name change

README.md

@@ -12,8 +12,8 @@ You can actually train them and reproduce the performance... not just to see how
 + [Fully-convolutional Network for Holistically-Nested Edge Detection](examples/HED)
 + [Spatial Transformer Networks on MNIST addition](examples/SpatialTransformer)
 + [Generative Adversarial Networks](examples/GAN)
-+ [Double DQN plays Atari games](examples/Atari2600)
-+ [Asynchronous Advantage Actor-Critic(A3C) with demos on OpenAI Gym Atari games](examples/OpenAIGym)
++ [DQN variants on Atari games](examples/Atari2600)
++ [Asynchronous Advantage Actor-Critic(A3C) with demos on OpenAI Gym](examples/OpenAIGym)
 + [char-rnn language model](examples/char-rnn)
 
 ## Features:

examples/GAN/celebA.py → examples/GAN/DCGAN-CelebA.py

 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
-# File: celebA.py
+# File: DCGAN-CelebA.py
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
 import numpy as np
@@ -68,8 +68,9 @@ class Model(ModelDesc):
     def _build_graph(self, input_vars):
         image_pos = input_vars[0]
         image_pos = image_pos / 128.0 - 1
-        z = tf.random_uniform(tf.pack([tf.shape(image_pos)[0], 100]), -1, 1, name='z')
-        z.set_shape([None, 100])    # issue#5680
+
+        z = tf.random_uniform([BATCH, 100], -1, 1, name='z_train')
+        z = tf.placeholder_with_default(z, [None, 100], name='z')
 
         with argscope([Conv2D, Deconv2D, FullyConnected],
                 W_init=tf.truncated_normal_initializer(stddev=0.02)):
@@ -110,7 +111,7 @@ def get_config():
         session_config=get_default_sess_config(0.5),
         model=Model(),
         step_per_epoch=300,
-        max_epoch=500,
+        max_epoch=300,
     )
 
 def sample(model_path):
@@ -119,7 +120,7 @@ def sample(model_path):
        model=Model(),
        input_names=['z'],
        output_names=['gen/gen'])
-    pred = SimpleDatasetPredictor(pred, RandomZData((128, 100)))
+    pred = SimpleDatasetPredictor(pred, RandomZData((100, 100)))
     for o in pred.get_result():
         o = o[0] + 1
         o = o * 128.0
@@ -141,6 +142,7 @@ if __name__ == '__main__':
     if args.sample:
         sample(args.load)
     else:
+        assert args.data
         config = get_config()
         if args.load:
             config.session_init = SaverRestore(args.load)

examples/HED/hed.py

@@ -70,7 +70,7 @@ class Model(ModelDesc):
         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_sigmoid_binary_class_cross_entropy(
+            xentropy = class_balanced_sigmoid_cross_entropy(
                 b, edgemap,
                 name='xentropy{}'.format(idx+1))
             costs.append(xentropy)

examples/Inception/inception-bn.py

@@ -124,6 +124,7 @@ def get_data(train_or_test):
     pp_mean = meta.get_per_pixel_mean()
 
     if isTrain:
+        # TODO use the augmentor in GoogleNet
         augmentors = [
             imgaug.Resize((256, 256)),
             imgaug.Brightness(30, False),

tensorpack/models/regularize.py

@@ -40,8 +40,8 @@ def regularize_cost(regex, func, name=None):
 
 
 @layer_register(log_shape=False)
-def Dropout(x, prob=0.5):
+def Dropout(x, keep_prob=0.5):
     is_training = get_current_tower_context().is_training
-    keep_prob = tf.constant(prob if is_training else 1.0)
+    keep_prob = tf.constant(keep_prob if is_training else 1.0)
     return tf.nn.dropout(x, keep_prob)
 

tensorpack/tfutils/symbolic_functions.py

@@ -27,18 +27,18 @@ def batch_flatten(x):
     """
     shape = x.get_shape().as_list()[1:]
     if None not in shape:
-        return tf.reshape(x, [-1, np.prod(shape)])
+        return tf.reshape(x, [-1, int(np.prod(shape))])
     return tf.reshape(x, tf.pack([tf.shape(x)[0], -1]))
 
-def class_balanced_binary_class_cross_entropy(pred, label, name='cross_entropy_loss'):
+def class_balanced_cross_entropy(pred, label, name='cross_entropy_loss'):
     """
-    The class-balanced cross entropy loss for binary classification,
+    The class-balanced cross entropy loss,
     as in `Holistically-Nested Edge Detection
     <http://arxiv.org/abs/1504.06375>`_.
 
     :param pred: size: b x ANYTHING. the predictions in [0,1].
     :param label: size: b x ANYTHING. the ground truth in {0,1}.
-    :returns: class-balanced binary classification cross entropy loss
+    :returns: class-balanced cross entropy loss
     """
     z = batch_flatten(pred)
     y = tf.cast(batch_flatten(label), tf.float32)
@@ -53,30 +53,34 @@ def class_balanced_binary_class_cross_entropy(pred, label, name='cross_entropy_l
     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'):
+def class_balanced_sigmoid_cross_entropy(logits, label, name='cross_entropy_loss'):
     """
-    The class-balanced cross entropy loss for binary classification,
+    The class-balanced cross entropy loss,
     as in `Holistically-Nested Edge Detection
     <http://arxiv.org/abs/1504.06375>`_.
+    This is more numerically stable than class_balanced_cross_entropy
 
-    :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
+    :param logits: size: the logits.
+    :param label: size: the ground truth in {0,1}, of the same shape as logits.
+    :returns: a scalar. class-balanced cross entropy loss
     """
-    z = batch_flatten(pred)
+    z = batch_flatten(logits)
     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(0.0, z)))
-    loss_neg = (1. - beta) * tf.reduce_mean((y - 1.) *
-            (logstable + tf.maximum(z, 0.0)))
-    cost = tf.sub(loss_pos, loss_neg, name=name)
+    pos_weight = beta / (1 - beta)
+    cost = tf.nn.weighted_cross_entropy_with_logits(z, y, pos_weight)
+    cost = tf.reduce_mean(cost * (1 - beta), name=name)
+
+    #logstable = tf.log(1 + tf.exp(-tf.abs(z)))
+    #loss_pos = -beta * tf.reduce_mean(-y *
+            #(logstable - tf.minimum(0.0, z)))
+    #loss_neg = (1. - beta) * tf.reduce_mean((y - 1.) *
+            #(logstable + tf.maximum(z, 0.0)))
+    #cost = tf.sub(loss_pos, loss_neg, name=name)
     return cost
 
 def print_stat(x, message=None):

tensorpack/utils/viz.py

@@ -98,7 +98,20 @@ def build_patch_list(patch_list,
 
 def dump_dataflow_images(df, index=0, batched=True,
         number=300, output_dir=None,
-        scale=1, resize=None, viz=None, flipRGB=False, exit_after=True):
+        scale=1, resize=None, viz=None,
+        flipRGB=False, exit_after=True):
+    """
+    :param df: a DataFlow
+    :param index: the index of the image component
+    :param batched: whether the component contains batched images or not
+    :param number: how many datapoint to take from the DataFlow
+    :param output_dir: output directory to save images, default to not save.
+    :param scale: scale the value, usually either 1 or 255
+    :param resize: (h, w) or Nne, resize the images
+    :param viz: (h, w) or None, visualize the images in grid with imshow
+    :param flipRGB: apply a RGB<->BGR conversion or not
+    :param exit_after: exit the process after this function
+    """
     if output_dir:
         mkdir_p(output_dir)
     if viz is not None: