update imgaug

tensorpack/dataflow/common.py

@@ -100,7 +100,6 @@ class MapData(DataFlow):
 
     def get_data(self):
         for dp in self.ds.get_data():
-            d = list(dp)
             dp[self.index] = self.func(dp[self.index])
             yield dp
 

tensorpack/dataflow/imgaug/_test.py

+#!/usr/bin/env python2
+# -*- coding: UTF-8 -*-
+# File: _test.py
+# Author: Yuxin Wu <ppwwyyxx@gmail.com>
+
+import cv2
+from . import *
+
+augmentors = AugmentorList([
+    Contrast((0.2,1.8)),
+    Flip(horiz=True)
+])
+
+img = cv2.imread('cat.jpg')
+img = Image(img)
+augmentors.augment(img)
+cv2.imshow(" ", img.arr)
+cv2.waitKey()

tensorpack/dataflow/imgaug/base.py

@@ -61,6 +61,6 @@ class AugmentorList(ImageAugmentor):
 
     def _augment(self, img):
         assert img.arr.ndim in [2, 3]
-        img.arr = img.arr.astype('float32') / 255.0
+        img.arr = img.arr.astype('float32')
         for aug in self.augs:
             aug.augment(img)

tensorpack/dataflow/imgaug/imgproc.py

@@ -19,7 +19,7 @@ class BrightnessAdd(ImageAugmentor):
     def _augment(self, img):
         v = self._rand_range(-self.delta, self.delta)
         img.arr += v
-        img.arr = np.clip(img.arr, 0, 1)
+        img.arr = np.clip(img.arr, 0, 255)
 
 class Contrast(ImageAugmentor):
     """
@@ -33,7 +33,7 @@ class Contrast(ImageAugmentor):
         r = self._rand_range(*self.factor_range)
         mean = np.mean(arr, axis=(0,1), keepdims=True)
         img.arr = (arr - mean) * r + mean
-        img.arr = np.clip(img.arr, 0, 1)
+        img.arr = np.clip(img.arr, 0, 255)
 
 class PerImageWhitening(ImageAugmentor):
     """
@@ -41,11 +41,16 @@ class PerImageWhitening(ImageAugmentor):
     x = (x - mean) / adjusted_stddev
     where adjusted_stddev = max(stddev, 1.0/sqrt(num_pixels * channels))
     """
-    def __init__(self):
+    def __init__(self, all_channel=True):
+        self.all_channel = all_channel
         pass
 
     def _augment(self, img):
-        mean = np.mean(img.arr, axis=(0,1), keepdims=True)
-        std = np.std(img.arr, axis=(0,1), keepdims=True)
+        if self.all_channel:
+            mean = np.mean(img.arr)
+            std = np.std(img.arr)
+        else:
+            mean = np.mean(img.arr, axis=(0,1), keepdims=True)
+            std = np.std(img.arr, axis=(0,1), keepdims=True)
         std = np.maximum(std, 1.0 / np.sqrt(np.prod(img.arr.shape)))
         img.arr = (img.arr - mean) / std

tensorpack/dataflow/imgaug/noname.py

@@ -28,6 +28,7 @@ class Flip(ImageAugmentor):
         self._init()
 
     def _augment(self, img):
+# TODO XXX prob is wrong for both mode
         if self._rand_range() < self.prob:
             img.arr = cv2.flip(img.arr, self.code)
             if img.coords:

tensorpack/train.py

@@ -161,8 +161,9 @@ def start_train(config):
     summary_grads(grads)
     avg_maintain_op = summary_moving_average(cost_var)
 
-    with tf.control_dependencies([avg_maintain_op]):
-        train_op = config.optimizer.apply_gradients(grads, get_global_step_var())
+    train_op = tf.group(
+        config.optimizer.apply_gradients(grads, get_global_step_var()),
+        avg_maintain_op)
 
     describe_model()
     sess = tf.Session(config=config.session_config)