bug fix in imgaug

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -25,6 +25,9 @@ You'll need tcmalloc to avoid large memory consumption: https://github.com/tenso
 This config, with (W,A,G)=(1,1,4), can reach 3.1~3.2% error after 150 epochs.
 With the GaussianDeform augmentor, it will reach 2.8~2.9%
 (we are not using this augmentor in the paper).
+
+with (W,A,G)=(1,2,4), error is 3.0~3.1%.
+with (W,A,G)=(32,32,32), error is about 2.9%.
 """
 
 BITW = 1

tensorpack/dataflow/imgaug/base.py

@@ -61,7 +61,7 @@ class ImageAugmentor(object):
             low, high = 0, low
         if size == None:
             size = []
-        return low + self.rng.rand(*size) * (high - low)
+        return self.rng.uniform(low, high, size)
 
 class AugmentorList(ImageAugmentor):
     """

tensorpack/dataflow/imgaug/crop.py

@@ -20,8 +20,18 @@ class RandomCrop(ImageAugmentor):
 
     def _get_augment_params(self, img):
         orig_shape = img.shape
-        h0 = self.rng.randint(0, orig_shape[0] - self.crop_shape[0])
-        w0 = self.rng.randint(0, orig_shape[1] - self.crop_shape[1])
+        assert orig_shape[0] >= self.crop_shape[0] \
+                and orig_shape[1] >= self.crop_shape[1], orig_shape
+        diffh = orig_shape[0] - self.crop_shape[0]
+        if diffh == 0:
+            h0 = 0
+        else:
+            h0 = self.rng.randint(diffh)
+        diffw = orig_shape[1] - self.crop_shape[1]
+        if diffw == 0:
+            w0 = 0
+        else:
+            w0 = self.rng.randint(diffw)
         return (h0, w0)
 
     def _augment(self, img, param):

tensorpack/dataflow/imgaug/noise.py

@@ -14,11 +14,11 @@ class JpegNoise(ImageAugmentor):
         self._init(locals())
 
     def _get_augment_params(self, img):
-        return self._rand_range(*self.quality_range)
+        return self.rng.randint(*self.quality_range)
 
     def _augment(self, img, q):
-        return cv2.imdecode(cv2.imencode('.jpg', img,
-            [cv2.IMWRITE_JPEG_QUALITY, q])[1], 1)
+        enc = cv2.imencode('.jpg', img, [cv2.IMWRITE_JPEG_QUALITY, q])[1]
+        return cv2.imdecode(enc, 1)
 
 
 class GaussianNoise(ImageAugmentor):

tensorpack/dataflow/imgaug/noname.py

@@ -3,6 +3,7 @@
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 from .base import ImageAugmentor
+from ...utils import logger
 import numpy as np
 import cv2
 
@@ -68,6 +69,7 @@ class RandomResize(ImageAugmentor):
         self._init(locals())
 
     def _get_augment_params(self, img):
+        cnt = 0
         while True:
             sx = self._rand_range(*self.xrange)
             sy = self._rand_range(*self.yrange)
@@ -78,6 +80,9 @@ class RandomResize(ImageAugmentor):
             diff = abs(newr - oldr) / oldr
             if diff <= self.aspect_ratio_thres:
                 return (destX, destY)
+            cnt += 1
+            if cnt > 50:
+                logger.warn("RandomResize failed to augment an image")
 
     def _augment(self, img, dsize):
         return cv2.resize(img, dsize, interpolation=cv2.INTER_CUBIC)