1 + 1e-6
+                np.ones(len(normalization_parameters), dtype=np.float32) * -1 + EPS
             )
 
             scaling_factor = (
-                (np.ones(len(normalization_parameters), dtype=np.float32) - 1e-6)
+                (np.ones(len(normalization_parameters), dtype=np.float32) - EPS)
                 * 2
                 / (serving_max_value - serving_min_value)
             )
@@ -275,7 +281,7 @@ def preprocess_blob(self, blob, normalization_parameters):
             blob = C2.Mul(blob, scaling_factor_blob, broadcast=1, axis=1)
             blob = C2.Add(blob, training_min_blob, broadcast=1, axis=1)
             if self.clip_anomalies:
-                blob = C2.Clip(blob, min=-1 + 1e-6, max=1 - 1e-6)
+                blob = C2.Clip(blob, min=-1 + EPS, max=1 - EPS)
         else:
             raise NotImplementedError("Invalid feature type: {}".format(feature_type))