bug fix with dataset.py

yann/special/datasets.py

@@ -1,7 +1,6 @@
 from yann.utils.dataset import setup_dataset
 
-def cook_mnist_normalized(  verbose = 1,
-				 **kwargs):
+def cook_mnist_normalized(  verbose = 1, **kwargs):
 	"""
 	Wrapper to cook mnist dataset. Will take as input,
 
@@ -40,7 +39,7 @@ def cook_mnist_normalized(  verbose = 1,
                             "GCN"           : False,
                             "ZCA"           : False,
                             "grayscale"     : False,
-							"mean_subtract" : False,
+							"zero_mean" 	: False,
                         }
 	else:
 		preprocess_params = kwargs['preprocess_params']
@@ -56,8 +55,7 @@ def cook_mnist_normalized(  verbose = 1,
 							verbose = 3)
 	return dataset
 
-def cook_mnist_normalized_mean_subtracted(  verbose = 1,
-				 **kwargs):
+def cook_mnist_normalized_zero_mean(  verbose = 1,	**kwargs):
 	"""
 	Wrapper to cook mnist dataset. Will take as input,
 
@@ -93,7 +91,7 @@ def cook_mnist_normalized_mean_subtracted(  verbose = 1,
                             "GCN"           : False,
                             "ZCA"           : False,
                             "grayscale"     : False,
-							"mean_subtract" : True,
+							"zero_mean"	 	: True,
                         }
 	else:
 		preprocess_params = kwargs['preprocess_params']
@@ -149,7 +147,7 @@ def cook_mnist_multi_load(  verbose = 1, **kwargs):
                             "GCN"           : False,
                             "ZCA"           : False,
                             "grayscale"     : False,
-                            "mean_subtract" : True,
+                            "zero_mean" 	: False,
                         }
 	else:
 		preprocess_params = kwargs['preprocess_params']
@@ -165,7 +163,7 @@ def cook_mnist_multi_load(  verbose = 1, **kwargs):
 							verbose = 3)
 	return dataset	
 
-def cook_cifar10_normalized_mean_subtracted(verbose = 1, **kwargs):
+def cook_cifar10_normalized(verbose = 1, **kwargs):
 	"""
 	Wrapper to cook cifar10 dataset. Will take as input,
 
@@ -201,7 +199,7 @@ def cook_cifar10_normalized_mean_subtracted(verbose = 1, **kwargs):
                             "GCN"           : False,
                             "ZCA"           : False,
                             "grayscale"     : False,
-                            "mean_subtract" : True,
+                            "mean_subtract" : False,
                         }
 	else:
 		preprocess_params = kwargs['preprocess_params']
@@ -219,8 +217,8 @@ def cook_cifar10_normalized_mean_subtracted(verbose = 1, **kwargs):
 
 
 # Just some wrappers
-cook_mnist = cook_mnist_normalized_mean_subtracted
-cook_cifar10 = cook_cifar10_normalized_mean_subtracted
+cook_mnist = cook_mnist_normalized
+cook_cifar10 = cook_cifar10_normalized
 
 if __name__ == '__main__':
 	pass

yann/utils/dataset.py

@@ -693,7 +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. 
+					"zero_mean" : Subtracts the mean of the image. 
 
 						} 
 
@@ -705,7 +705,7 @@ def preprocessing( data, height, width, channels, args):
 	GCN 			= args [ "GCN" ]      
 	ZCA 			= args [ "ZCA" ]	 
 	gray 			= args [ "grayscale" ]
-	mean_subtract   = args [ "mean_subtract" ]
+	zero_mean	    = args [ "zero_mean" ]
 
 	# Assume that the data is already resized on height and width and all ... 
 	if len(data.shape) == 2 and channels > 1: 	
@@ -731,14 +731,12 @@ def preprocessing( data, height, width, channels, args):
 
 	# from here on data is processed as a 2D matrix
 	data = numpy.reshape(data,out_shp)
-	if mean_subtract is True:
-		if normalize is True or ZCA is True: 
-			data = data  / (data.max() + 1e-7)
-			data = data - data.mean()
-			# do this normalization thing in batch mode.	
-	else:
-    		if normalize is True or ZCA is True:
-			data = data / (data.max() + 1e-7)
+
+	if normalize is True or ZCA is True:
+		data = data / (data.max() + 1e-7)
+
+	if zero_mean is True:  # This will make the data go from (-1,1)
+		data = ( data - 0.5 ) * 2
 
 	if ZCA is True: