Changes to make gan_pytorch.py use only squared diffs rather than squared diffs + original data

gistfile1.txt

diff --git a/gan_pytorch.py b/gan_pytorch.py
index 0bff38c..802d6cb 100755
--- a/gan_pytorch.py
+++ b/gan_pytorch.py
@@ -31,7 +31,7 @@ g_steps = 1
 
 # ### Uncomment only one of these
 #(name, preprocess, d_input_func) = ("Raw data", lambda data: data, lambda x: x)
-(name, preprocess, d_input_func) = ("Data and variances", lambda data: decorate_with_diffs(data, 2.0), lambda x: x * 2)
+(name, preprocess, d_input_func) = ("Data and variances", lambda data: decorate_with_diffs(data, 2.0), lambda x: x)
 
 print "Using data [%s]" % (name)
 
@@ -79,7 +79,8 @@ def decorate_with_diffs(data, exponent):
     mean = torch.mean(data.data, 1)
     mean_broadcast = torch.mul(torch.ones(data.size()), mean.tolist()[0][0])
     diffs = torch.pow(data - Variable(mean_broadcast), exponent)
-    return torch.cat([data, diffs], 1)
+    #return torch.cat([data, diffs], 1)
+    return diffs
 
 d_sampler = get_distribution_sampler(data_mean, data_stddev)
 gi_sampler = get_generator_input_sampler()

Hi， i have the same question about the function of decorate_with_diffs. I didn't understand why you cat the data and diffs. I have run the scrip with using these two methods .After training, i found without cating the data and diffs, the fake data has more negative mean values than the first method. Can you explain it? Thanks very much for helping me.

My own understanding about it is using both the data and diffs， the G can generate much more better fake data than only using diffs. It just looks like the features extracting from images, more dimensions of feature, more accuracy of the nueral network.