upsample layer

scripts/dump_train_config.py

@@ -11,6 +11,7 @@ import tqdm
 import os
 from tensorpack.utils import logger
 from tensorpack.utils.fs import mkdir_p
+from tensorpack.dataflow import *
 
 
 parser = argparse.ArgumentParser()
@@ -44,13 +45,13 @@ if args.output:
 
 NR_DP_TEST = args.number
 logger.info("Testing dataflow speed:")
+ds = RepeatedData(config.dataset, -1)
 with tqdm.tqdm(total=NR_DP_TEST, leave=True, unit='data points') as pbar:
-    for idx, dp in enumerate(config.dataset.get_data()):
+    for idx, dp in enumerate(ds.get_data()):
         del dp
         if idx > NR_DP_TEST:
             break
         pbar.update()
-    from IPython import embed; embed()
 
 
 

tensorpack/dataflow/common.py

@@ -10,8 +10,8 @@ from .base import DataFlow, ProxyDataFlow
 from ..utils import *
 
 __all__ = ['BatchData', 'FixedSizeData', 'FakeData', 'MapData',
-           'MapDataComponent', 'RandomChooseData', 'RandomMixData',
-           'JoinData', 'ConcatData', 'SelectComponent']
+           'RepeatedData', 'MapDataComponent', 'RandomChooseData',
+           'RandomMixData', 'JoinData', 'ConcatData', 'SelectComponent']
 
 class BatchData(ProxyDataFlow):
     def __init__(self, ds, batch_size, remainder=False):

tensorpack/dataflow/dataset/bsds500.py

@@ -24,7 +24,7 @@ class BSDS500(DataFlow):
 
     Produce (image, label) pair, where image has shape (321, 481, 3) and
     ranges in [0,255]. Label is binary and has shape (321, 481).
-    Those pixels annotated as boundaries by >= 3 out of 5 annotators are
+    Those pixels annotated as boundaries by >= 3 annotators are
     considered positive examples. This is used in `Holistically-Nested Edge Detection
     <http://arxiv.org/abs/1504.06375>`_.
     """
@@ -71,7 +71,8 @@ class BSDS500(DataFlow):
             imgid = os.path.basename(f).split('.')[0]
             gt_file = os.path.join(gt_dir, imgid)
             gt = loadmat(gt_file)['groundTruth'][0]
-            gt = sum(gt[k]['Boundaries'][0][0] for k in range(5))
+            n_annot = gt.shape[0]
+            gt = sum(gt[k]['Boundaries'][0][0] for k in range(n_annot))
             gt[gt < 3] = 0
             gt[gt != 0] = 1
             if gt.shape[0] > gt.shape[1]:

tensorpack/dataflow/format.py

@@ -11,7 +11,7 @@ from six.moves import range
 try:
     import h5py
 except ImportError:
-    logger.error("Error in 'import h5py'. HDF5Data won't be available.")
+    logger.warn("Error in 'import h5py'. HDF5Data won't be available.")
     __all__ = []
 else:
     __all__ = ['HDF5Data']

tensorpack/dataflow/imgaug/base.py

@@ -4,6 +4,7 @@
 
 from abc import abstractmethod, ABCMeta
 from ...utils import get_rng
+from six.moves import zip
 
 __all__ = ['ImageAugmentor', 'AugmentorList']
 

tensorpack/dataflow/prefetch.py

@@ -17,7 +17,7 @@ from ..utils import logger
 try:
     import zmq
 except ImportError:
-    logger.error("Error in 'import zmq'. PrefetchDataZMQ won't be available.")
+    logger.warn("Error in 'import zmq'. PrefetchDataZMQ won't be available.")
     __all__ = ['PrefetchData']
 else:
     __all__ = ['PrefetchData', 'PrefetchDataZMQ']

tensorpack/models/pool.py

@@ -69,7 +69,7 @@ def FixedUnPooling(x, shape, unpool_mat=None):
 
     :param input: NHWC tensor
     :param shape: int or [h, w]
-    :param unpool_mat: a tf matrix with size=shape. If None, will use a mat
+    :param unpool_mat: a tf/np matrix with size=shape. If None, will use a mat
         with 1 at top-left corner.
     :returns: NHWC tensor
     """
@@ -80,6 +80,8 @@ def FixedUnPooling(x, shape, unpool_mat=None):
         mat = np.zeros(shape, dtype='float32')
         mat[0][0] = 1
         unpool_mat = tf.Variable(mat, trainable=False, name='unpool_mat')
+    elif isinstance(unpool_mat, np.ndarray):
+        unpool_mat = tf.Variable(unpool_mat, trainable=False, name='unpool_mat')
     assert unpool_mat.get_shape().as_list() == list(shape)
 
     # perform a tensor-matrix kronecker product
@@ -96,6 +98,41 @@ def FixedUnPooling(x, shape, unpool_mat=None):
                             input_shape[3]])
     return prod
 
+@layer_register()
+def BilinearUpSample(x, shape):
+    """
+    Bilinear upsample the input images.
+    :param x: input NHWC tensor
+    :param shape: an integer
+    """
+    def bilinear_conv_filler(s):
+        """
+        s: width, height of the conv filter
+        See https://github.com/BVLC/caffe/blob/master/include%2Fcaffe%2Ffiller.hpp#L244
+        """
+        f = np.ceil(float(s) / 2)
+        c = float(2 * f - 1 - f % 2) / (2 * f)
+        ret = np.zeros((s, s), dtype='float32')
+        for x in range(s):
+            for y in range(s):
+                ret[x,y] = (1 - abs(x / f - c)) * (1 - abs(y / f - c))
+        return ret
+
+    ch = x.get_shape().as_list()[3]
+    shape = int(shape)
+    filter_shape = 2 * shape
+    w = bilinear_conv_filler(filter_shape)
+    w = np.repeat(w, ch * ch).reshape((filter_shape, filter_shape, ch, ch))
+    weight_var = tf.constant(w,
+                             tf.float32,
+                             shape=(filter_shape, filter_shape, ch, ch))
+
+    unpool_mat = np.zeros((shape, shape), dtype='float32')
+    unpool_mat[-1,-1] = 1
+    x = FixedUnPooling('unpool', x, shape, unpool_mat)
+    output = tf.nn.conv2d(x, weight_var, [1,1,1,1], padding='SAME')
+    return output
+
 from ._test import TestModel
 class TestPool(TestModel):
     def test_fixed_unpooling(self):
@@ -113,3 +150,23 @@ class TestPool(TestModel):
         # the rest are zeros
         res[0,::2,::2,0] = 0
         self.assertTrue((res == 0).all())
+
+    def test_upsample(self):
+        h, w = 5, 5
+        scale = 2
+
+        mat = np.random.rand(h, w).astype('float32')
+        inp = self.make_variable(mat)
+        inp = tf.reshape(inp, [1, h, w, 1])
+
+        output = BilinearUpSample('upsample', inp, scale)
+        res = self.run_variable(output)
+
+        from skimage.transform import rescale
+        res2 = rescale(mat, scale)
+
+        diff = np.abs(res2 - res[0,:,:,0])
+
+        # not equivalent at corner
+        diff[0,0] = diff[-1,-1] = 0
+        self.assertTrue(diff.max() < 1e-4)