random paste and image sampling

tensorpack/dataflow/common.py

@@ -115,7 +115,6 @@ class BatchDataByShape(BatchData):
     def get_data(self):
         for dp in self.ds.get_data():
             shp = dp[self.idx].shape
-            print(shp, len(self.holder))
             holder = self.holder[shp]
             holder.append(dp)
             if len(holder) == self.batch_size:

tensorpack/dataflow/dataset/mnist.py

@@ -15,8 +15,6 @@ from ..base import RNGDataFlow
 
 __all__ = ['Mnist']
 
-""" This file is mostly copied from tensorflow example """
-
 SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
 
 def maybe_download(filename, work_directory):

tensorpack/dataflow/imgaug/noise.py

@@ -7,7 +7,7 @@ from .base import ImageAugmentor
 import numpy as np
 import cv2
 
-__all__ = ['JpegNoise', 'GaussianNoise']
+__all__ = ['JpegNoise', 'GaussianNoise', 'SaltPepperNoise']
 
 class JpegNoise(ImageAugmentor):
     def __init__(self, quality_range=(40, 100)):

tensorpack/dataflow/imgaug/paste.py

@@ -8,7 +8,8 @@ from .base import ImageAugmentor
 from abc import abstractmethod
 import numpy as np
 
-__all__ = [ 'CenterPaste', 'BackgroundFiller', 'ConstantBackgroundFiller']
+__all__ = ['CenterPaste', 'BackgroundFiller', 'ConstantBackgroundFiller',
+        'RandomPaste']
 
 
 class BackgroundFiller(object):
@@ -36,7 +37,7 @@ class ConstantBackgroundFiller(BackgroundFiller):
         self.value = value
 
     def _fill(self, background_shape, img):
-        assert img.ndim in [3, 1]
+        assert img.ndim in [3, 2]
         if img.ndim == 3:
             return_shape = background_shape + (3,)
         else:
@@ -63,12 +64,30 @@ class CenterPaste(ImageAugmentor):
 
         background = self.background_filler.fill(
             self.background_shape, img)
-        h0 = int((self.background_shape[0] - img_shape[0]) * 0.5)
-        w0 = int((self.background_shape[1] - img_shape[1]) * 0.5)
-        background[h0:h0+img_shape[0], w0:w0+img_shape[1]] = img
-        img = background
-        return img
+        y0 = int((self.background_shape[0] - img_shape[0]) * 0.5)
+        x0 = int((self.background_shape[1] - img_shape[1]) * 0.5)
+        background[y0:y0+img_shape[0], x0:x0+img_shape[1]] = img
+        return background
 
     def _fprop_coord(self, coord, param):
         raise NotImplementedError()
 
+class RandomPaste(CenterPaste):
+    """
+    Randomly paste the image onto a background convas
+    """
+    def _get_augment_params(self, img):
+        img_shape = img.shape[:2]
+        assert self.background_shape[0] > img_shape[0] and self.background_shape[1] > img_shape[1]
+
+        y0 = self._rand_range(self.background_shape[0] - img_shape[0])
+        x0 = self._rand_range(self.background_shape[1] - img_shape[1])
+        return int(x0), int(y0)
+
+    def _augment(self, img, loc):
+        x0, y0 = loc
+        img_shape = img.shape[:2]
+        background = self.background_filler.fill(
+            self.background_shape, img)
+        background[y0:y0+img_shape[0], x0:x0+img_shape[1]] = img
+        return background

tensorpack/models/image_sample.py

@@ -18,11 +18,11 @@ def sample(img, coords):
     :param coords: bxh2xw2x2 (y, x) integer
     :return: bxh2xw2xc image
     """
+    coords = tf.cast(coords, tf.int32)
     shape = img.get_shape().as_list()[1:]
     shape2 = coords.get_shape().as_list()[1:3]
-    max_coor = tf.constant([shape[0] - 1, shape[1] - 1])
-    coords = tf.minimum(coords, max_coor)
-    coords = tf.maximum(coords, tf.constant(0))
+    max_coor = tf.constant([shape[0] - 1, shape[1] - 1], dtype=tf.int32)
+    coords = tf.clip_by_value(coords, 0, max_coor)
 
     w = shape[1]
     coords = tf.reshape(coords, [-1, 2])
@@ -46,8 +46,8 @@ def ImageSample(inputs):
     It mimics the same behavior described in:
     `Spatial Transformer Networks <http://arxiv.org/abs/1506.02025>`_.
 
-    :param input: [template, mapping]. template of shape NHWC. mapping of
-        shape NHW2, where each pair of the last dimension is a (y, x) real-value
+    :param input: [template, mapping]. template of shape NHWC.
+        mapping of shape NHW2, where each pair of the last dimension is a (y, x) real-value
         coordinate.
     :returns: a NHWC output tensor.
     """
@@ -55,13 +55,10 @@ def ImageSample(inputs):
     assert template.get_shape().ndims == 4 and mapping.get_shape().ndims == 4
 
     mapping = tf.maximum(mapping, 0.0)
-    lcoor = tf.cast(mapping, tf.int32)  # floor
+    lcoor = tf.floor(mapping)
     ucoor = lcoor + 1
 
-    # has to cast to int32 and then cast back
-    # tf.floor have gradient 1 w.r.t input
-    # TODO bug fixed in #951
-    diff = mapping - tf.cast(lcoor, tf.float32)
+    diff = mapping - lcoor
     neg_diff = 1.0 - diff   #bxh2xw2x2
 
     lcoory, lcoorx = tf.split(3, 2, lcoor)