Refactored

ExperimentExecution.py

@@ -33,7 +33,7 @@
     training_patterns_heterogeneous.append([uppercase_letter, lowercase_letter])
 
 turnover_rate = 0.50
-weighting_dg = current_dg_weighting % 25
+weighting_dg = current_dg_weighting % 26
 # print "TRIAL #", trial, "turnover rate:", turnover_rate
 # dims,
 # connection_rate_input_ec, perforant_path, mossy_fibers,
@@ -45,24 +45,23 @@
           0.7, 100.0, 0.1, turnover_rate,  # gamma, epsilon, nu, turnover rate
           0.10, 0.95, 0.8, 2.0, weighting_dg)  # k_m, k_r, a_i, alpha. alpha is 2 in 4.1
 
-for three_trials in range(3):
-    hpc.reset_hpc_module()
-    for train_set_size_ctr in range(3, 4):
+# hpc.reset_hpc_module()
+for train_set_size_ctr in range(3, 4):
 
-        hipp_chaotic_pats, _ = experiment_4_x_1(hpc, train_set_size_ctr, training_patterns_associative)
-        Tools.save_experiment_4_1_results(hpc, hipp_chaotic_pats, "train_set_size_"+str(train_set_size_ctr)+"_exp_1"+
-                                          "turnover_rate_" + str(turnover_rate) +
-                                          "weighting_" + str(hpc._weighting_dg))
+    hipp_chaotic_pats, _ = experiment_4_x_1(hpc, train_set_size_ctr, training_patterns_associative)
+    Tools.save_experiment_4_1_results(hpc, hipp_chaotic_pats, "train_set_size_"+str(train_set_size_ctr)+"_exp_1"+
+                                      "turnover_rate_" + str(turnover_rate) +
+                                      "weighting_" + str(hpc._weighting_dg))
 
-        # ann = SimpleNeocorticalNetwork(io_dim, 30, io_dim, 0.01, 0.9)
+    # ann = SimpleNeocorticalNetwork(io_dim, 30, io_dim, 0.01, 0.9)
 
-        # print "Starting experiment 4_2..."
-        # This also saves the experiment_4_x_1 results!
-        # information_vector = experiment_4_x_2(hpc, ann, train_set_size_ctr,
-        #                                       training_patterns_associative[:5 * train_set_size_ctr])
-        # print "Saving the results."
-        # Tools.save_experiment_4_2_results(information_vector, "train_set_size_" + str(train_set_size_ctr) +
-        #                                   "_exp_2_")
+    # print "Starting experiment 4_2..."
+    # This also saves the experiment_4_x_1 results!
+    # information_vector = experiment_4_x_2(hpc, ann, train_set_size_ctr,
+    #                                       training_patterns_associative[:5 * train_set_size_ctr])
+    # print "Saving the results."
+    # Tools.save_experiment_4_2_results(information_vector, "train_set_size_" + str(train_set_size_ctr) +
+    #                                   "_exp_2_")
 
-        # For now, this is the ONLY place where the counter is incremented.
-        Tools.increment_experiment_counter()
+    # For now, this is the ONLY place where the counter is incremented.
+    Tools.increment_experiment_counter()

Experiments_4_x.py

@@ -37,7 +37,7 @@ def training_and_recall_hpc_helper(hpc, training_set_size, train_set_num, origin
     # print "Neuronal turnover completed in", "{:7.3f}".format(t1-t0), "seconds."
     # hpc.re_wire_fixed_input_to_ec_weights()
     print "Learning patterns in training set..."
-    hpc_learn_patterns_wrapper(hpc, patterns=training_set, max_training_iterations=30)  # when training is fixed,
+    hpc_learn_patterns_wrapper(hpc, patterns=training_set, max_training_iterations=50)  # when training is fixed,
     # convergence should occur after one or two iterations?
 
     # extract by chaotic recall:

HPC.py

@@ -71,28 +71,28 @@ def __init__(self, dims, connection_rate_input_ec, perforant_path, mossy_fibers,
                                            borrow=True)
 
         # randomly assign about 25 % of the weights to a random connection weight
-        ec_dg_weights = Tools.binomial_f(dims[1], dims[2], self.PP) * np.random.normal(0.5, np.sqrt(0.25), (dims[1], dims[2]))
+        ec_dg_weights = Tools.binomial_f(dims[1], dims[2], self.PP) * np.random.normal(0.5, np.sqrt(0.25), (dims[1], dims[2])).astype(np.float32)
         self.ec_dg_weights = theano.shared(name='ec_dg_weights', value=ec_dg_weights.astype(theano.config.floatX),
                                            borrow=True)
 
         # randomly assign all weights between the EC and CA3
-        ec_ca3_weights = np.random.normal(0.5, np.sqrt(0.25), (dims[1], dims[3]))# * Tools.binomial_f(dims[1], dims[3], self.PP)
+        ec_ca3_weights = np.random.normal(0.5, np.sqrt(0.25), (dims[1], dims[3])).astype(np.float32)# * Tools.binomial_f(dims[1], dims[3], self.PP)
         self.ec_ca3_weights = theano.shared(name='ec_ca3_weights', value=ec_ca3_weights.astype(theano.config.floatX),
                                             borrow=True)
 
         # randomly assign about 4 % of the weights to random connection weights
         dg_ca3_weights = Tools.binomial_f(dims[2], dims[3], self.MF) * \
-                         np.random.normal(0.9, np.sqrt(0.01), (dims[2], dims[3]))  # elemwise
+                         np.random.normal(0.9, np.sqrt(0.01), (dims[2], dims[3])).astype(np.float32)  # elemwise
         self.dg_ca3_weights = theano.shared(name='dg_ca3_weights', value=dg_ca3_weights.astype(theano.config.floatX),
                                             borrow=True)
 
         # randomly assign 100 % of the weights between CA3 and CA3
-        ca3_ca3_weights = np.random.normal(0.5, np.sqrt(0.25), (dims[3], dims[3]))
+        ca3_ca3_weights = np.random.normal(0.5, np.sqrt(0.25), (dims[3], dims[3])).astype(np.float32)
         self.ca3_ca3_weights = theano.shared(name='ca3_ca3_weights', value=ca3_ca3_weights.astype(theano.config.floatX),
                                              borrow=True)
 
         # random weight assignment, full connection rate CA3-out
-        ca3_output_weights = np.random.normal(0., np.sqrt(0.5), (dims[3], dims[4]))
+        ca3_output_weights = np.random.normal(0., np.sqrt(0.5), (dims[3], dims[4])).astype(np.float32)
         self.ca3_out_weights = theano.shared(name='ca3_out_weights',
                                              value=ca3_output_weights.astype(theano.config.floatX), borrow=True)
 

Tools.py

@@ -114,15 +114,15 @@ def get_pattern_correlation_slow(pattern_1, pattern_2):
 def save_experiment_4_1_results(hpc, chaotically_recalled_patterns, custom_name):
     experiment_dir = get_experiment_dir()
 
-    hpc_f = file(experiment_dir+'/hpc_'+custom_name+'.save', 'wb')
-    cPickle.dump(hpc, hpc_f, protocol=cPickle.HIGHEST_PROTOCOL)
-    hpc_f.close()
+    # hpc_f = file(experiment_dir+'/hpc_'+custom_name+'.save', 'wb')
+    # cPickle.dump(hpc, hpc_f, protocol=cPickle.HIGHEST_PROTOCOL)
+    # hpc_f.close()
 
-    save_images_from(chaotically_recalled_patterns, experiment_dir+'/images')
+    save_images_from(chaotically_recalled_patterns, experiment_dir+'/images'+custom_name)
 
-    f2 = file(experiment_dir+'/_chaotically_recalled_patterns.save', 'wb')
-    cPickle.dump(chaotically_recalled_patterns, f2, protocol=cPickle.HIGHEST_PROTOCOL)
-    f2.close()
+    # f2 = file(experiment_dir+'/_chaotically_recalled_patterns.save', 'wb')
+    # cPickle.dump(chaotically_recalled_patterns, f2, protocol=cPickle.HIGHEST_PROTOCOL)
+    # f2.close()
 
 
 def save_experiment_4_2_results(information_vector, custom_name):