[WIP] trigger_step with fetch

examples/A3C-Gym/README.md

@@ -6,7 +6,7 @@ Implemented A3C in [Asynchronous Methods for Deep Reinforcement Learning](http:/
 
 `./train-atari.py --env Breakout-v0 --gpu 0`
 
-It should run at a speed of 6~10 iteration/s on 1 GPU.
+It should run at a speed of 6~10 iteration/s on 1 GPU plus 12+ CPU cores.
 Training with a significant slower speed (e.g. on CPU) will give bad performance,
 probably because of async issues.
 The pre-trained models are all trained with 4 GPUs for about 2 days.

examples/README.md

@@ -13,6 +13,7 @@ Training examples with __reproducible__ and meaningful performance.
 + [Fully-convolutional Network for Holistically-Nested Edge Detection(HED)](HED)
 + [Spatial Transformer Networks on MNIST addition](SpatialTransformer)
 + [Visualize Saliency Maps by Guided ReLU](Saliency)
++ [Similarity Learning on MNIST](SimilarityLearning)
 + Load a pre-trained [AlexNet](load-alexnet.py) or [VGG16](load-vgg16.py) model.
 + Load a pre-trained [Convolutional Pose Machines](ConvolutionalPoseMachines/).
 

examples/SpatialTransformer/README.md

@@ -11,7 +11,9 @@ and warped them separately.
 
 <p align="center"> <img src="./demo.jpg" width="400"> </p>
 
-Left: input image; Middle: output of the first STN branch (which localizes the second digit); Right: output of the second STN branch.
+* Left: input image.
+* Middle: output of the first STN branch (which localizes the second digit).
+* Right: output of the second STN branch.
 
 To train (takes about 300 epochs to reach 8.8% error):
 ```bash

tensorpack/callbacks/base.py

@@ -5,6 +5,7 @@
 import tensorflow as tf
 from abc import ABCMeta
 import six
+from ..tfutils.common import get_op_or_tensor_by_name
 
 __all__ = ['Callback', 'PeriodicCallback', 'ProxyCallback', 'CallbackFactory']
 
@@ -49,12 +50,42 @@ class Callback(object):
     def _before_train(self):
         pass
 
-    def trigger_step(self):
+    def trigger_step(self, *args):
         """
-        Callback to be triggered after every step (every backpropagation)
+        Callback to be triggered after every step (every backpropagation).
+
+        Args:
+            args: a list of values corresponding to :meth:`extra_fetches`.
 
         Could be useful to apply some tricks on parameters (clipping, low-rank, etc)
         """
+        self._trigger_step(*args)
+
+    def _trigger_step(self, *args):
+        pass
+
+    def extra_fetches(self):
+        """
+        Returns:
+            list: a list of elements to be fetched in every step and
+                passed to :meth:`trigger_step`. Elements can be
+                Operations/Tensors, or names of Operations/Tensors.
+
+        This function will be called only after the graph is finalized.
+
+        This function should be a pure function (i.e. no side-effect when called)
+        """
+        fetches = self._extra_fetches()
+        ret = []
+        for f in fetches:
+            if isinstance(f, (tf.Tensor, tf.Operation)):
+                ret.append(f)
+            else:
+                ret.append(get_op_or_tensor_by_name(f))
+        return ret
+
+    def _extra_fetches(self):
+        return []
 
     def trigger_epoch(self):
         """
@@ -110,8 +141,6 @@ class ProxyCallback(Callback):
 class PeriodicCallback(ProxyCallback):
     """
     Wrap a callback so that it is triggered after every ``period`` epochs.
-
-    Doesn't work for ``trigger_step``.
     """
 
     def __init__(self, cb, period):

tensorpack/callbacks/group.py

@@ -4,6 +4,7 @@
 
 import tensorflow as tf
 from contextlib import contextmanager
+from collections import defaultdict
 import time
 
 from .base import Callback
@@ -67,6 +68,7 @@ class Callbacks(Callback):
             raise ValueError("Callbacks must contain StatPrinter for stat and writer to work properly!")
 
         self.cbs = cbs
+        self._extra_fetches_cache = None
 
     def _setup_graph(self):
         with tf.name_scope(None):
@@ -81,9 +83,30 @@ class Callbacks(Callback):
         for cb in self.cbs:
             cb.after_train()
 
-    def trigger_step(self):
-        for cb in self.cbs:
-            cb.trigger_step()
+    def _extra_fetches(self):
+        if self._extra_fetches_cache is not None:
+            return self._extra_fetches_cache
+        # TODO use dispatch mechanism to avoid duplication
+        self._cbid_to_fetchid = defaultdict(list)
+        ret = []
+        for idx, cb in enumerate(self.cbs):
+            fetch = cb.extra_fetches()
+            if len(fetch) == 0:
+                continue
+            for f in fetch:
+                ret.append(f)
+                self._cbid_to_fetchid[idx].append(len(ret)-1)
+        self._extra_fetches_cache = ret
+        return ret
+
+    def _trigger_step(self, *args):
+        for idx, cb in enumerate(self.cbs):
+            fid = self._cbid_to_fetchid[idx]
+            if len(fid) == 0:
+                cb.trigger_step()
+            else:
+                data = [args[k] for k in fid]
+                cb.trigger_step(*data)
 
     def _trigger_epoch(self):
         tm = CallbackTimeLogger()

tensorpack/tfutils/summary.py

@@ -137,8 +137,8 @@ def summary_moving_average(tensors=None):
     with tf.name_scope(None):
         averager = tf.train.ExponentialMovingAverage(
             0.95, num_updates=get_global_step_var(), name='EMA')
-    avg_maintain_op = averager.apply(tensors)
-    for idx, c in enumerate(tensors):
-        name = re.sub('tower[p0-9]+/', '', c.op.name)
-        tf.summary.scalar(name + '-summary', averager.average(c))
+        avg_maintain_op = averager.apply(tensors)
+        for idx, c in enumerate(tensors):
+            name = re.sub('tower[p0-9]+/', '', c.op.name)
+            tf.summary.scalar(name + '-summary', averager.average(c))
     return avg_maintain_op