stage 1 covariance matrices are similar when normalized

compared my cov1 to og matlab cov1, where each was divided by its max
element so all values are from 0 to 1. Max difference was 1e-2
Max difference when compares to randn() was 1.5, so I think I got
covariance of patches right. Not sure why the scale is different. Might
be because of batch size

dataset_utils.py

@@ -52,7 +52,7 @@ def load(name):
 
     info = datasets[name]
 
-    batch_size = 500
+    batch_size = 100
 
     with tf.name_scope('input'):
         filename_queue = tf.train.string_input_producer([info.TRAIN_RECORD_PATH], name='train_queue')

pcanet.py

@@ -202,6 +202,10 @@ def main():
     train_pcanet_features, train_labels, summary = sess.run([m.output_features, train_label_batch, merged_summary])
     writer.add_summary(summary, 0)
 
+    cov = sess.run(m.patches_covariance1)
+    np.savetxt('cov.csv', np.squeeze(cov))
+    exit(0)
+
     # train linear SVM
     svm = LinearSVC(C=1, fit_intercept=False)
     svm.fit(train_pcanet_features, train_labels)