reran everything to make sure its reproducible and added a few plots

Author: Qwlouse

Create Animations.ipynb

@@ -1,28 +1,11 @@
 # Binding by Reconstruction Clustering
 
 This is the code repository complementing the paper ["Binding by Reconstruction Clustering"](http://arxiv.org/abs/1511.06418).
+Everything from the paper starting from the datasets all the way to the plots can be
+reproduced from this repository.
 
-## Demo Images
-Here are some animations of the convergence behaviour of Reconstruction Clustering.
-
-### Shapes
 ![shapes animation](animations/shapes.gif)
 
-### Multi MNIST
-![Multi-MNIST animation](animations/multi_mnist.gif)
-
-### Corners
-![Corners animation](animations/corners.gif)
-
-### Bars
-![Bars animation](animations/bars.gif)
-
-### MNIST + Shape
-![MNIST + Shape animation](animations/mnist_shape.gif)
-
-### Simple Superposition
-![Simple Superposition animation](animations/simple_superpos.gif)
-
 ## Dependencies and Setup
 
 * brainstorm == 0.5
@@ -59,7 +42,7 @@ To get the best networks we used in the paper for each dataset run `run_best_net
 It will save a `Networks/DATASET_best_dae.h5` network for each dataset. This shouldn't take more than half an hour.
 
 Alternatively you can use your own best results from the random search by running the 
-last cell in the `Get_Search_Results.ipynb`.
+corresponding cells in the `Get_Search_Results.ipynb`.
 
 These files are needed for the following steps.
 
@@ -68,10 +51,62 @@ Next we use these networks for Reconstruction Clustering and store all of the re
 
     run_evaluation.py
 
-NOTE: This should take about an hour and use about 26 GBytes of disk space.
+NOTE: This should take about an hour and use ca 21 GBytes of disk space.
 
 ## Plots
 The `Plots.ipynb` notebook generates all the figures used in the paper.
 Once you've run all the other steps you should be able to able to generate them
 all yourself.
 
+## Database Dump
+With the file ``dump.zip`` we've included a dump of the MongoDB that contains all the information about
+all the experimental runs we did.
+
+
+
+## Demo Images
+
+### Regular (soft) Reconstruction Clustering 
+These animations show the cluster assignment during a run of RC on 120 
+different test images. 
+For each dataset we used the best DAE trained on single object images.
+
+#### Shapes
+![shapes animation](animations/shapes.gif)
+
+#### Multi MNIST
+![Multi-MNIST animation](animations/multi_mnist.gif)
+
+#### Corners
+![Corners animation](animations/corners.gif)
+
+#### Bars
+![Bars animation](animations/bars.gif)
+
+#### MNIST + Shape
+![MNIST + Shape animation](animations/mnist_shape.gif)
+
+#### Simple Superposition
+![Simple Superposition animation](animations/simple_superpos.gif)
+
+
+### Hard Reconstruction Clustering
+These animations show the hard cluster assignment during a run of RC on 120 
+different test images. 
+To improve visibility we toned down the brightness on the background pixels. 
+Here we used the best DAE trained on **multi object** images.
+
+#### Shapes
+![shapes animation](animations/shapes_train_multi.gif)
+
+#### Multi MNIST
+![Multi-MNIST animation](animations/multi_mnist_train_multi.gif)
+
+#### Corners
+![Corners animation](animations/corners_train_multi.gif)
+
+#### Bars
+![Bars animation](animations/bars_train_multi.gif)
+
+#### MNIST + Shape
+![MNIST + Shape animation](animations/mnist_shape_train_multi.gif)

Get Results.ipynb

@@ -16,16 +16,16 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": 5,
    "metadata": {
     "collapsed": true
    },
    "outputs": [],
    "source": [
-    "cluster_nodes = ['your', 'dispy', 'cluster', 'nodes']\n",
+    "cluster_nodes = ['put', 'your', 'nodes', 'here']\n",
     "\n",
     "mongo_db = {\n",
-    "    'url': 'INSERT_MONGO_DB_IP_HERE:27017',\n",
+    "    'url': 'DATABASE_IP_HERE:27017',\n",
     "    'db': 'binding_via_rc'\n",
     "}"
    ]
@@ -41,8 +41,8 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "2016-01-15 19:47:31,652 - dispy - Storing fault recovery information in \"_dispy_20160115194731\"\n",
-      "INFO:dispy:Storing fault recovery information in \"_dispy_20160115194731\"\n"
+      "2016-01-16 17:57:08,627 - dispy - Storing fault recovery information in \"_dispy_20160116175708\"\n",
+      "INFO:dispy:Storing fault recovery information in \"_dispy_20160116175708\"\n"
      ]
     }
    ],
@@ -60,7 +60,7 @@
     {
      "data": {
       "text/plain": [
-       "ClusterStatus(nodes=[<dispy.DispyNode object at 0x7f65016af908>], jobs_pending=0)"
+       "ClusterStatus(nodes=[<dispy.DispyNode object at 0x7fb865659d30>], jobs_pending=0)"
       ]
      },
      "execution_count": 4,

Networks/best_bars_dae.h5

@@ -3,18 +3,17 @@
 from __future__ import division, print_function, unicode_literals
 from dae import ex
 
-
 @ex.named_config
 def best_bars():
     dataset = {
         'name': 'bars',
-        'salt_n_pepper': 0.1
+        'salt_n_pepper': 0.0
     }
     training = {
-        'learning_rate': 0.0788887811150988
+        'learning_rate': 0.768014586935404
     }
-    seed = 915841892
-    network_spec = "Ft1000"
+    seed = 459182787
+    network_spec = "Fr100"
     net_filename = 'Networks/best_bars_dae.h5'
 
 ex.run(named_configs=['best_bars'])
@@ -24,13 +23,13 @@ def best_bars():
 def best_corners():
     dataset = {
         'name': 'corners',
-        'salt_n_pepper': 0.2
+        'salt_n_pepper': 0.0
     }
     training = {
-        'learning_rate': 0.11338088391400022
+        'learning_rate': 0.0019199822609484764
     }
-    seed = 820141270
-    network_spec = "Fs500"
+    seed = 158253144
+    network_spec = "Fr100"
     net_filename = 'Networks/best_corners_dae.h5'
 
 ex.run(named_configs=['best_corners'])
@@ -40,13 +39,13 @@ def best_corners():
 def best_shapes():
     dataset = {
         'name': 'shapes',
-        'salt_n_pepper': 0.3
+        'salt_n_pepper': 0.4
     }
     training = {
-        'learning_rate': 0.06701641168189125
+        'learning_rate': 0.08314720669724956
     }
-    seed = 533867354
-    network_spec = "Ft1000"
+    seed = 845841083
+    network_spec = "Ft500"
     net_filename = 'Networks/best_shapes_dae.h5'
 
 ex.run(named_configs=['best_shapes'])
@@ -56,13 +55,13 @@ def best_shapes():
 def best_multi_mnist():
     dataset = {
         'name': 'multi_mnist',
-        'salt_n_pepper': 0.8
+        'salt_n_pepper': 0.6
     }
     training = {
-        'learning_rate': 0.004194304163542453
+        'learning_rate': 0.011361917579645924
     }
-    seed = 708383804
-    network_spec = "Fr500"
+    seed = 498470020
+    network_spec = "Fr1000"
     net_filename = 'Networks/best_multi_mnist_dae.h5'
 
 ex.run(named_configs=['best_multi_mnist'])
@@ -72,13 +71,13 @@ def best_multi_mnist():
 def best_mnist_shape():
     dataset = {
         'name': 'mnist_shape',
-        'salt_n_pepper': 0.5
+        'salt_n_pepper': 0.6
     }
     training = {
-        'learning_rate': 0.011252506237215505
+        'learning_rate': 0.0316848152096582
     }
-    seed = 17521051
-    network_spec = "Fr250"
+    seed = 166717815
+    network_spec = "Fs250"
     net_filename = 'Networks/best_mnist_shape_dae.h5'
 
 ex.run(named_configs=['best_mnist_shape'])
@@ -91,10 +90,10 @@ def best_simple_superpos():
         'salt_n_pepper': 0.1
     }
     training = {
-        'learning_rate': 0.20713360779738232
+        'learning_rate': 0.36662702472680564
     }
-    seed = 563758549
-    network_spec = "Fr500"
+    seed = 848588405
+    network_spec = "Fr100"
     net_filename = 'Networks/best_simple_superpos_dae.h5'
 
 ex.run(named_configs=['best_simple_superpos'])
@@ -105,13 +104,13 @@ def best_bars_train_multi():
     dataset = {
         'name': 'bars',
         'train_set': 'train_multi',
-        'salt_n_pepper': 0.9
+        'salt_n_pepper': 0.8
     }
     training = {
-        'learning_rate': 0.07009646428972773
+        'learning_rate': 0.01219213699462807
     }
-    seed = 100618159
-    network_spec = "Ft100"
+    seed = 141786426
+    network_spec = "Fs100"
     net_filename = 'Networks/best_bars_dae_train_multi.h5'
 
 ex.run(named_configs=['best_bars_train_multi'])
@@ -122,13 +121,13 @@ def best_corners_train_multi():
     dataset = {
         'name': 'corners',
         'train_set': 'train_multi',
-        'salt_n_pepper': 0.5
+        'salt_n_pepper': 0.7
     }
     training = {
-        'learning_rate': 0.07329485079040306
+        'learning_rate': 0.02603487482829947
     }
-    seed = 242923369
-    network_spec = "Fr500"
+    seed = 872544498
+    network_spec = "Fr100"
     net_filename = 'Networks/best_corners_dae_train_multi.h5'
 
 ex.run(named_configs=['best_corners_train_multi'])
@@ -142,10 +141,10 @@ def best_shapes_train_multi():
         'salt_n_pepper': 0.9
     }
     training = {
-        'learning_rate': 0.023890872422966537
+        'learning_rate': 0.049401835193689486
     }
-    seed = 900380945
-    network_spec = "Ft500"
+    seed = 702200962
+    network_spec = "Fs100"
     net_filename = 'Networks/best_shapes_dae_train_multi.h5'
 
 ex.run(named_configs=['best_shapes_train_multi'])
@@ -159,9 +158,9 @@ def best_multi_mnist_train_multi():
         'salt_n_pepper': 0.9
     }
     training = {
-        'learning_rate': 0.0040886980876061
+        'learning_rate': 0.001785591525476118
     }
-    seed = 66747176
+    seed = 632224571
     network_spec = "Fs250"
     net_filename = 'Networks/best_multi_mnist_dae_train_multi.h5'
 
@@ -176,11 +175,12 @@ def best_mnist_shape_train_multi():
         'salt_n_pepper': 0.6
     }
     training = {
-        'learning_rate': 0.014119758147665773
+        'learning_rate': 0.033199614969711265
     }
-    seed = 767374347
+    seed = 900543563
     network_spec = "Fr1000"
     net_filename = 'Networks/best_mnist_shape_dae_train_multi.h5'
 
 ex.run(named_configs=['best_mnist_shape_train_multi'])
 
+