some fixes

README.md

@@ -6,7 +6,7 @@ Still in development, but usable.
 See some interesting [examples](examples) to learn about the framework:
 
 + [DoReFa-Net: training binary / low bitwidth CNN](examples/DoReFa-Net)
-+ [Double-DQN for playing Atari games](examples/Atari2600)
++ [Double-DQN and A3C for playing Atari games](examples/Atari2600)
 + [ResNet for Cifar10 classification](examples/ResNet)
 + [IncpetionV3 on ImageNet](examples/Inception/inceptionv3.py)
 + [char-rnn language model](examples/char-rnn)

examples/HED/hed.py

@@ -88,17 +88,17 @@ class Model(ModelDesc):
             l = Conv2D('conv5_3', l, 512)
             b5 = branch('branch5', l, 16)
 
-        final_map = Conv2D('convfcweight',
-                tf.concat(3, [b1, b2, b3, b4, b5]), 1, 1,
-                W_init=tf.constant_initializer(0.2), use_bias=False)
-        final_map = tf.squeeze(final_map, [3], name='predmap')
-        #final_map = tf.squeeze(tf.mul(0.2, b1 + b2 + b3 + b4 + b5),
-                #[3], name='predmap')
+        #final_map = Conv2D('convfcweight',
+                #tf.concat(3, [b1, b2, b3, b4, b5]), 1, 1,
+                #W_init=tf.constant_initializer(0.2), use_bias=False)
+        #final_map = tf.squeeze(final_map, [3], name='predmap')
+        final_map = tf.squeeze(tf.mul(0.2, b1 + b2 + b3 + b4 + b5),
+                [3], name='predmap')
         costs = []
         for idx, b in enumerate([b1, b2, b3, b4, b5, final_map]):
             output = tf.nn.sigmoid(b, name='output{}'.format(idx+1))
-            xentropy = class_balanced_binary_class_cross_entropy(
-                output, edgemap,
+            xentropy = class_balanced_sigmoid_binary_class_cross_entropy(
+                b, edgemap,
                 name='xentropy{}'.format(idx+1))
             costs.append(xentropy)
 
@@ -138,6 +138,10 @@ def get_data(name):
             h0, w0, newh, neww = param
             return img[h0:h0+newh,w0:w0+neww]
 
+    def f(m):
+        m[m>=0.50] = 1
+        m[m<0.50] = 0
+        return m
     if isTrain:
         shape_aug = [
             imgaug.RandomResize(xrange=(0.7,1.5), yrange=(0.7,1.5),
@@ -152,18 +156,12 @@ def get_data(name):
         IMAGE_SHAPE = (320, 480)
         shape_aug = [imgaug.CenterCrop(IMAGE_SHAPE)]
     ds = AugmentImageComponents(ds, shape_aug, (0, 1))
-
-    def f(m):
-        m[m>=0.51] = 1
-        m[m<0.51] = 0
-        return m
     ds = MapDataComponent(ds, f, 1)
 
     if isTrain:
         augmentors = [
             imgaug.Brightness(63, clip=False),
             imgaug.Contrast((0.4,1.5)),
-            imgaug.GaussianNoise(),
         ]
         ds = AugmentImageComponent(ds, augmentors)
         ds = BatchDataByShape(ds, 8, idx=0)
@@ -212,7 +210,7 @@ def get_config():
 
 def run(model_path, image_path):
     pred_config = PredictConfig(
-            model=Model(False),
+            model=Model(),
             input_data_mapping=[0],
             session_init=get_model_loader(model_path),
             output_var_names=['output' + str(k) for k in range(1, 7)])

examples/README.md

@@ -5,7 +5,7 @@ Examples with __reproducible__ and meaningful performance.
 
 + [An illustrative mnist example](mnist-convnet.py)
 + [A tiny SVHN ConvNet with 97.5% accuracy](svhn-digit-convnet.py)
-+ [Reproduce some reinforcement learning papers](Atari2600)
++ Reinforcement learning (DQN, A3C) on [Atari games](Atari2600) and [demos on OpenAI Gym](OpenAIGym).
 + [char-rnn for fun](char-rnn)
 + [DisturbLabel, because I don't believe the paper](DisturbLabel)
 + [DoReFa-Net: binary / low-bitwidth CNN on ImageNet](DoReFa-Net)

tensorpack/callbacks/common.py

@@ -47,11 +47,11 @@ class ModelSaver(Callback):
             if name not in var_dict:
                 if name != v.name:
                     logger.info(
-                        "{} renamed to {} when saving model.".format(v.name, name))
+                        "[ModelSaver] {} renamed to {} when saving model.".format(v.name, name))
                 var_dict[name] = v
             else:
-                logger.warn("Variable {} won't be saved \
-because {} will be saved".format(v.name, var_dict[name].name))
+                logger.info("[ModelSaver] Variable {} won't be saved \
+due to an alternative in a different tower".format(v.name, var_dict[name].name))
         return var_dict
 
     def _trigger_epoch(self):

tensorpack/dataflow/imgaug/noise.py

@@ -23,9 +23,9 @@ class JpegNoise(ImageAugmentor):
 
 
 class GaussianNoise(ImageAugmentor):
-    def __init__(self, scale=10, clip=True):
+    def __init__(self, sigma=1, clip=True):
         """
-        Add a gaussian noise of the same shape to img.
+        Add a gaussian noise N(0, sigma^2) of the same shape to img.
         """
         super(GaussianNoise, self).__init__()
         self._init(locals())
@@ -34,7 +34,7 @@ class GaussianNoise(ImageAugmentor):
         return self.rng.randn(*img.shape)
 
     def _augment(self, img, noise):
-        ret = img + noise
+        ret = img + noise * self.sigma
         if self.clip:
             ret = np.clip(ret, 0, 255)
         return ret

tensorpack/dataflow/imgaug/noname.py

@@ -44,10 +44,9 @@ class Flip(ImageAugmentor):
     def _fprop_coord(self, coord, param):
         raise NotImplementedError()
 
-
 class Resize(ImageAugmentor):
     """ Resize image to a target size"""
-    def __init__(self, shape):
+    def __init__(self, shape, interp=cv2.INTER_CUBIC):
         """
         :param shape: shape in (h, w)
         """
@@ -56,7 +55,7 @@ class Resize(ImageAugmentor):
     def _augment(self, img, _):
         return cv2.resize(
             img, self.shape[::-1],
-            interpolation=cv2.INTER_CUBIC)
+            interpolation=self.interp)
 
 class RandomResize(ImageAugmentor):
     """ randomly rescale w and h of the image"""

tensorpack/models/batch_norm.py

@@ -42,9 +42,10 @@ def BatchNorm(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
 
     n_out = shape[-1]  # channel
     assert n_out is not None
-    beta = tf.get_variable('beta', [n_out])
+    beta = tf.get_variable('beta', [n_out],
+            initializer=tf.zeros_initializer)
     gamma = tf.get_variable('gamma', [n_out],
-        initializer=tf.ones_initializer)
+            initializer=tf.ones_initializer)
 
     if len(shape) == 2:
         batch_mean, batch_var = tf.nn.moments(x, [0], keep_dims=False)

tensorpack/tfutils/symbolic_functions.py

@@ -46,12 +46,38 @@ def class_balanced_binary_class_cross_entropy(pred, label, name='cross_entropy_l
     count_pos = tf.reduce_sum(y)
     beta = count_neg / (count_neg + count_pos)
 
-    eps = 1e-8
+    eps = 1e-12
     loss_pos = -beta * tf.reduce_mean(y * tf.log(z + eps))
     loss_neg = (1. - beta) * tf.reduce_mean((1. - y) * tf.log(1. - z + eps))
     cost = tf.sub(loss_pos, loss_neg, name=name)
     return cost
 
+def class_balanced_sigmoid_binary_class_cross_entropy(pred, label, name='cross_entropy_loss'):
+    """
+    The class-balanced cross entropy loss for binary classification,
+    as in `Holistically-Nested Edge Detection
+    <http://arxiv.org/abs/1504.06375>`_.
+
+    :param pred: size: b x ANYTHING. the logits.
+    :param label: size: b x ANYTHING. the ground truth in {0,1}.
+    :returns: class-balanced binary classification cross entropy loss
+    """
+    z = batch_flatten(pred)
+    y = tf.cast(batch_flatten(label), tf.float32)
+
+    count_neg = tf.reduce_sum(1. - y)
+    count_pos = tf.reduce_sum(y)
+    beta = count_neg / (count_neg + count_pos)
+
+    #eps = 1e-12
+    logstable = tf.log(1 + tf.exp(-tf.abs(z)))
+    loss_pos = -beta * tf.reduce_mean(-y *
+            (logstable - tf.minimum(0, z)))
+    loss_neg = (1. - beta) * tf.reduce_mean((y - 1.) *
+            (logstable + tf.maximum(z, 0)))
+    cost = tf.sub(loss_pos, loss_neg, name=name)
+    return cost
+
 def print_stat(x, message=None):
     """ a simple print op.
         Use it like: x = print_stat(x)

tensorpack/utils/utils.py

@@ -112,7 +112,6 @@ def get_dataset_path(*args):
     assert os.path.isdir(d), d
     return os.path.join(d, *args)
 
-
 def get_tqdm_kwargs(**kwargs):
     default = dict(
             smoothing=0.5,