update and fix grad bug

examples/cifar10_convnet.py

@@ -67,8 +67,7 @@ class Model(ModelDesc):
         l = FullyConnected('fc1', l, out_dim=512,
                            b_init=tf.constant_initializer(0.1))
         # fc will have activation summary by default. disable for the output layer
-        logits = FullyConnected('linear', l, out_dim=10, summary_activation=False,
-                                nl=tf.identity)
+        logits = FullyConnected('linear', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
         y = one_hot(label, 10)

examples/cifar10_resnet.py

@@ -102,8 +102,7 @@ class Model(ModelDesc):
         l = tf.nn.relu(l)
         # 8,c=64
         l = GlobalAvgPooling('gap', l)
-        logits = FullyConnected('linear', l, out_dim=10, summary_activation=False,
-                                nl=tf.identity)
+        logits = FullyConnected('linear', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
         y = one_hot(label, 10)

examples/load_alexnet.py

@@ -56,7 +56,7 @@ class Model(ModelDesc):
         l = FullyConnected('fc6', l, 4096)
         l = FullyConnected('fc7', l, out_dim=4096)
         # fc will have activation summary by default. disable this for the output layer
-        logits = FullyConnected('fc8', l, out_dim=1000, summary_activation=False, nl=tf.identity)
+        logits = FullyConnected('fc8', l, out_dim=1000, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
         y = one_hot(label, 1000)

examples/load_vgg16.py

@@ -64,7 +64,7 @@ class Model(ModelDesc):
         l = tf.nn.dropout(l, keep_prob)
         l = FullyConnected('fc7', l, 4096)
         l = tf.nn.dropout(l, keep_prob)
-        logits = FullyConnected('fc8', l, out_dim=1000, summary_activation=False, nl=tf.identity)
+        logits = FullyConnected('fc8', l, out_dim=1000, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
         y = one_hot(label, 1000)

examples/mnist_convnet.py

@@ -54,8 +54,7 @@ class Model(ModelDesc):
         l = tf.nn.dropout(l, keep_prob)
 
         # fc will have activation summary by default. disable this for the output layer
-        logits = FullyConnected('fc1', l, out_dim=10,
-                             summary_activation=False, nl=tf.identity)
+        logits = FullyConnected('fc1', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='prob')
 
         y = one_hot(label, 10)

examples/svhn_digit_convnet.py

@@ -45,8 +45,7 @@ class Model(ModelDesc):
         l = FullyConnected('fc0', l, 512,
                            b_init=tf.constant_initializer(0.1))
         # fc will have activation summary by default. disable for the output layer
-        logits = FullyConnected('linear', l, out_dim=10, summary_activation=False,
-                                nl=tf.identity)
+        logits = FullyConnected('linear', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
         y = one_hot(label, 10)

tensorpack/dataflow/dataset/ilsvrc.py

@@ -11,12 +11,15 @@ from ...utils import logger, get_rng
 from ..base import DataFlow
 from ...utils.fs import mkdir_p, download
 
-__all__ = ['ILSVRCMeta']
+__all__ = ['ILSVRCMeta', 'ILSVRC12']
 
 CAFFE_ILSVRC12_URL = "http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz"
 CAFFE_PROTO_URL = "https://github.com/BVLC/caffe/raw/master/src/caffe/proto/caffe.proto"
 
 class ILSVRCMeta(object):
+    """
+    Provide metadata for ILSVRC dataset.
+    """
     def __init__(self, dir=None):
         if dir is None:
             dir = os.path.join(os.path.dirname(__file__), 'ilsvrc_metadata')
@@ -24,15 +27,18 @@ class ILSVRCMeta(object):
         mkdir_p(self.dir)
         self.caffe_pb_file = os.path.join(self.dir, 'caffe_pb2.py')
         if not os.path.isfile(self.caffe_pb_file):
-            self.download_caffe_meta()
+            self._download_caffe_meta()
 
     def get_synset_words_1000(self):
+        """
+        :returns a dict of {cls_number: cls_name}
+        """
         fname = os.path.join(self.dir, 'synset_words.txt')
         assert os.path.isfile(fname)
         lines = [x.strip() for x in open(fname).readlines()]
         return dict(enumerate(lines))
 
-    def download_caffe_meta(self):
+    def _download_caffe_meta(self):
         fpath = download(CAFFE_ILSVRC12_URL, self.dir)
         tarfile.open(fpath, 'r:gz').extractall(self.dir)
 
@@ -41,6 +47,10 @@ class ILSVRCMeta(object):
         assert ret == 0, "caffe proto compilation failed!"
 
     def get_image_list(self, name):
+        """
+        :param name: 'train' or 'val' or 'test'
+        :returns list of image filenames
+        """
         assert name in ['train', 'val', 'test']
         fname = os.path.join(self.dir, name + '.txt')
         assert os.path.isfile(fname)
@@ -51,10 +61,9 @@ class ILSVRCMeta(object):
                 ret.append((name, int(cls)))
             return ret
 
-    def load_mean(self):
+    def get_per_pixel_mean(self):
         """
-        return per-pixel mean as an array of shape
-         (3, 256, 256) in range [0, 255]
+        :returns per-pixel mean as an array of shape (3, 256, 256) in range [0, 255]
         """
         import imp
         caffepb = imp.load_source('caffepb', self.caffe_pb_file)
@@ -83,9 +92,15 @@ class ILSVRC12(DataFlow):
         return len(self.imglist)
 
     def reset_state(self):
+        """
+        reset rng for shuffle
+        """
         self.rng = get_rng(self)
 
     def get_data(self):
+        """
+        Produce original images or shape [h, w, 3], and label
+        """
         idxs = np.arange(len(self.imglist))
         if self.shuffle:
             self.rng.shuffle(idxs)
@@ -99,7 +114,7 @@ class ILSVRC12(DataFlow):
 
 if __name__ == '__main__':
     meta = ILSVRCMeta()
-    print meta.load_mean()
+    print meta.get_per_pixel_mean()
     #print(meta.get_synset_words_1000())
 
     #ds = ILSVRC12('/home/wyx/data/imagenet', 'val')

tensorpack/dataflow/imgaug/deform.py

@@ -3,6 +3,7 @@
 # Author: Yuxin Wu <ppwwyyxx@gmail.com>
 
 from .base import ImageAugmentor
+from ...utils import logger
 import numpy as np
 
 __all__ = ['GaussianDeform', 'GaussianMap']
@@ -63,6 +64,7 @@ class GaussianDeform(ImageAugmentor):
         :param sigma: sigma for Gaussian weight
         :param randrange: default to shape[0] / 8
         """
+        logger.warn("GaussianDeform is slow. Consider using it with 4 or more prefetching processes.")
         super(GaussianDeform, self).__init__()
         self.anchors = anchors
         self.K = len(self.anchors)

tensorpack/models/conv2d.py

@@ -10,7 +10,7 @@ from ._common import *
 
 __all__ = ['Conv2D']
 
-@layer_register(summary_activation=True)
+@layer_register()
 def Conv2D(x, out_channel, kernel_shape,
            padding='SAME', stride=1,
            W_init=None, b_init=None,

tensorpack/models/fc.py

@@ -11,7 +11,7 @@ from ..tfutils.symbolic_functions import *
 
 __all__ = ['FullyConnected']
 
-@layer_register(summary_activation=True)
+@layer_register()
 def FullyConnected(x, out_dim,
                    W_init=None, b_init=None,
                    nl=tf.nn.relu, use_bias=True):

tensorpack/tfutils/gradproc.py

@@ -21,7 +21,7 @@ class GradientProcessor(object):
         :param grads: list of (grad, var)
         :returns: symbolic gradients with the same type as input
         """
-        self._process(grads)
+        return self._process(grads)
 
     @abstractmethod
     def _process(self, grads):

tensorpack/train/base.py

@@ -57,7 +57,7 @@ class Trainer(object):
             logger.LOG_DIR, graph=self.sess.graph)
         self.summary_op = tf.merge_all_summaries()
         # create an empty StatHolder
-        self.stat_holder = StatHolder(logger.LOG_DIR, [])
+        self.stat_holder = StatHolder(logger.LOG_DIR)
 
     def _process_summary(self, summary_str):
         summary = tf.Summary.FromString(summary_str)