discretizer

examples/Atari2600/README.md

@@ -11,4 +11,5 @@ To run:
 ./DQN.py --rom breakout.rom --gpu 0
 ```
 
+Can reproduce the claimed performance, on games I've tested with (curves will be available soon).
 A demo trained with Double-DQN on breakout is available at [youtube](https://youtu.be/o21mddZtE5Y).

examples/ResNet/README.md

 
 ## ResNet
 
-Implement the paper "Deep Residual Learning for Image Recognition", [http://arxiv.org/abs/1512.03385](http://arxiv.org/abs/1512.03385)
+Reproduce the results in paper "Deep Residual Learning for Image Recognition", [http://arxiv.org/abs/1512.03385](http://arxiv.org/abs/1512.03385)
 with the variants proposed in "Identity Mappings in Deep Residual Networks", [https://arxiv.org/abs/1603.05027](https://arxiv.org/abs/1603.05027).
 
 The train error shown here is a moving average of the error rate of each batch in training.

tensorpack/utils/discretize.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: discretize.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+from . import logger, memoized
+from abc import abstractmethod, ABCMeta
+import numpy as np
+from six.moves import range
+
+__all__ = ['UniformDiscretizer1D']
+
+@memoized
+def log_once(s):
+    logger.warn(s)
+
+# just placeholder
+class Discretizer(object):
+    __metaclass__ = ABCMeta
+
+    @abstractmethod
+    def get_nr_bin(self):
+        pass
+
+    @abstractmethod
+    def get_bin(self, v):
+        pass
+
+class Discretizer1D(Discretizer):
+    pass
+
+class UniformDiscretizer1D(Discretizer1D):
+    def __init__(self, minv, maxv, spacing):
+        self.minv = float(minv)
+        self.maxv = float(maxv)
+        self.spacing = float(spacing)
+        self.nr_bin = np.ceil((self.maxv - self.minv) / self.spacing)
+
+    def get_nr_bin(self):
+        return self.nr_bin
+
+    def get_bin(self, v):
+        if v < self.minv:
+            log_once("UniformDiscretizer1D: value smaller than min!")
+            return 0
+        if v > self.maxv:
+            log_once("UniformDiscretizer1D: value larger than max!")
+            return self.nr_bin - 1
+        return int(np.clip(
+                (v - self.minv) / self.spacing,
+                0, self.nr_bin - 1))
+
+    def get_distribution(self, v, smooth_factor=0.05, smooth_radius=2):
+        b = self.get_bin(v)
+        ret = np.zeros((self.nr_bin, ), dtype='float32')
+        ret[b] = 1.0
+        if v >= self.maxv or v <= self.minv:
+            return ret
+        try:
+            for k in range(1, smooth_radius+1):
+                ret[b+k] = smooth_factor ** k
+        except IndexError:
+            pass
+        for k in range(1, min(smooth_radius+1, b+1)):
+            ret[b-k] = smooth_factor ** k
+        ret /= ret.sum()
+        return ret
+
+
+if __name__ == '__main__':
+    u = UniformDiscretizer1D(-10, 10, 0.12)
+    import IPython as IP;
+    IP.embed(config=IP.terminal.ipapp.load_default_config())
+