imgaug

tensorpack/dataflow/common.py

@@ -324,10 +324,10 @@ class LocallyShuffleData(ProxyDataFlow, RNGDataFlow):
             self.rng.shuffle(self.q)
             for _ in range(self.q.maxlen):
                 yield self.q.popleft()
+                self.q.append(next(self.ds_itr))
                 cnt += 1
                 if cnt == self.size():
                     return
-                self.q.append(next(self.ds_itr))
 
 
 def SelectComponent(ds, idxs):

tensorpack/dataflow/dataset/ilsvrc.py

@@ -90,7 +90,7 @@ class ILSVRCMeta(object):
         obj = caffepb.BlobProto()
 
         mean_file = os.path.join(self.dir, 'imagenet_mean.binaryproto')
-        with open(mean_file) as f:
+        with open(mean_file, 'rb') as f:
             obj.ParseFromString(f.read())
         arr = np.array(obj.data).reshape((3, 256, 256)).astype('float32')
         arr = np.transpose(arr, [1,2,0])
@@ -105,6 +105,7 @@ class ILSVRC12(DataFlow):
         """
         assert name in ['train', 'test', 'val']
         self.full_dir = os.path.join(dir, name)
+        assert os.path.isdir(self.full_dir), self.full_dir
         self.shuffle = shuffle
         self.meta = ILSVRCMeta(meta_dir)
         self.imglist = self.meta.get_image_list(name)

tensorpack/dataflow/imgaug/meta.py

@@ -6,7 +6,11 @@
 
 from .base import ImageAugmentor
 
-__all__ = ['RandomChooseAug', 'MapImage']
+__all__ = ['RandomChooseAug', 'MapImage', 'Identity']
+
+class Identity(ImageAugmentor):
+    def _augment(self, img, _):
+        return img
 
 class RandomChooseAug(ImageAugmentor):
     def __init__(self, aug_lists):

tensorpack/dataflow/imgaug/noise.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: noise.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+from .base import ImageAugmentor
+import numpy as np
+import cv2
+
+__all__ = ['JpegNoise', 'GaussianNoise']
+
+class JpegNoise(ImageAugmentor):
+    def __init__(self, quality_range=(40, 100)):
+        self._init(locals())
+
+    def _get_augment_params(self, img):
+        return self._rand_range(*self.quality_range)
+
+    def _augment(self, img, q):
+        return cv2.imdecode(cv2.imencode('.jpg', img,
+            [cv2.IMWRITE_JPEG_QUALITY, q])[1], 1)
+
+
+class GaussianNoise(ImageAugmentor):
+    def __init__(self, scale=10, clip=True):
+        self._init(locals())
+
+    def _get_augment_params(self, img):
+        return self.rng.randn(img.shape)
+
+    def _augment(self, img, noise):
+        ret = img + noise
+        if self.clip:
+            ret = np.clip(ret, 0, 255)
+        return ret

tensorpack/dataflow/imgaug/noname.py

@@ -6,7 +6,7 @@ from .base import ImageAugmentor
 import numpy as np
 import cv2
 
-__all__ = ['Flip', 'Resize', 'RandomResize', 'JpegNoise']
+__all__ = ['Flip', 'Resize', 'RandomResize']
 
 class Flip(ImageAugmentor):
     """
@@ -58,7 +58,7 @@ class Resize(ImageAugmentor):
 
 class RandomResize(ImageAugmentor):
     """ randomly rescale w and h of the image"""
-    def __init__(self, xrange, yrange, minimum=None, aspect_ratio_thres=0.2):
+    def __init__(self, xrange, yrange, minimum=(0,0), aspect_ratio_thres=0.15):
         """
         :param xrange: (min, max) scaling ratio
         :param yrange: (min, max) scaling ratio
@@ -82,14 +82,3 @@ class RandomResize(ImageAugmentor):
     def _augment(self, img, dsize):
         return cv2.resize(img, dsize, interpolation=cv2.INTER_CUBIC)
 
-
-class JpegNoise(ImageAugmentor):
-    def __init__(self, quality_range=(40, 100)):
-        self._init(locals())
-
-    def _get_augment_params(self, img):
-        return self._rand_range(*self.quality_range)
-
-    def _augment(self, img, q):
-        return cv2.imdecode(cv2.imencode('.jpg', img,
-            [cv2.IMWRITE_JPEG_QUALITY, q])[1], 1)