randomcroprandomshape

examples/cifar10_convnet.py

@@ -38,9 +38,6 @@ class Model(ModelDesc):
         keep_prob = tf.constant(0.5 if is_training else 1.0)
 
         if is_training:
-            image, label = tf.train.shuffle_batch(
-                [image, label], BATCH_SIZE, CAPACITY, MIN_AFTER_DEQUEUE,
-                num_threads=6, enqueue_many=True)
             tf.image_summary("train_image", image, 10)
 
         image = image / 4.0     # just to make range smaller

examples/cifar10_resnet.py

@@ -81,7 +81,6 @@ class Model(ModelDesc):
                 l = c2 + l
                 return l
 
-
         l = conv('conv0', image, 16, 1)
         l = BatchNorm('bn0', l, is_training)
         l = tf.nn.relu(l)

tensorpack/dataflow/imgaug/_test.py

@@ -5,14 +5,16 @@
 
 import sys
 import cv2
-from . import AugmentorList, Flip, GaussianDeform, Image
+from . import AugmentorList, Image
+from .crop import *
 
 
 anchors = [(0.2, 0.2), (0.8, 0.8), (0.5, 0.5), (0.2, 0.5)]
 augmentors = AugmentorList([
     #Contrast((0.2,1.8)),
     #Flip(horiz=True),
-    GaussianDeform(anchors, (360,480), 1, randrange=10)
+    #GaussianDeform(anchors, (360,480), 1, randrange=10)
+    RandomCropRandomShape(0.3)
 ])
 
 while True:

tensorpack/dataflow/imgaug/crop.py

@@ -3,12 +3,12 @@
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 from .base import ImageAugmentor
+from ...utils.rect import Rect
 
+from six.moves import range
 import numpy as np
-from abc import abstractmethod
 
-__all__ = ['RandomCrop', 'CenterCrop', 'FixedCrop', 'CenterPaste',
-           'ConstantBackgroundFiller']
+__all__ = ['RandomCrop', 'CenterCrop', 'FixedCrop', 'RandomCropRandomShape']
 
 class RandomCrop(ImageAugmentor):
     """ Randomly crop the image into a smaller one """
@@ -44,79 +44,76 @@ class CenterCrop(ImageAugmentor):
 
 class FixedCrop(ImageAugmentor):
     """ Crop a rectangle at a given location"""
-    def __init__(self, rangex, rangey):
+    def __init__(self, rect):
         """
         Two arguments defined the range in both axes to crop, min inclued, max excluded.
 
-        :param rangex: like (xmin, xmax).
-        :param rangey: like (ymin, ymax).
+        :param rect: a `Rect` instance
         """
         self._init(locals())
 
     def _augment(self, img):
         orig_shape = img.arr.shape
-        img.arr = img.arr[self.rangey[0]:self.rangey[1],
-                          self.rangex[0]:self.rangex[1]]
+        img.arr = img.arr[self.rect.y0: self.rect.y1+1,
+                          self.rect.x0: self.rect.x0+1]
         if img.coords:
             raise NotImplementedError()
 
-class BackgroundFiller(object):
-    """ Base class for all BackgroundFiller"""
-    def fill(self, background_shape, img):
+def perturb_BB(image_shape, bb, max_pertub_pixel,
+        rng=None, max_aspect_ratio_diff=0.3,
+        max_try=100):
     """
-        Return a proper background image of background_shape, given img
-
-        :param background_shape: a shape of [h, w]
-        :param img: an image
-        :returns: a background image
-        """
-        return self._fill(background_shape, img)
-
-    @abstractmethod
-    def _fill(self, background_shape, img):
-        pass
-
-class ConstantBackgroundFiller(BackgroundFiller):
-    """ Fill the background by a constant """
-    def __init__(self, value):
-        """
-        :param value: the value to fill the background.
+    Perturb a bounding box.
+    :param image_shape: [h, w]
+    :param bb: a `Rect` instance
+    :param max_pertub_pixel: pertubation on each coordinate
+    :param max_aspect_ratio_diff: result can't have an aspect ratio too different from the original
+    :param max_try: if cannot find a valid bounding box, return the original
+    :returns: new bounding box
     """
-        self.value = value
-
-    def _fill(self, background_shape, img):
-        assert img.ndim in [3, 1]
-        if img.ndim == 3:
-            return_shape = background_shape + (3,)
-        else:
-            return_shape = background_shape
-        return np.zeros(return_shape) + self.value
-
-class CenterPaste(ImageAugmentor):
+    orig_ratio = bb.h * 1.0 / bb.w
+    if rng is None:
+        rng = np.random.RandomState()
+    for _ in range(max_try):
+        p = rng.randint(-max_pertub_pixel, max_pertub_pixel, [4])
+        newbb = bb.copy()
+        newbb.x += p[0]
+        newbb.y += p[1]
+        newx1 = bb.x1 + p[2]
+        newy1 = bb.y1 + p[3]
+        newbb.w = newx1 - newbb.x
+        newbb.h = newy1 - newbb.y
+        if not newbb.validate(image_shape):
+            continue
+        new_ratio = newbb.h * 1.0 / newbb.w
+        diff = abs(new_ratio - orig_ratio)
+        if diff / orig_ratio > max_aspect_ratio_diff:
+            continue
+        return newbb
+    return bb
+
+
+class RandomCropRandomShape(ImageAugmentor):
     """
-    Paste the image onto the center of a background canvas.
+    Crop a box around a bounding box
     """
-    def __init__(self, background_shape, background_filler=None):
+    def __init__(self, perturb_ratio, max_aspect_ratio_diff=0.3):
         """
-        :param background_shape: shape of the background canvas.
-        :param background_filler: a `BackgroundFiller` instance. Default to zero-filler.
+        :param perturb_ratio: perturb distance will be in [0, perturb_ratio * sqrt(w * h)]
+        :param max_aspect_ratio_diff: keep aspect ratio within the range
         """
-        if background_filler is None:
-            background_filler = ConstantBackgroundFiller(0)
-
         self._init(locals())
 
     def _augment(self, img):
-        img_shape = img.arr.shape[:2]
-        assert self.background_shape[0] > img_shape[0] and self.background_shape[1] > img_shape[1]
-
-        background = self.background_filler.fill(
-            self.background_shape, img.arr)
-        h0 = (self.background_shape[0] - img_shape[0]) * 0.5
-        w0 = (self.background_shape[1] - img_shape[1]) * 0.5
-        background[h0:h0+img_shape[0], w0:w0+img_shape[1]] = img.arr
-        img.arr = background
+        shape = img.arr.shape[:2]
+        box = Rect(0, 0, shape[1] - 1, shape[0] - 1)
+        dist = self.perturb_ratio * np.sqrt(shape[0]*shape[1])
+        newbox = perturb_BB(shape, box, dist,
+                self.rng, self.max_aspect_ratio_diff)
+
+        img.arr = newbox.roi(img.arr)
         if img.coords:
             raise NotImplementedError()
 
-
+if __name__ == '__main__':
+    print(perturb_BB([100, 100], Rect(3, 3, 50, 50), 50))

tensorpack/dataflow/imgaug/paste.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: paste.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+from .base import ImageAugmentor
+
+from abc import abstractmethod
+import numpy as np
+
+__all__ = [ 'CenterPaste', 'ConstantBackgroundFiller']
+
+
+class BackgroundFiller(object):
+    """ Base class for all BackgroundFiller"""
+    def fill(self, background_shape, img):
+        """
+        Return a proper background image of background_shape, given img
+
+        :param background_shape: a shape of [h, w]
+        :param img: an image
+        :returns: a background image
+        """
+        return self._fill(background_shape, img)
+
+    @abstractmethod
+    def _fill(self, background_shape, img):
+        pass
+
+class ConstantBackgroundFiller(BackgroundFiller):
+    """ Fill the background by a constant """
+    def __init__(self, value):
+        """
+        :param value: the value to fill the background.
+        """
+        self.value = value
+
+    def _fill(self, background_shape, img):
+        assert img.ndim in [3, 1]
+        if img.ndim == 3:
+            return_shape = background_shape + (3,)
+        else:
+            return_shape = background_shape
+        return np.zeros(return_shape) + self.value
+
+class CenterPaste(ImageAugmentor):
+    """
+    Paste the image onto the center of a background canvas.
+    """
+    def __init__(self, background_shape, background_filler=None):
+        """
+        :param background_shape: shape of the background canvas.
+        :param background_filler: a `BackgroundFiller` instance. Default to zero-filler.
+        """
+        if background_filler is None:
+            background_filler = ConstantBackgroundFiller(0)
+
+        self._init(locals())
+
+    def _augment(self, img):
+        img_shape = img.arr.shape[:2]
+        assert self.background_shape[0] > img_shape[0] and self.background_shape[1] > img_shape[1]
+
+        background = self.background_filler.fill(
+            self.background_shape, img.arr)
+        h0 = (self.background_shape[0] - img_shape[0]) * 0.5
+        w0 = (self.background_shape[1] - img_shape[1]) * 0.5
+        background[h0:h0+img_shape[0], w0:w0+img_shape[1]] = img.arr
+        img.arr = background
+        if img.coords:
+            raise NotImplementedError()
+
+

tensorpack/utils/rect.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: rect.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+
+class Rect(object):
+    """
+    A Rectangle.
+    Note that x1 = x+w, not x+w-1 or something
+    """
+    __slots__ = ['x', 'y', 'w', 'h']
+
+    def __init__(self, x=0, y=0, w=0, h=0):
+        self.x = x
+        self.y = y
+        self.w = w
+        self.h = h
+        assert min(self.x, self.y, self.w, self.h) >= 0
+
+    @property
+    def x0(self):
+        return self.x
+
+    @property
+    def y0(self):
+        return self.y
+
+    @property
+    def x1(self):
+        return self.x + self.w
+
+    @property
+    def y1(self):
+        return self.y + self.h
+
+    def copy(self):
+        new = type(self)()
+        for i in self.__slots__:
+            setattr(new, i, getattr(self, i))
+        return new
+
+    def __str__(self):
+        return 'Rect(x={}, y={}, w={}, h={})'.format(self.x, self.y, self.w, self.h)
+
+    def area(self):
+        return self.w * self.h
+
+    def validate(self, shape=None):
+        """
+        Is a valid bounding box within this shape
+        :param shape: [h, w]
+        :returns: boolean
+        """
+        if min(self.x, self.y) < 0:
+            return False
+        if min(self.w, self.h) <= 0:
+            return False
+        if shape is None:
+            return True
+        if self.x1 > shape[1] - 1:
+            return False
+        if self.y1 > shape[0] - 1:
+            return False
+        return True
+
+    def roi(self, img):
+        assert self.validate(img.shape[:2])
+        return img[self.y0:self.y1+1, self.x0:self.x1+1]
+
+    __repr__ = __str__

tensorpack/utils/utils.py

@@ -4,6 +4,7 @@
 
 import os, sys
 from contextlib import contextmanager
+from datetime import datetime
 import time
 import collections
 import numpy as np