pyutil

diff --git a/src/caffe/pyutil/__init__.py b/src/caffe/pyutil/__init__.py
new file mode 100644 (file)
index 0000000..e69de29

diff --git a/src/caffe/pyutil/convert_ilsvrc.py b/src/caffe/pyutil/convert_ilsvrc.py
new file mode 100644 (file)
index 0000000..c494c01
--- /dev/null
+++ b/src/caffe/pyutil/convert_ilsvrc.py
@@ -0,0 +1,63 @@
+"""This script converts images stored in the ILSVRC format to a leveldb,
+converting every image to a 256*256 image as well as converting them to channel
+first storage. The output will be shuffled - so that a sequential read will
+result in pseudo-random minibatches.
+"""
+
+from decaf import util
+from decaf.util import transform
+import glob
+import leveldb
+import numpy as np
+import random
+import os
+from skimage import io
+import sys
+
+from caffe.proto import caffe_pb2
+
+def main(argv):
+  root = argv[0]
+  db_name = argv[1]
+  db = leveldb.LevelDB(db_name)
+  synsets = glob.glob(os.path.join(root, "n????????"))
+  synsets.sort()
+  print 'A total of %d synsets' % len(synsets)
+  all_files = [glob.glob(os.path.join(root, synset, "*.JPEG"))
+      for synset in synsets]
+  all_labels = [[i] * len(files) for i, files in enumerate(all_files)]
+  all_files = sum(all_files, [])
+  all_labels = sum(all_labels, [])
+  print 'A total of %d files' % len(all_files)
+  random_indices = list(range(len(all_files)))
+  random.shuffle(random_indices)
+  datum = caffe_pb2.Datum()
+  datum.blob.num = 1
+  datum.blob.channels = 3
+  datum.blob.height = 256
+  datum.blob.width = 256
+  my_timer = util.Timer()
+  batch = leveldb.WriteBatch()
+  for i in range(len(all_files))[:1281]:
+    filename = all_files[random_indices[i]]
+    basename = os.path.basename(filename)
+    label = all_labels[random_indices[i]]
+    image = io.imread(filename)
+    image = transform.scale_and_extract(transform.as_rgb(image), 256)
+    image = np.ascontiguousarray(image.swapaxes(1,2).swapaxes(0,1))
+    del datum.blob.data[:]
+    datum.blob.data.extend(list(image.flatten()))
+    datum.label = label
+    batch.Put('%d_%d_%s' % (i, label, basename),
+        datum.SerializeToString())
+    print '(%d %s) Wrote file %s' % (i, my_timer.total(), basename)
+    if (i % 256 and i > 0):
+      # write and start a new batch
+      db.Write(batch)
+      batch = leveldb.WriteBatch()
+
+if __name__ == '__main__':
+  if len(sys.argv) != 3:
+    print 'Usage: convert_ilsvrc.py DATA_ROOT OUTPUT_DB'
+  else:
+    main(sys.argv[1:])
\ No newline at end of file