update dorefa to use floor instead of round

examples/DoReFa-Net/README.md

@@ -3,7 +3,8 @@ Code and model for the paper:
 [DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients](http://arxiv.org/abs/1606.06160), by Zhou et al.
 
 We hosted a demo at CVPR16 on behalf of Megvii, Inc, running a real-time 1/4-VGG size DoReFa-Net on ARM and half-VGG size DoReFa-Net on FPGA.
-We're not planning to release those runtime bit-op libraries for now. In this repo, bit operations are run in float32.
+We're not planning to release our C++ runtime for bit-operations.
+In this repo, bit operations are performed through `tf.float32`.
 
 Pretrained model for 1-2-6-AlexNet is available at
 [google drive](https://drive.google.com/a/%20megvii.com/folderview?id=0B308TeQzmFDLa0xOeVQwcXg1ZjQ).
@@ -13,7 +14,7 @@ It's provided in the format of numpy dictionary, so it should be very easy to po
 
 To use the script. You'll need:
 
-+ TensorFlow 0.10,0.11rc1,0.11rc2. 0.11 is not supported due to [TF bug](https://github.com/tensorflow/tensorflow/issues/5888)
++ TensorFlow >= 0.10
 
 + OpenCV bindings for Python
 

examples/DoReFa-Net/dorefa.py

@@ -17,7 +17,7 @@ def get_dorefa(bitW, bitA, bitG):
     def quantize(x, k):
         n = float(2**k-1)
         with G.gradient_override_map({"Floor": "Identity"}):
-            return tf.round(x * n) / n
+            return tf.floor(x * n + 0.5) / n
 
     def fw(x):
         if bitW == 32:

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -30,7 +30,7 @@ Accuracy:
     With (W,A,G)=(32,32,32), error is about 2.9%.
 
 Speed:
-    About 18 iteration/s on 1 Tesla M40. (4721 iterations / epoch)
+    30~35 iteration/s on 1 TitanX Pascal. (4721 iterations / epoch)
 
 To Run:
     ./svhn-digit-dorefa.py --dorefa 1,2,4
@@ -45,8 +45,9 @@ class Model(ModelDesc):
         return [InputVar(tf.float32, [None, 40, 40, 3], 'input'),
                 InputVar(tf.int32, [None], 'label') ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         image, label = input_vars
+        is_training = get_current_tower_context().is_training
 
         fw, fa, fg = get_dorefa(BITW, BITA, BITG)
         # monkey-patch tf.get_variable to apply fw

tensorpack/__init__.py

@@ -15,3 +15,6 @@ from tensorpack.predict import *
 
 if int(numpy.__version__.split('.')[1]) < 9:
     logger.warn("Numpy < 1.9 could be extremely slow on some tasks.")
+
+if get_tf_version() < 10:
+    logger.error("tensorpack requires TensorFlow >= 0.10")

tensorpack/dataflow/dftools.py

@@ -8,6 +8,7 @@ import multiprocessing as mp
 import six
 from six.moves import range, map
 
+from .base import DataFlow
 from ..utils import get_tqdm, logger
 from ..utils.concurrency import DIE
 from ..utils.serialize import dumps
@@ -43,6 +44,7 @@ def dump_dataset_images(ds, dirname, max_count=None, index=0):
         cv2.imwrite(os.path.join(dirname, "{}.jpg".format(i)), img)
 
 def dump_dataflow_to_lmdb(ds, lmdb_path):
+    assert isinstance(ds, DataFlow), type(ds)
     isdir = os.path.isdir(lmdb_path)
     if isdir:
         assert not os.path.isfile(os.path.join(lmdb_path, 'data.mdb')), "LMDB file exists!"
@@ -52,16 +54,20 @@ def dump_dataflow_to_lmdb(ds, lmdb_path):
     db = lmdb.open(lmdb_path, subdir=isdir,
             map_size=1099511627776 * 2, readonly=False,
             meminit=False, map_async=True)    # need sync() at the end
-    with get_tqdm(total=ds.size()) as pbar:
+    try:
+        sz = ds.size()
+    except NotImplementedError:
+        sz = 0
+    with get_tqdm(total=sz) as pbar:
         with db.begin(write=True) as txn:
             for idx, dp in enumerate(ds.get_data()):
                 txn.put(six.binary_type(idx), dumps(dp))
                 pbar.update()
             keys = list(map(six.binary_type, range(idx + 1)))
             txn.put('__keys__', dumps(keys))
-    logger.info("Flushing database ...")
-    db.sync()
-    db.close()
+
+            logger.info("Flushing database ...")
+            db.sync()
 
 
 def dataflow_to_process_queue(ds, size, nr_consumer):

tensorpack/models/batch_norm.py

@@ -9,6 +9,7 @@ from tensorflow.python.training import moving_averages
 from copy import copy
 import re
 
+from ..tfutils.common import get_tf_version
 from ..tfutils.tower import get_current_tower_context
 from ..utils import logger
 from ._common import layer_register
@@ -177,4 +178,8 @@ def BatchNormV2(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
     else:
         return tf.identity(xn, name='output')
 
-BatchNorm = BatchNormV2
+if get_tf_version() >= 11:
+    BatchNorm = BatchNormV2
+else:
+    logger.warn("BatchNorm might be faster if you update TensorFlow")
+    BatchNorm = BatchNormV1

tensorpack/tfutils/common.py

@@ -19,7 +19,8 @@ __all__ = ['get_default_sess_config',
            'backup_collection',
            'restore_collection',
            'clear_collection',
-           'freeze_collection']
+           'freeze_collection',
+           'get_tf_version']
 
 def get_default_sess_config(mem_fraction=0.99):
     """
@@ -104,3 +105,6 @@ def freeze_collection(keys):
     backup = backup_collection(keys)
     yield
     restore_collection(backup)
+
+def get_tf_version():
+    return int(tf.__version__.split('.')[1])

tensorpack/utils/fs.py

@@ -8,7 +8,7 @@ from six.moves import urllib
 import errno
 from . import logger
 
-__all__ = ['mkdir_p', 'download']
+__all__ = ['mkdir_p', 'download', 'recursive_walk']
 
 def mkdir_p(dirname):
     """ make a dir recursively, but do nothing if the dir exists"""
@@ -21,7 +21,6 @@ def mkdir_p(dirname):
         if e.errno != errno.EEXIST:
             raise e
 
-
 def download(url, dir):
     mkdir_p(dir)
     fname = url.split('/')[-1]
@@ -46,5 +45,10 @@ def download(url, dir):
     print('Succesfully downloaded ' + fname + " " + str(size) + ' bytes.')
     return fpath
 
+def recursive_walk(rootdir):
+    for r, dirs, files in os.walk(rootdir):
+        for f in files:
+            yield os.path.join(r, f)
+
 if __name__ == '__main__':
     download('http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz', '.')