v3.2.4

Harry24k · Dec 23, 2021 · 6fec117 · 6fec117
1 parent 3d18299
commit 6fec117
Show file tree

Hide file tree

Showing 4 changed files with 109 additions and 93 deletions.
diff --git a/UPDATE_HISTORY.md b/UPDATE_HISTORY.md
@@ -1,58 +1,67 @@
-
 ### ~~v0.3~~
+
 * ~~**New Attacks** : FGSM, IFGSM, IterLL, RFGSM, CW(L2), PGD are added.~~
 * ~~**Demos** are uploaded.~~
 
 ### ~~v0.4~~
+
 * ~~**DO NOT USE** : 'init.py' is omitted.~~
 
 ### ~~v0.5~~
+
 * ~~**Package name changed** : 'attacks' is changed to 'torchattacks'.~~
 * ~~**New Attack** : APGD is added.~~
 * ~~**attack.py** : 'update_model' method is added.~~
 
 ### ~~v0.6~~
+
 * ~~**Error Solved** :~~ 
-    * ~~Before this version, even after getting an adversarial image, the model remains evaluation mode.~~
-    * ~~To solve this, below methods are modified.~~
-        * ~~'_switch_model' method is added into **attack.py**. It will automatically change model mode to the previous mode after getting adversarial images. When getting adversarial images, model is switched to evaluation mode.~~
-        * ~~'__call__' methods in all attack changed to forward. Instead of this, '__call__' method is added into 'attack.py'~~
+  * ~~Before this version, even after getting an adversarial image, the model remains evaluation mode.~~
+  * ~~To solve this, below methods are modified.~~
+    * ~~'_switch_model' method is added into **attack.py**. It will automatically change model mode to the previous mode after getting adversarial images. When getting adversarial images, model is switched to evaluation mode.~~
+    * ~~'__call__' methods in all attack changed to forward. Instead of this, '__call__' method is added into 'attack.py'~~
 * ~~**attack.py** : To provide ease of changing images to uint8 from float, 'set_mode' and '_to_uint' is added.~~
-    * ~~'set_mode' determines returning all outputs as 'int' OR 'flaot' through '_to_uint'.~~
-    * ~~'_to_uint' changes all outputs into uint8.~~
+  * ~~'set_mode' determines returning all outputs as 'int' OR 'flaot' through '_to_uint'.~~
+  * ~~'_to_uint' changes all outputs into uint8.~~
 
 ### ~~v0.7~~
+
 * ~~**All attacks are modified**~~
-    * ~~clone().detach() is used instead of .data~~
-    * ~~torch.autograd.grad is used instead of .backward() and .grad :~~
-        * ~~It showed 2% reduction of computation time.~~
-    
+  * ~~clone().detach() is used instead of .data~~
+  * ~~torch.autograd.grad is used instead of .backward() and .grad :~~
+    * ~~It showed 2% reduction of computation time.~~
+
 ### ~~v0.8~~
+
 * ~~**New Attack** : RPGD is added.~~
 * ~~**attack.py** : 'update_model' method is depreciated. Because torch models are passed by call-by-reference, we don't need to update models.~~
-    * ~~**cw.py** : In the process of cw attack, now masked_select uses a mask with dtype torch.bool instead of a mask with dtype torch.uint8.~~
+  * ~~**cw.py** : In the process of cw attack, now masked_select uses a mask with dtype torch.bool instead of a mask with dtype torch.uint8.~~
 
 ### ~~v0.9~~
+
 * ~~**New Attack** : DeepFool is added.~~
 * ~~**Some attacks are renamed** :~~
-    * ~~I-FGSM -> BIM~~
-    * ~~IterLL -> StepLL~~
+  * ~~I-FGSM -> BIM~~
+  * ~~IterLL -> StepLL~~
 
 ### ~~v1.0~~
+
 * ~~**attack.py** :~~
-    * ~~**load** : Load is depreciated. Instead, use TensorDataset and DataLoader.~~
-    * ~~**save** : The problem of calculating invalid accuracy when the mode of the attack set to 'int' is solved.~~
+  * ~~**load** : Load is depreciated. Instead, use TensorDataset and DataLoader.~~
+  * ~~**save** : The problem of calculating invalid accuracy when the mode of the attack set to 'int' is solved.~~
 
 ### ~~v1.1~~
+
 * ~~**DeepFool** :~~
-    * ~~[**Error solved**](https://github.com/Harry24k/adversairal-attacks-pytorch/issues/2).~~
-    
+  * ~~[**Error solved**](https://github.com/Harry24k/adversairal-attacks-pytorch/issues/2).~~
+
 ### ~~v1.2~~
+
 * ~~**Description has been added for each module.**~~
 * ~~**Sphinx Document uploaded**~~ 
 * ~~**attack.py** : 'device' will be decided by [next(model.parameters()).device](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/3#issue-602571865).~~
 * ~~**Two attacks are merged** :~~
-    * ~~RPGD, PGD -> PGD~~
+  * ~~RPGD, PGD -> PGD~~
 
 
 
@@ -144,93 +153,94 @@
 ### v2.11.0
 
   * ``CW``
-      * Now it outputs the best L2 adversarial images.
-      * Faster computation.
+    * Now it outputs the best L2 adversarial images.
+    * Faster computation.
   * `DeepFool`
-      * Make the codes cleaner.
+    * Make the codes cleaner.
   * `BIM`
-      * Bug fixed: Wrong cliping.
+    * Bug fixed: Wrong cliping.
   * `MIFGSM`
-      * Bug fixed: Wrong cliping.
-      * Bug fixed: [Gradient Norm](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/12).
+    * Bug fixed: Wrong cliping.
+    * Bug fixed: [Gradient Norm](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/12).
   * Demo Added
-      * Performance Comparison (CIFAR10)
+    * Performance Comparison (CIFAR10)
 
 
 
 ### v2.12.1
 
   * `DeepFool`
-      * Deprecated.
+    * Deprecated.
   * `Attack._targeted`
-      * ._targeted is set to 1 when targeted mode is activated. [Issue](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/14).
-          * All attacks supporting targeted mode is change.
+    * ._targeted is set to 1 when targeted mode is activated. [Issue](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/14).
+      * All attacks supporting targeted mode is change.
   * `Attack.set_attack_mode`
-      * To provide various attack mode, it uses following methods.
-          * `set_default_mode`: default mode.
-          * `set_targeted_mode`: targeted mode. Now supporting `target_map_function=None` for pre-generated targeted labels.
-          * `set_least_likely_mode`: least likely targeted mode. Now supporting k-th smallest probability targeted mode by `kth_min`.
+    * To provide various attack mode, it uses following methods.
+      * `set_default_mode`: default mode.
+      * `set_targeted_mode`: targeted mode. Now supporting `target_map_function=None` for pre-generated targeted labels.
+      * `set_least_likely_mode`: least likely targeted mode. Now supporting k-th smallest probability targeted mode by `kth_min`.
   * `Attack.save`
-      * Bug fixed: When `verbose=True`, it now use model.eval() and torch.no_grad().
+    * Bug fixed: When `verbose=True`, it now use model.eval() and torch.no_grad().
 
 
 
 ### v2.12.1
 
   * `DeepFool`
-      * Deprecated.
+    * Deprecated.
   * `Attack._targeted`
-      * ._targeted is set to 1 when targeted mode is activated. [Issue](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/14).
-          * All attacks supporting targeted mode is change.
+    * ._targeted is set to 1 when targeted mode is activated. [Issue](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/14).
+      * All attacks supporting targeted mode is change.
   * `Attack.set_attack_mode`
-      * To provide various attack mode, it uses following methods.
-          * `set_default_mode`: default mode.
-          * `set_targeted_mode`: targeted mode. Now supporting `target_map_function=None` for pre-generated targeted labels.
-          * `set_least_likely_mode`: least likely targeted mode. Now supporting k-th smallest probability targeted mode by `kth_min`.
+    * To provide various attack mode, it uses following methods.
+      * `set_default_mode`: default mode.
+      * `set_targeted_mode`: targeted mode. Now supporting `target_map_function=None` for pre-generated targeted labels.
+      * `set_least_likely_mode`: least likely targeted mode. Now supporting k-th smallest probability targeted mode by `kth_min`.
   * `Attack.save`
-      * Bug fixed: When `verbose=True`, it now use model.eval() and torch.no_grad().
+    * Bug fixed: When `verbose=True`, it now use model.eval() and torch.no_grad().
 
 
 
 ### v2.12.2
 
   * `PGDL2`
-      * PGD with L2 distance measure.
+    * PGD with L2 distance measure.
   * `Attack.save`
-      * Print L2 distance between adversarial examples and the original examples.
+    * Print L2 distance between adversarial examples and the original examples.
 
 
 
 ### v2.12.3
 
   * `PGDL2`
+
     * Initialization perturbation is changed.
+
 
-
-
+
 ### v2.13.1
 
   * `Attack.set_attack_mode`
-      * Deprecated. Use following built-in functions.
-          * `set_mode_default`: default mode.
-          * `set_mode_targeted`: targeted mode. Now supporting `target_map_function=None` for pre-generated targeted labels.
-          * `set_mode_least_likely`: least likely targeted mode. Now supporting k-th smallest probability targeted mode by `kth_min`.
+    * Deprecated. Use following built-in functions.
+      * `set_mode_default`: default mode.
+      * `set_mode_targeted`: targeted mode. Now supporting `target_map_function=None` for pre-generated targeted labels.
+      * `set_mode_least_likely`: least likely targeted mode. Now supporting k-th smallest probability targeted mode by `kth_min`.
   * `APGD` is changed to `EOTPGD`.
   * `PGDDLR` is added.
   * `APGD`, `APGDT`, `Square`, `FAB`
-      * Modified from https://github.com/fra31/auto-attack.
-          * `n_iters` is changed to `steps`.
-          * `n_target_classes` is calculated based on `n_claases`.
-          * `reduce=False` is erased because it is enough with `reduction='none'`.
+    * Modified from https://github.com/fra31/auto-attack.
+      * `n_iters` is changed to `steps`.
+      * `n_target_classes` is calculated based on `n_claases`.
+      * `reduce=False` is erased because it is enough with `reduction='none'`.
   * `AutoAttack`
-      * Created based on `APGD`, `APGDT`, `Square`, `FAB`.
+    * Created based on `APGD`, `APGDT`, `Square`, `FAB`.
 
 
 
 ### v2.13.2
 
   * `Attack.save`
-      * Don't use an additional memory if `save_path=None`
+    * Don't use an additional memory if `save_path=None`
 
 
 
@@ -245,8 +255,8 @@
 ### v2.14.1
 
   * `Attack.set_training_mode`
-      * The method to support changing the model to training mode.
-      * Note that RNN requires model.training=True to calculate gradient.
+    * The method to support changing the model to training mode.
+    * Note that RNN requires model.training=True to calculate gradient.
 
 
 
@@ -272,7 +282,7 @@
 ### v2.14.4
 
   * `Square` is fixed.
-      * If idx_to_fool is empty, then terminate an iteration.
+    * If idx_to_fool is empty, then terminate an iteration.
 
 
 
@@ -281,9 +291,9 @@
 ### v2.14.5
 
   * `MIFGSM` is fixed.
-      * https://github.com/Harry24k/adversarial-attacks-pytorch/issues/33
+    * https://github.com/Harry24k/adversarial-attacks-pytorch/issues/33
   * `CW` is fixed.
-      * https://github.com/Harry24k/adversarial-attacks-pytorch/issues/32
+    * https://github.com/Harry24k/adversarial-attacks-pytorch/issues/32
 
 
 
@@ -297,18 +307,18 @@
 
   * Targeted mode is officially supported.
 
-      * `Attack` & `Attacks.*`
+    * `Attack` & `Attacks.*`
 
-          * `set_mode_default`
-          * `set_mode_targeted_by_function`
-          * `set_mode_targeted_least_likely`
-          * `set_mode_targeted_random`
-          * `_get_target_label`
-          * `_get_least_likely_label`
-          * `_get_random_target_label`
+      * `set_mode_default`
+      * `set_mode_targeted_by_function`
+      * `set_mode_targeted_least_likely`
+      * `set_mode_targeted_random`
+      * `_get_target_label`
+      * `_get_least_likely_label`
+      * `_get_random_target_label`
 
-        * `self._supported_mode`
-        * `self._targeted`
+      * `self._supported_mode`
+      * `self._targeted`
 
 * `UPGD` created.
 
@@ -331,7 +341,7 @@
 ### v3.1.0
 
   * `TIFGSM` is added.
-      * https://github.com/Harry24k/adversarial-attacks-pytorch/pull/29/commits
+    * https://github.com/Harry24k/adversarial-attacks-pytorch/pull/29/commits
 
 
 
@@ -341,18 +351,20 @@
 ### v3.2.0
 
   * `Jitter` is added.
+
   * `Attack.*`
+
     * `set_training_mode`: Now supports changing training mode of `Batchnorm` and `Dropout`.
     * `save`: Now supports return values of the last verbose information.
 
   * `MultiAttack`
 
-      * Fixed the verbose function.
-      * Now supports return values of the last verbose information
+    * Fixed the verbose function.
+    * Now supports return values of the last verbose information
 
-      
+
 
-      
+
 
 
 ### v3.2.1
@@ -367,15 +379,16 @@
 ### v3.2.2
 
   * `SparseFool`: [bug fixed](https://github.com/Harry24k/adversarial-attacks-pytorch/pull/43).
+
   * `MultiAttack`: [bug fixed](https://github.com/Harry24k/adversarial-attacks-pytorch/issues/44).
 
 
 
 
 
-      
+      
 
-      
+      
 
 
 ### v3.2.3
@@ -384,7 +397,11 @@
 
 
 
+      
 
-
+      
 
-
+
+### v3.2.4
+
+  * `save`, `MultiAttack`: `return_verbose` can be `True `even if `verbose=False`.
diff --git a/torchattacks/__init__.py b/torchattacks/__init__.py
@@ -24,4 +24,4 @@
 from .attacks.tifgsm import TIFGSM
 from .attacks.jitter import Jitter
 
-__version__ = '3.2.3'
+__version__ = '3.2.4'
diff --git a/torchattacks/attack.py b/torchattacks/attack.py
@@ -157,10 +157,7 @@ def save(self, data_loader, save_path=None, verbose=True, return_verbose=False,
             return_verbose (bool): True for returning detailed information. (Default: False)
             save_pred (bool): True for saving predicted labels (Default: False)
 
-        """
-        if (verbose==False) and (return_verbose==True):
-            raise ValueError("Verobse should be True if return_verbose==True.")
-
+        """            
         if save_path is not None:
             image_list = []
             label_list = []
@@ -183,7 +180,7 @@ def save(self, data_loader, save_path=None, verbose=True, return_verbose=False,
 
             batch_size = len(images)
 
-            if verbose:
+            if verbose or return_verbose:
                 with torch.no_grad():
                     if given_training:
                         self.model.eval()
@@ -200,7 +197,8 @@ def save(self, data_loader, save_path=None, verbose=True, return_verbose=False,
                     l2 = torch.cat(l2_distance).mean().item()
                     progress = (step+1)/total_batch*100
                     elapsed_time = end-start
-                    self._save_print(progress, rob_acc, l2, elapsed_time, end='\r')
+                    if verbose:
+                        self._save_print(progress, rob_acc, l2, elapsed_time, end='\r')
 
             if save_path is not None:
                 if given_return_type == 'int':