DL updates.

Daniel McDuff · Daniel McDuff · commit 7b6b66e16f13 · 2015-02-05T11:34:31.000-05:00
diff --git a/code/DBN_face.py b/code/DBN_face.py
@@ -342,7 +342,7 @@ def test_DBN(finetune_lr=0.1, pretraining_epochs=3,
     if not os.path.isdir(output_dir+'/dbn_plots'):
             os.makedirs(output_dir+'/dbn_plots')
     
-    # compute number of minibatches for training, validation and testing
+    # compute number of minibatches for pretraining:
     n_train_batches = pre_train_set_x.get_value(borrow=True).shape[0] / batch_size
 
     # numpy random generator
@@ -409,6 +409,9 @@ def test_DBN(finetune_lr=0.1, pretraining_epochs=3,
     # FINETUNING THE MODEL #
     ########################
 
+    # compute number of minibatches for training, validation and testing:
+    n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
+
     # get the training, validation and testing function for the model
     print '... getting the finetuning functions'
     train_fn, validate_model, test_model = dbn.build_finetune_functions(
diff --git a/code/load_faces.py b/code/load_faces.py
@@ -16,6 +16,8 @@
 import theano.tensor as T
 from theano.tensor.shared_randomstreams import RandomStreams
 import pandas
+import matplotlib.pyplot as plt
+import matplotlib.cm as cm
 
 def logistic_transform(A, mu, sigma):
     A[numpy.where(A == 0)] = 0.1
@@ -63,14 +65,21 @@ def import_data(label, data_dir, image_dim):
             #if neutral.iloc[0]:
             if test_target>=-1:
                 test_image = numpy.array(scipy.misc.imread(f))
+
                 if (len(test_image.flatten())!=(image_dim*image_dim)):
                     continue
                 #for i, row in enumerate(test_targets.iloc[0].values):
                 #    print i + str(test_targets.iloc[0][i])
 
+                #test_image2 = test_image.astype(float)
+                #temp = logistic_transform(test_image2.flatten(), 140, 0.05)
+                #plt.imshow(temp.reshape(image_dim,image_dim), cmap = cm.Greys_r)
+                #plt.show()
+                #plt.show(block=False)
+
                 if test_target > 50:
                         test_image2 = test_image.astype(float)
-                        temp = logistic_transform(test_image2.flatten(), 120, 0.1)
+                        temp = logistic_transform(test_image2.flatten(), 140, 0.05)
                         if numpy.isnan(temp).any():
                             print "NaN found :("
                             continue
@@ -79,7 +88,7 @@ def import_data(label, data_dir, image_dim):
                             target = numpy.append(target, [1], axis=0)
                 elif test_target == 0:
                         test_image2 = test_image.astype(float)
-                        temp = logistic_transform(test_image2.flatten(), 120, 0.1)
+                        temp = logistic_transform(test_image2.flatten(), 140, 0.05)
                         if numpy.isnan(temp).any():
                             print "NaN found :("
                             continue
@@ -88,7 +97,7 @@ def import_data(label, data_dir, image_dim):
                             target = numpy.append(target, [0], axis=0)
                 elif test_target == -1:
                         test_image2 = test_image.astype(float)
-                        temp = logistic_transform(test_image2.flatten(), 120, 0.1)
+                        temp = logistic_transform(test_image2.flatten(), 140, 0.05)
                         if numpy.isnan(temp).any():
                             print "NaN found :("
                             continue
diff --git a/code/rbm.py b/code/rbm.py
@@ -151,12 +151,16 @@ def sample_h_given_v(self, v0_sample):
         # the visibles
         pre_sigmoid_h1, h1_mean = self.propup(v0_sample)
 
+
+        # LARGER mu IS MORE SPARSE.
+        mu = 0.000001 # mu = 0.01 is probably too small.
+        # LOOKED AT THE CODE HERE: http://lrn2cre8.ofai.at/lrn2/doc/_modules/lrn2/models/srbm_goh.html#SRBM_Goh
         ## DAN ADDED:#########################
         rank_0 = ((h1_mean.argsort(axis=0)).argsort(axis=0).astype(theano.config.floatX) + 1.)/T.shape(h1_mean)[0].astype(theano.config.floatX)
 
         rank_1 = ((h1_mean.argsort(axis=1)).argsort(axis=1).astype(theano.config.floatX) + 1.)/T.shape(h1_mean)[1].astype(theano.config.floatX)
 
-        h1_mean = (1.-0.5)*(rank_0**((1./0.05)-1.))+0.5*(rank_1**((1./0.05)-1.))
+        h1_mean = (1.-0.5)*(rank_0**((1./mu)-1.))+0.5*(rank_1**((1./mu)-1.))
 
         #pre_sigmoid_h1_bin = T.log(h1_mean) - T.log(1. - h1_mean)
         #pre_sigmoid_h1 = pre_sigmoid_h1_bin 
