Added --save_test_mistakes.
authorFrancois Fleuret <francois@fleuret.org>
Mon, 26 Jun 2017 13:40:52 +0000 (15:40 +0200)
committerFrancois Fleuret <francois@fleuret.org>
Mon, 26 Jun 2017 13:40:52 +0000 (15:40 +0200)
cnn-svrt.py

index ade87ce..3fe50d8 100755 (executable)
@@ -85,6 +85,9 @@ parser.add_argument('--compress_vignettes',
                     type = distutils.util.strtobool, default = 'True',
                     help = 'Use lossless compression to reduce the memory footprint')
 
+parser.add_argument('--save_test_mistakes',
+                    type = distutils.util.strtobool, default = 'False')
+
 parser.add_argument('--model',
                     type = str, default = 'deepnet',
                     help = 'What model to use')
@@ -338,7 +341,7 @@ class DeepNet3(nn.Module):
 
 ######################################################################
 
-def nb_errors(model, data_set):
+def nb_errors(model, data_set, mistake_filename_pattern = None):
     ne = 0
     for b in range(0, data_set.nb_batches):
         input, target = data_set.get_batch(b)
@@ -348,6 +351,12 @@ def nb_errors(model, data_set):
         for i in range(0, data_set.batch_size):
             if wta_prediction[i] != target[i]:
                 ne = ne + 1
+                if mistake_filename_pattern is not None:
+                    img = input[i].clone()
+                    img.sub_(img.min())
+                    img.div_(img.max())
+                    torchvision.utils.save_image(img,
+                                                 mistake_filename_pattern.format(b + i, target[i]))
 
     return ne
 
@@ -550,7 +559,8 @@ for problem_number in map(int, args.problems.split(',')):
                                args.nb_test_samples, args.batch_size,
                                cuda = torch.cuda.is_available())
 
-        nb_test_errors = nb_errors(model, test_set)
+        nb_test_errors = nb_errors(model, test_set,
+                                   mistake_filename_pattern = 'mistake_{:d}_{:06d}.png')
 
         log_string('test_error {:d} {:.02f}% {:d} {:d}'.format(
             problem_number,