GAN bug fix

pantry/tutorials/gan.py

@@ -34,7 +34,7 @@ def simple_gan ( dataset= None, verbose = 1 ):
     visualizer_params = {
                     "root"       : '.',
                     "frequency"  : 1,
-                    "sample_size": 144,
+                    "sample_size": 225,
                     "rgb_filters": False,
                     "debug_functions" : False,
                     "debug_layers": True,  
@@ -85,7 +85,7 @@ def simple_gan ( dataset= None, verbose = 1 ):
     net.add_layer ( type = "dot_product",
                     id = "D(x)",
                     origin = "x",
-                    num_neurons = 512,
+                    num_neurons = 1024,
                     dropout_rate = 0,                    
                     activation = ('maxout','maxout',2),
                     verbose = verbose
@@ -97,7 +97,7 @@ def simple_gan ( dataset= None, verbose = 1 ):
                     id = "D(G(z))",
                     origin = "G(z)",
                     dropout_rate = 0,                    
-                    num_neurons = 512,
+                    num_neurons = 1024,
                     input_params = net.dropout_layers["D(x)"].params, # must be the same params, 
                                                         # this way it remains the same network.
                     activation = ('maxout','maxout',2),
@@ -168,6 +168,7 @@ def simple_gan ( dataset= None, verbose = 1 ):
     from yann.utils.graph import draw_network
     draw_network(net.graph, filename = 'gan.png')    
     net.pretty_print()
+
     net.cook (  objective_layers = ["classifier_obj","real_obj","fake_obj"],
                 optimizer_params = optimizer_params,
                 generator_layers = ["G(z)"], 
@@ -177,8 +178,9 @@ def simple_gan ( dataset= None, verbose = 1 ):
                 verbose = verbose )
 
     learning_rates = (0.05, 0.001, 0.0001)  
+
     net.train( epochs = (50, 50), 
-               k = 1,  # refer to Ian Goodfellow's paper Algorithm 1.
+               k = 5,  # refer to Ian Goodfellow's paper Algorithm 1.
                validate_after_epochs = 1,
                training_accuracy = True,
                show_progress = True,

yann/utils/dataset.py

@@ -693,6 +693,7 @@ def preprocessing( data, height, width, channels, args):
 					"GCN"	: True for global contrast normalization
 					"ZCA"	: True, kind of like a PCA representation (not fully tested)
 					"grayscale"  : Convert the image to grayscale
+					"mean_subtract" : Subtracts the mean of the image. 
 
 						} 
 
@@ -704,6 +705,7 @@ def preprocessing( data, height, width, channels, args):
 	GCN 			= args [ "GCN" ]      
 	ZCA 			= args [ "ZCA" ]	 
 	gray 			= args [ "grayscale" ]
+	mean_subtract   = args [ "mean_subtract" ]
 
 	# Assume that the data is already resized on height and width and all ... 
 	if len(data.shape) == 2 and channels > 1: 	
@@ -731,7 +733,8 @@ def preprocessing( data, height, width, channels, args):
 	data = numpy.reshape(data,out_shp)
 	if normalize is True or ZCA is True: 
 		data = data / (data.max(axis = 0) + 1e-7)
-		# do this normalization thing in batch mode.		
+		# do this normalization thing in batch mode.	
+	if mean_subtract is True or ZCA is True:    			
 		data = data - data.mean(axis = 0)		
 
 	if ZCA is True:		
@@ -1195,10 +1198,11 @@ def cook_mnist(  verbose = 1,
 
     # parameters relating to preprocessing.
 		preprocess_params = { 
-                            "normalize"     : False,
+                            "normalize"     : True,
                             "GCN"           : False,
                             "ZCA"           : False,
                             "grayscale"     : False,
+							"mean_subtract" : False,
                         }
 	else:
 		preprocess_params = kwargs['preprocess_params']