change ConcatWith.

CHANGES.md

@@ -8,6 +8,10 @@ so you won't need to look at here very often.
 Here are a list of things that were changed, starting from an early version.
 TensorFlow itself also changes API and those are not listed here.
 
+* 2017/01/25. Argument order of `models.ConcatWith` is changed to follow the API change in
+	TensorFlow upstream.
+* 2017/01/25. `TrainConfig(callbacks=)` now takes a list of `Callback` instances. See [commit](https://github.com/ppwwyyxx/tensorpack/commit/243e957fe6d62a0cfb5728bd77fb3e005d6603e4)
+	on how to change your code.
 * 2017/01/06. `summary.add_moving_summary` now takes any number of positional arguments instead of a list.
 	See [commit](https://github.com/ppwwyyxx/tensorpack/commit/bbf41d9e58053f843d0471e6d2d87ff714a79a90) to change your code.
 * 2017/01/05. The argument `TrainConfig(dataset=)` is renamed to `TrainConfig(dataflow=)`.

examples/GAN/Image2Image.py

@@ -68,21 +68,21 @@ class Model(GANModelDesc):
                 return (LinearWrap(e8)
                         .Deconv2D('deconv1', NF * 8)
                         .Dropout()
-                        .ConcatWith(3, e7)
+                        .ConcatWith(e7, 3)
                         .Deconv2D('deconv2', NF * 8)
                         .Dropout()
-                        .ConcatWith(3, e6)
+                        .ConcatWith(e6, 3)
                         .Deconv2D('deconv3', NF * 8)
                         .Dropout()
-                        .ConcatWith(3, e5)
+                        .ConcatWith(e5, 3)
                         .Deconv2D('deconv4', NF * 8)
-                        .ConcatWith(3, e4)
+                        .ConcatWith(e4, 3)
                         .Deconv2D('deconv5', NF * 4)
-                        .ConcatWith(3, e3)
+                        .ConcatWith(e3, 3)
                         .Deconv2D('deconv6', NF * 2)
-                        .ConcatWith(3, e2)
+                        .ConcatWith(e2, 3)
                         .Deconv2D('deconv7', NF * 1)
-                        .ConcatWith(3, e1)
+                        .ConcatWith(e1, 3)
                         .Deconv2D('deconv8', OUT_CH, nl=tf.tanh)())
 
     def discriminator(self, inputs, outputs):

tensorpack/models/shapes.py

@@ -10,15 +10,15 @@ __all__ = ['ConcatWith']
 
 
 @layer_register(use_scope=False, log_shape=False)
-def ConcatWith(x, dim, tensor):
+def ConcatWith(x, tensor, dim):
     """
-    A wrapper around ``tf.concat`` to cooperate with :class:`LinearWrap`.
+    A wrapper around ``tf.concat_v2`` to cooperate with :class:`LinearWrap`.
 
     Args:
         x (tf.Tensor): input
-        dim (int): the dimension along which to concatenate
         tensor (list[tf.Tensor]): a tensor or list of tensors to concatenate with x.
             x will be at the beginning
+        dim (int): the dimension along which to concatenate
 
     Returns:
         tf.Tensor: ``tf.concat_v2([x] + tensor, dim)``

tensorpack/tfutils/distributions.py

@@ -116,8 +116,13 @@ class Distribution(object):
 
     @class_scope
     def encoder_activation(self, dist_param):
-        """ An activation function to produce
-            feasible distribution parameters from unconstrained raw network output.
+        """ An activation function which transform unconstrained raw network output
+            to a vector of feasible distribution parameters.
+
+            Note that for each distribution,
+            there are many feasible ways to design this function and it's hard to say which is better.
+            The default implementations in the distrbution classes here is
+            just one reasonable way to do this.
 
         Args:
             dist_param: output from a network, of shape (batch, param_dim).
@@ -218,8 +223,9 @@ class GaussianDistribution(Distribution):
             return dist_param
         else:
             mean, stddev = tf.split(dist_param, 2, axis=1)
-            # this is from https://github.com/openai/InfoGAN. don't know why
-            stddev = tf.sqrt(tf.exp(stddev))
+            stddev = tf.exp(stddev)  # just make it positive and assume it's stddev
+            # OpenAI code assumes exp(input) is variance. https://github.com/openai/InfoGAN.
+            # not sure if there is any theory about this.
             return tf.concat_v2([mean, stddev], axis=1)
 
     def _sample(self, batch_size, theta):