[opengm] 192/386: -impmeneted batch adding of python learnable functions, NICE AS FUCK! - removed pure 'python-function' from typelist - improved const correctness

[Ghislain Vaillant]

This is an automated email from the git hooks/post-receive script.

ghisvail-guest pushed a commit to branch debian/master
in repository opengm.

commit 4caa21cf7b753ac1edc93ffb90060a39f6d79b07
Author: DerThorsten <thorsten.beier at iwr.uni-heidelberg.de>
Date:   Mon Dec 22 13:09:14 2014 +0100

    -impmeneted batch adding of python learnable functions, NICE AS FUCK!
    - removed pure 'python-function' from typelist
    - improved const correctness
---
 fubar/pascal_voc/convertPascalVOCNew.py | 107 ++++++++++++++++++++++++++++++++
 1 file changed, 107 insertions(+)

diff --git a/fubar/pascal_voc/convertPascalVOCNew.py b/fubar/pascal_voc/convertPascalVOCNew.py
new file mode 100644
index 0000000..8c7315f
--- /dev/null
+++ b/fubar/pascal_voc/convertPascalVOCNew.py
@@ -0,0 +1,107 @@
+import cPickle as pickle
+import numpy as np
+import opengm
+from opengm import learning
+import h5py
+
+# download the Pascal VOC datasets from
+# 
+# http://www.ais.uni-bonn.de/deep_learning/downloads.html
+# 
+
+# converter from pystruct EdgeFeatureGraphCRF to opengm learnable
+fns = ['./data_train.pickle', './data_val.pickle']
+ds_suffixes = ['_train']#, '_val']
+ogm_dss = [None, None]
+ww = [None, None]
+num_samples = None
+out_dir = './'
+out_prefix = 'pascal_voc'
+
+num_labels = 21
+
+# get label weights from training set:
+# loss should be weighted inversely by the number of 
+# occurrences of each class in the training set    
+Y = pickle.load(open(fns[0], 'r'))['Y'][:num_samples]
+Y = np.hstack(Y)
+Y[Y==-1]=0 # FIXME: introduce a void label, so long: make the void label background 
+label_weights = 1. / np.bincount(Y)
+label_weights= np.ones(27,dtype=opengm.value_type)
+label_weights[np.isinf(label_weights)] = 0
+label_weights *= 1. / np.sum(label_weights)
+
+for ii, (fn, ds_suffix) in enumerate(zip(fns, ds_suffixes)):
+    ds = pickle.load(open(fn, 'r'))
+
+    # X is a list of instances of a dataset where (for EdgeFeatureGraphCRF)
+    # each instance is a tuple of (unary_feats, edges, edge_feats)
+    X = ds['X'][:num_samples]
+
+    # the ground truth labels
+    Y = ds['Y'][:num_samples]
+
+    # superpixels (for reference)
+    #superpixels_train = ds['superpixels'][:num_samples]
+
+    # filenames (for reference)
+    #fns_train = ds['file_names'][:num_samples]
+
+    num_edge_feats = X[0][2].shape[1]
+    num_unary_feats = num_labels * X[0][0].shape[1]
+    num_weights = num_unary_feats + num_edge_feats
+    # create and initialize weights
+    print 'num_weights =', num_weights
+    print 'num_instances =', len(X)
+
+    ogm_dss[ii] = learning.createDataset(num_weights, numInstances=len(X), loss="generalized-hamming")
+    #ogm_ds = ogm_dss[ii]
+    ww[ii] = ogm_dss[ii].getWeights()
+
+    for idx, (x, y) in enumerate(zip(X, Y)):
+        y[y==-1]=0  # FIXME: introduce a void label, so long: make the void label background 
+        unary_feats, edges, edge_feats = x
+        num_vars = unary_feats.shape[0]
+
+        states = np.ones(num_vars, dtype=opengm.label_type) * num_labels
+        
+        gm = opengm.gm(states, operator='adder')
+
+        lossParam = learning.GeneralizedHammingLossParameter()
+        lossParam.setLabelLossMultiplier(label_weights)
+
+        # add unary factors
+        weight_ids = np.arange(0, num_labels * unary_feats.shape[1]).reshape((num_labels, -1))
+
+            
+            
+        # the features are different for each function instance
+        # but a single feature vector is shared between all
+        # labels for one particular instance.
+        # The weights the same for all function instances
+        # but each label has a particular set of weights
+        lUnaries = learning.lUnaryFunctions(weights = ww[ii],numberOfLabels = num_labels, 
+                                            features=unary_feats,weightIds = weight_ids,
+                                            featurePolicy= learning.FeaturePolicy.sharedBetweenLabels)
+        gm.addFactors(gm.addFunctions(lUnaries), np.arange(num_vars)) 
+
+
+
+        # add all pairwise factors at once
+        weight_ids = np.arange(num_unary_feats, num_unary_feats+num_edge_feats)
+        lp = learning.lPottsFunctions(weights=ww[ii], numberOfLabels=num_labels,
+                                     features=edge_feats, weightIds=weight_ids)
+        gm.addFactors(gm.addFunctions(lp), edges) 
+
+
+        # add the model to the dataset
+        ogm_dss[ii].setInstanceWithLossParam(idx, gm, y.astype(dtype=opengm.label_type), lossParam)
+
+
+    ogm_dss[ii].save(out_dir, out_prefix + ds_suffix + '_')
+
+
+
+
+
+

-- 
Alioth's /usr/local/bin/git-commit-notice on /srv/git.debian.org/git/debian-science/packages/opengm.git