Makes suggested corrections to JOSS paper. Corrects accidental copy/p…

…aste mistake in gen dataset defs.

cyphercat/datadefs/cyphercat_dataset.py

@@ -73,5 +73,5 @@ def get_preload_split_fn(name=''):
         fn = PRELOAD_SPLIT_FN_DICT[name]
         return fn
     else:
-        raise ValueError('Invalid test statistic, {}, entered. Must be '
+        raise ValueError('Invalid dataset, {}, entered. Must be '
                          'in {}'.format(name, PRELOAD_SPLIT_FN_DICT.keys()))

docs/joss_paper/paper.md

@@ -1,5 +1,5 @@
 ---
-title: 'Cyphercat: A Python Package for Reproduceably Evaluating Adversarial Robustness'
+title: 'Cyphercat: A Python Package for Reproduceably Evaluating Robustness Against Privacy Attacks'
 tags:
   - Python
   - machine learning
@@ -65,8 +65,8 @@ Cyphercat is an extensible Python package for benchmarking privacy
 attack and defense efficacy in a reproducible environment.
 The Cyphercat application programming interface (API) allows users to test the 
 robustness of a specified target model against several well-documented privacy
-attacks [@mlleaks][@fredrikson2015model], which extract details of the training data from the model, 
-including the option to assess defenses.
+attacks [@mlleaks][@fredrikson2015model], which aim to extract detailed of the training data from the model.
+Also included is the option to further assess the efficacy of several implemented defense methods.
 The API is built on the PyTorch [@pytorch] machine learning library and 
 provides access to well known image, audio, and text benchmark datasets used for machine learning applications.
 The Cyphercat API includes the option to train on commonly used architectures, 

scripts/test_train.py

@@ -80,7 +80,9 @@ def main():
     ccatset = CCATDataset(path=data_struct.save_path, name=data_name, splits=splits, transforms=[train_transform])
     trainset = ccatset.get_split_n(0)
     testset = ccatset.get_split_n(1)
-    n_classes = 10
+    n_classes = data_struct.n_classes
+    img_size = data_struct.height
+    n_in = data_struct.channels
 
     ## Define pyTorch ingestors for training and testing
     trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=2)
@@ -98,7 +100,7 @@ def main():
     #imshow(torchvision.utils.make_grid(imgs))  
 
     # Prepare the model for training
-    model = get_predef_model(model_name)(n_in=3, n_classes=n_classes, n_filters=64, size=250)
+    model = get_predef_model(model_name)(n_in=n_in, n_classes=n_classes, n_filters=64, size=img_size)
     model.to(device)
     model.apply(weights_init)
     model_optim = optim.Adam(model.parameters(), lr=learnrate/10)