up

dataset/base.py

@@ -8,8 +8,8 @@ import torch
 import random
 from PIL import Image
 from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
-from model.simple_tokenizer import SimpleTokenizer as Tokenizer
-
+# from model.simple_tokenizer import SimpleTokenizer as Tokenizer
+from transformers import AutoTokenizer
 
 class BaseDataset(Dataset):
 
@@ -19,7 +19,7 @@ class BaseDataset(Dataset):
             indexs: dict,
             labels: dict,
             is_train=True,
-            tokenizer=Tokenizer(),
+            tokenizer=AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32", model_max_length=77, truncation=True),
             maxWords=32,
             imageResolution=224,
             npy=False):

dataset/dataloader.py

@@ -10,7 +10,7 @@ def split_data(captions, indexs, labels, query_num=5000, train_num=10000, seed=N
     random_index = np.random.permutation(range(len(indexs)))
     query_index = random_index[: query_num]
     train_index = random_index[query_num: query_num + train_num]
-    retrieval_index = random_index[query_num:]
+    retrieval_index = random_index[query_num:-100000]
 
     query_indexs = indexs[query_index]
     query_captions = captions[query_index]

main.py

@@ -1,11 +1,11 @@
-from train.text_train import Trainer
-
+# from train.text_train import Trainer
+from train.hash_train import Trainer
 
 if __name__ == "__main__":
 
     engine=Trainer()
     engine.test()
-    engine.train_epoch()
+    # engine.train_epoch()
     
     
     # engine.train()

model/simple_tokenizer.py

@@ -131,11 +131,16 @@ class SimpleTokenizer(object):
         text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
         return text
 
-    def my_decode(self, tokens):
-        tokens=[item for item in tokens if item in self.decoder]
-        text = ''.join([self.decoder[token] for token in tokens])
-        text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
-        return text
+    # def my_decode(self, tokens):
+    #     print(tokens)
+    #     tem_token=[]
+    #     for i in tokens:
+    #         if i in self.decoder.keys():
+    #             tem_token.append(self.decoder[i])
+    #     print(tem_token)
+    #     text = ''.join(tem_token)
+    #     text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
+    #     return text
     
     def tokenize(self, text):
         tokens = []
@@ -147,3 +152,6 @@ class SimpleTokenizer(object):
 
     def convert_tokens_to_ids(self, tokens):
         return [self.encoder[bpe_token] for bpe_token in tokens]
+
+    def convert_ids_to_tokens(self, ids):
+        return [self.decoder[id] for id in ids]

train/hash_train.py

@@ -183,18 +183,18 @@ class Trainer(TrainBase):
 
         mAP_t=cal_map(adv_img,adv_labels,retrieval_txt,retrieval_labels)
         # pr=cal_pr(retrieval_txt,adv_img,query_labels,retrieval_labels)
-        # pr_t=cal_pr(retrieval_txt,adv_img,adv_labels,retrieval_labels)
+        pr_t=cal_pr(retrieval_txt,adv_img,retrieval_labels,adv_labels)
         self.logger.info(f">>>>>> MAP_t: {mAP_t}")
         result_dict = {
             'adv_img': adv_img,
             'r_txt': retrieval_txt,
             'adv_l': adv_labels,
-            'r_l': retrieval_labels
+            'r_l': retrieval_labels,
             # 'q_l':query_labels
             # 'pr': pr,
-            # 'pr_t': pr_t
+            'pr_t': pr_t
         }
-        scio.savemat(os.path.join(save_dir, str(self.args.output_dim) + "-adv-" + self.args.dataset  + ".mat"), result_dict)
+        scio.savemat(os.path.join(save_dir, str(self.args.victim).replace("/", "_") + "-adv-" + self.args.dataset  + ".mat"), result_dict)
         self.logger.info(">>>>>> save all data!")
         
         
@@ -267,8 +267,8 @@ class Trainer(TrainBase):
         retrieval_labels = self.retrieval_labels.numpy()
         mAPi2t = cal_map(query_img,query_labels,retrieval_txt,retrieval_labels)
         mAPt2i =cal_map(query_txt,query_labels,retrieval_img,retrieval_labels)
-        # pr_i2t=cal_pr(retrieval_txt,query_img,query_labels,retrieval_labels)
-        # pr_t2i=cal_pr(retrieval_img,query_txt,query_labels,retrieval_labels)
+        pr_i2t=cal_pr(retrieval_txt,query_img,retrieval_labels,query_labels)
+        pr_t2i=cal_pr(retrieval_img,query_txt,retrieval_labels,query_labels)
         self.max_mapt2i = max(self.max_mapt2i, mAPi2t)
         self.logger.info(f">>>>>> MAP(i->t): {mAPi2t}, MAP(t->i): {mAPt2i}")
         result_dict = {
@@ -277,11 +277,11 @@ class Trainer(TrainBase):
             'r_img': retrieval_img,
             'r_txt': retrieval_txt,
             'q_l': query_labels,
-            'r_l': retrieval_labels
-            # 'pr_i2t': pr_i2t,
-            # 'pr_t2i': pr_t2i
+            'r_l': retrieval_labels,
+            'pr_i2t': pr_i2t,
+            'pr_t2i': pr_t2i
         }
-        scio.savemat(os.path.join(save_dir, str(self.args.output_dim) + "-ours-" + self.args.dataset  + ".mat"), result_dict)
+        scio.savemat(os.path.join(save_dir, str(self.args.victim).replace("/", "_") + "-ours-" + self.args.dataset  + ".mat"), result_dict)
         self.logger.info(">>>>>> save all data!")
 
 

train/text_train.py

@@ -16,10 +16,10 @@ from utils import get_args, calc_neighbor, cosine_similarity, euclidean_similari
 from utils.calc_utils import cal_map, cal_pr
 from dataset.dataloader import dataloader
 import clip
-from model.simple_tokenizer import SimpleTokenizer as Tokenizer
+import re
 from transformers import BertForMaskedLM
 from model.bert_tokenizer import BertTokenizer
-# from transformers import BertModel
+from transformers import AutoTokenizer
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
@@ -51,6 +51,14 @@ filter_words = ['a', 'about', 'above', 'across', 'after', 'afterwards', 'again',
                 '@', '%', '^', '*', '(', ')', "-", '-', '+', '=', '<', '>', '|', ':', ";", '～', '·']
 filter_words = set(filter_words)
 
+
+def text_filter(text):
+    text = re.findall(r"<|startoftext|>(.+)<|endoftext|>", text)
+    text = text[2]
+    text = re.sub(r'</w>', ' ', text)
+    return text
+
+
 class GoalFunctionStatus(object):
     SUCCEEDED = 0  # attack succeeded
     SEARCHING = 1  # In process of searching for a success
@@ -82,67 +90,6 @@ class GoalFunctionResult(object):
     def __hash__(self):
         return hash(self.text)
 
-# def get_bpe_substitues(substitutes, tokenizer, mlm_model):
-#     # substitutes L, k
-#     # device = mlm_model.device
-#     substitutes = substitutes[0:12, 0:4]  # maximum BPE candidates
-#
-#     # find all possible candidates
-#
-#     all_substitutes = []
-#     for i in range(substitutes.size(0)):
-#         if len(all_substitutes) == 0:
-#             lev_i = substitutes[i]
-#             all_substitutes = [[int(c)] for c in lev_i]
-#         else:
-#             lev_i = []
-#             for all_sub in all_substitutes:
-#                 for j in substitutes[i]:
-#                     lev_i.append(all_sub + [int(j)])
-#             all_substitutes = lev_i
-#
-#     # all substitutes  list of list of token-id (all candidates)
-#     c_loss = nn.CrossEntropyLoss(reduction='none')
-#     word_list = []
-#     # all_substitutes = all_substitutes[:24]
-#     all_substitutes = torch.tensor(all_substitutes)  # [ N, L ]
-#     all_substitutes = all_substitutes[:24].to(device)
-#     # print(substitutes.size(), all_substitutes.size())
-#     N, L = all_substitutes.size()
-#     word_predictions = mlm_model(all_substitutes)[0]  # N L vocab-size
-#     ppl = c_loss(word_predictions.view(N * L, -1), all_substitutes.view(-1))  # [ N*L ]
-#     ppl = torch.exp(torch.mean(ppl.view(N, L), dim=-1))  # N
-#     _, word_list = torch.sort(ppl)
-#     word_list = [all_substitutes[i] for i in word_list]
-#     final_words = []
-#     for word in word_list:
-#         tokens = [tokenizer._convert_id_to_token(int(i)) for i in word]
-#         text = tokenizer.convert_tokens_to_string(tokens)
-#         final_words.append(text)
-#     return final_words
-
-# def get_substitues(substitutes, tokenizer, mlm_model, use_bpe, substitutes_score=None, threshold=3.0):
-#     # substitues L,k
-#     # from this matrix to recover a word
-#     words = []
-#     sub_len, k = substitutes.size()  # sub-len, k
-#
-#     if sub_len == 0:
-#         return words
-#
-#     elif sub_len == 1:
-#         for (i, j) in zip(substitutes[0], substitutes_score[0]):
-#             if threshold != 0 and j < threshold:
-#                 break
-#             words.append(tokenizer._convert_id_to_token(int(i)))
-#     else:
-#         if use_bpe == 1:
-#             words = get_bpe_substitues(substitutes, tokenizer, mlm_model)
-#         else:
-#             return words
-#     #
-#     # print(words)
-#     return words
 
 class Trainer(TrainBase):
 
@@ -155,9 +102,10 @@ class Trainer(TrainBase):
         self.image_mean = image_mean
         self.image_var = image_var
         self.device = rank
-        self.clip_tokenizer=Tokenizer()
+        self.clip_tokenizer = AutoTokenizer.from_pretrained("openai/clip-vit-base-patch32", model_max_length=77,
+                                                            truncation=True)
         self.bert_tokenizer = BertTokenizer.from_pretrained(self.args.text_encoder, do_lower_case=True)
-        self.ref_net = BertForMaskedLM.from_pretrained(self.args.text_encoder)
+        self.ref_net = BertForMaskedLM.from_pretrained(self.args.text_encoder).to(device)
         self.attack_thred = self.args.attack_thred
         # self.run()
 
@@ -233,14 +181,17 @@ class Trainer(TrainBase):
 
         masked_embeds = []
         for i in range(0, len(masked_texts), batch_size):
-            masked_text_input = self.bert_tokenizer(masked_texts[i:i+batch_size], padding='max_length', truncation=True, max_length=max_length, return_tensors='pt').to(device)
+            masked_text_input = self.bert_tokenizer(masked_texts[i:i + batch_size], padding='max_length',
+                                                    truncation=True, max_length=max_length, return_tensors='pt').to(
+                device)
             masked_embed = self.ref_net(masked_text_input.text_inputs, attention_mask=masked_text_input.attention_mask)
             masked_embeds.append(masked_embed)
         masked_embeds = torch.cat(masked_embeds, dim=0)
 
         criterion = torch.nn.KLDivLoss(reduction='none')
 
-        import_scores = criterion(masked_embeds.log_softmax(dim=-1), origin_embeds.softmax(dim=-1).repeat(len(masked_texts), 1))
+        import_scores = criterion(masked_embeds.log_softmax(dim=-1),
+                                  origin_embeds.softmax(dim=-1).repeat(len(masked_texts), 1))
 
         return import_scores.sum(dim=-1)
 
@@ -265,14 +216,14 @@ class Trainer(TrainBase):
     def get_word_predictions(self, text):
         _, _, keys = self._tokenize(text)
 
-        inputs = self.bert_tokenizer.encode_plus(text, add_special_tokens=True, max_length=self.max_text_len,
+        inputs = self.bert_tokenizer.encode_plus(text, add_special_tokens=True, max_length=self.args.max_words,
                                                  truncation=True, return_tensors="pt")
         input_ids = inputs["input_ids"].to(self.device)
         attention_mask = inputs['attention_mask']
         with torch.no_grad():
-            word_predictions = self.ref_net(input_ids)['logits'].squeeze()  # (seq_len, vocab_size)
-
-        word_pred_scores_all, word_predictions = torch.topk(word_predictions, self.max_candidate, -1)
+            word_predictions = self.ref_net(input_ids)['logits'].squeeze(0)  # (seq_len, vocab_size)
+        # print(self.ref_net(input_ids)['logits'].shape)
+        word_pred_scores_all, word_predictions = torch.topk(word_predictions, self.args.max_candidate, -1)
 
         word_predictions = word_predictions[1:-1, :]  # remove [CLS] and [SEP]
         word_pred_scores_all = word_pred_scores_all[1:-1, :]
@@ -345,17 +296,25 @@ class Trainer(TrainBase):
 
     def get_goal_results(self, trans_texts, negetive_code, negetive_mean, negative_var, positive_code, positive_mean,
                          positive_var, beta=10, temperature=0.05):
-        trans_feature= self.clip_tokenizer(trans_texts)
+        # print(trans_texts)
+        trans_feature = clip.tokenize(trans_texts,context_length=77,truncate=True).to(device)
         anchor = self.model.encode_text(trans_feature)
-        loss1 = F.triplet_margin_with_distance_loss(anchor, positive_code, negetive_code,
+        batch_size=anchor.shape[0]
+        loss1 = F.triplet_margin_with_distance_loss(anchor, positive_code.repeat(batch_size, 1), negetive_code.repeat(batch_size, 1),
                                                     distance_function=nn.CosineSimilarity(), reduction='none')
-        negative_dist = (anchor - negetive_mean) ** 2 / negative_var
-        positive_dist = (anchor - positive_mean) ** 2 / positive_var
+        sim=F.cosine_similarity(anchor,positive_code.unsqueeze(0), dim=1, eps=1e-8).unsqueeze(1)
+        negative_dist = (anchor - negetive_mean) ** 2 / (negative_var+ 1e-5)
+        positive_dist = (anchor - positive_mean) ** 2 / (positive_var+ 1e-5)
         negatives = torch.exp(negative_dist / temperature)
         positives = torch.exp(positive_dist / temperature)
-        loss = torch.log(positives / (positives + negatives)) + beta * loss1
-        return loss
-
+        loss = torch.log(positives / (positives + negatives)).mean(dim=1, keepdim=True) + beta * loss1
+        results = []
+        # print(loss.shape)
+        for i in range(len(trans_texts)):
+            if loss[i].shape[0] >1 or sim[i] >self.args.sim_threshold:
+                continue
+            results.append(GoalFunctionResult(trans_texts[i], score=loss[i], similarity=sim[i]))
+        return results
 
     def generate_mapping(self):
         image_train = []
@@ -370,8 +329,12 @@ class Trainer(TrainBase):
         image_train = np.concatenate(image_train, axis=0)
         label_train = np.concatenate(label_train, axis=0)
         label_unipue = np.unique(label_train, axis=0)
-        image_centroids =np.stack([image_train[find_indices(label_train,label_unipue[i])].mean(axis=0) for i in range(len(label_unipue))], axis=0)
-        image_var=np.stack([image_train[find_indices(label_train,label_unipue[i])].var(axis=0) for i in range(len(label_unipue))], axis=0)
+        image_centroids = np.stack(
+            [image_train[find_indices(label_train, label_unipue[i])].mean(axis=0) for i in range(len(label_unipue))],
+            axis=0)
+        image_var = np.stack(
+            [image_train[find_indices(label_train, label_unipue[i])].var(axis=0) for i in range(len(label_unipue))],
+            axis=0)
 
         image_representation = {}
         image_var_representation = {}
@@ -380,10 +343,9 @@ class Trainer(TrainBase):
             image_var_representation[str(centroid.astype(int))] = image_var[i]
         return image_representation, image_var_representation
 
-    
-
     def target_adv(self, raw_text, negetive_code, negetive_mean, negative_var, positive_code, positive_mean,
                    positive_var, beta=10, temperature=0.05):
+        # print(raw_text)
         keys, word_predictions, word_pred_scores_all, mask = self.get_word_predictions(raw_text)
 
         #clean state
@@ -391,7 +353,6 @@ class Trainer(TrainBase):
         cur_result = GoalFunctionResult(raw_text)
         mask_idx = np.where(mask.cpu().numpy() == 1)[0]
 
-
         for idx in mask_idx:
             predictions = word_predictions[keys[idx][0]: keys[idx][1]]
             predictions_socre = word_pred_scores_all[keys[idx][0]: keys[idx][1]]
@@ -427,9 +388,6 @@ class Trainer(TrainBase):
             cur_result.status = GoalFunctionStatus.FAILED
         return cur_result
 
-
-
-    
     def train_epoch(self):
         # self.change_state(mode="valid")
         save_dir = os.path.join(self.args.save_dir, "adv_PR_cruve")
@@ -443,7 +401,6 @@ class Trainer(TrainBase):
             print(times)
             image.float()
 
-            raw_text=[self.clip_tokenizer.my_decode(token) for token in text]
             image = image.to(self.rank, non_blocking=True)
             text = text.to(self.rank, non_blocking=True)
             negetive_mean = np.stack([self.image_mean[str(i.astype(int))] for i in label.detach().cpu().numpy()])
@@ -464,11 +421,17 @@ class Trainer(TrainBase):
             positive_mean = torch.from_numpy(positive_mean).to(self.rank, non_blocking=True)
             positive_var = torch.from_numpy(positive_var).to(self.rank, non_blocking=True)
             positive_code = self.model.encode_image(target_image)
-            
-        
-            final_adverse=self.target_adv(text, raw_text,negetive_code,negetive_mean,negative_var,
-                                  positive_code,positive_mean,positive_var)
-            final_text=self.clip_tokenizer.tokenize(final_adverse.text).to(self.rank, non_blocking=True)
+            # print(self.clip_tokenizer.my_encode('This day is good!'))
+            raw_text = [self.clip_tokenizer.convert_ids_to_tokens(token.cpu()) for token in text]
+            raw_text = [text_filter(self.clip_tokenizer.convert_tokens_to_string(txt)) for txt in raw_text]
+            final_texts=[]
+            for i in range(self.args.batch_size):
+                adv_txt=self.target_adv( raw_text[i], negetive_code[i], negetive_mean[i], negative_var[i],
+                                            positive_code[i], positive_mean[i], positive_var[i])
+                final_texts.append(adv_txt.text)
+            # final_adverse = self.target_adv( raw_text, negetive_code, negetive_mean, negative_var,
+            #                                 positive_code, positive_mean, positive_var)
+            final_text = clip.tokenize(final_texts,context_length=77,truncate=True).to(self.rank, non_blocking=True)
             adv_code = self.model.encode_text(final_text)
             adv_codes.append(adv_code.cpu().detach().numpy())
             adv_label.append(target_label.numpy())
@@ -477,12 +440,9 @@ class Trainer(TrainBase):
 
         _, retrieval_txt = self.get_code(self.retrieval_loader, self.args.retrieval_num)
 
-        
-        
         retrieval_txt = retrieval_txt.cpu().detach().numpy()
         retrieval_labels = self.retrieval_labels.numpy()
 
-
         mAP_t = cal_map(adv_img, adv_labels, retrieval_txt, retrieval_labels)
         self.logger.info(f">>>>>> MAP_t: {mAP_t}")
         result_dict = {
@@ -494,14 +454,10 @@ class Trainer(TrainBase):
             # 'pr': pr,
             # 'pr_t': pr_t
         }
-        scio.savemat(os.path.join(save_dir, str(self.args.output_dim) + "-adv-" + self.args.dataset  + ".mat"), result_dict)
+        scio.savemat(os.path.join(save_dir, str(self.args.victim).replace("/", "_") + "-adv-" + self.args.dataset + ".mat"),
+                     result_dict)
         self.logger.info(">>>>>> save all data!")
 
-        
-        
-
-        
-
     def train(self):
         self.logger.info("Start train.")
 
@@ -510,9 +466,8 @@ class Trainer(TrainBase):
             self.valid(epoch)
             self.save_model(epoch)
 
-        self.logger.info(f">>>>>>> FINISHED >>>>>> Best epoch, I-T: {self.best_epoch_i}, mAP: {self.max_mapi2t}, T-I: {self.best_epoch_t}, mAP: {self.max_mapt2i}")
-
-    
+        self.logger.info(
+            f">>>>>>> FINISHED >>>>>> Best epoch, I-T: {self.best_epoch_i}, mAP: {self.max_mapi2t}, T-I: {self.best_epoch_t}, mAP: {self.max_mapt2i}")
 
     def make_hash_code(self, code: list) -> torch.Tensor:
 
@@ -542,15 +497,10 @@ class Trainer(TrainBase):
 
         return img_buffer, text_buffer  # img_buffer.to(self.rank), text_buffer.to(self.rank)
 
-   
-        
-    
     def valid_attack(self, adv_images, texts, adv_labels):
         save_dir = os.path.join(self.args.save_dir, "adv_PR_cruve")
         os.makedirs(save_dir, exist_ok=True)
 
-    
-
     def test(self, mode_name="i2t"):
         self.logger.info("Valid Clean.")
         save_dir = os.path.join(self.args.save_dir, "PR_cruve")
@@ -558,7 +508,6 @@ class Trainer(TrainBase):
         query_img, query_txt = self.get_code(self.query_loader, self.args.query_num)
         retrieval_img, retrieval_txt = self.get_code(self.retrieval_loader, self.args.retrieval_num)
 
-
         query_img = query_img.cpu().detach().numpy()
         query_txt = query_txt.cpu().detach().numpy()
         retrieval_img = retrieval_img.cpu().detach().numpy()
@@ -579,31 +528,11 @@ class Trainer(TrainBase):
             'q_l': query_labels,
             'r_l': retrieval_labels
         }
-        scio.savemat(os.path.join(save_dir, str(self.args.output_dim) + "-ours-" + self.args.dataset  + ".mat"), result_dict)
+        scio.savemat(os.path.join(save_dir, str(self.args.victim).replace("/", "_") + "-ours-" + self.args.dataset + ".mat"),
+                     result_dict)
         self.logger.info(">>>>>> save all data!")
 
 
-    # def valid(self, epoch):
-    #     self.logger.info("Valid.")
-    #     self.change_state(mode="valid")
-    #     query_img, query_txt = self.get_code(self.query_loader, self.args.query_num) if self.args.hash_layer == "select" else super().get_code(self.query_loader, self.args.query_num)
-    #     retrieval_img, retrieval_txt = self.get_code(self.retrieval_loader, self.args.retrieval_num) if self.args.hash_layer == "select" else super().get_code(self.retrieval_loader, self.args.retrieval_num)
-    #     # print("get all code")
-    #     mAPi2t = calc_map_k(query_img, retrieval_txt, self.query_labels, self.retrieval_labels, None, self.rank)
-    #     # print("map map")
-    #     mAPt2i = calc_map_k(query_txt, retrieval_img, self.query_labels, self.retrieval_labels, None, self.rank)
-    #     mAPi2i = calc_map_k(query_img, retrieval_img, self.query_labels, self.retrieval_labels, None, self.rank)
-    #     mAPt2t = calc_map_k(query_txt, retrieval_txt, self.query_labels, self.retrieval_labels, None, self.rank)
-    #     if self.max_mapi2t < mAPi2t:
-    #         self.best_epoch_i = epoch
-    #         self.save_mat(query_img, query_txt, retrieval_img, retrieval_txt, mode_name="i2t")
-    #     self.max_mapi2t = max(self.max_mapi2t, mAPi2t)
-    #     if self.max_mapt2i < mAPt2i:
-    #         self.best_epoch_t = epoch
-    #         self.save_mat(query_img, query_txt, retrieval_img, retrieval_txt, mode_name="t2i")
-    #     self.max_mapt2i = max(self.max_mapt2i, mAPt2i)
-    #     self.logger.info(f">>>>>> [{epoch}/{self.args.epochs}], MAP(i->t): {mAPi2t}, MAP(t->i): {mAPt2i}, MAP(t->t): {mAPt2t}, MAP(i->i): {mAPi2i}, \
-    #                 MAX MAP(i->t): {self.max_mapi2t}, MAX MAP(t->i): {self.max_mapt2i}")
 
     def save_mat(self, query_img, query_txt, retrieval_img, retrieval_txt, mode_name="i2t"):
 
@@ -625,7 +554,7 @@ class Trainer(TrainBase):
             'q_l': query_labels,
             'r_l': retrieval_labels
         }
-        scio.savemat(os.path.join(save_dir, str(self.args.output_dim) + "-ours-" + self.args.dataset + "-" + mode_name + ".mat"), result_dict)
+        scio.savemat(
+            os.path.join(save_dir, str(self.args.victim).replace("/", "_") + "-ours-" + self.args.dataset + "-" + mode_name + ".mat"),
+            result_dict)
         self.logger.info(f">>>>>> save best {mode_name} data!")
-
-

utils/get_args.py

@@ -9,20 +9,22 @@ def get_args():
     parser.add_argument("--save-dir", type=str, default="./result/64-bit")
     parser.add_argument("--clip-path", type=str, default="./ViT-B-32.pt", help="pretrained clip path.")
     parser.add_argument("--pretrained", type=str, default="")
-    parser.add_argument("--dataset", type=str, default="flickr25k", help="choise from [coco, mirflckr25k, nuswide]")
+    parser.add_argument("--dataset", type=str, default="coco", help="choise from [coco, mirflckr25k, nuswide]")
     parser.add_argument("--index-file", type=str, default="index.mat")
     parser.add_argument("--caption-file", type=str, default="caption.mat")
     parser.add_argument("--label-file", type=str, default="label.mat")
     parser.add_argument("--similarity-function", type=str, default="euclidean", help="choise form [cosine, euclidean]")
     parser.add_argument("--loss-type", type=str, default="l2", help="choise form [l1, l2]")
-    parser.add_argument('--victim', default='ViT-B/16', choices=['ViT-L/14', 'ViT-B/16', 'ViT-B/32', 'RN50', 'RN101'])
+    parser.add_argument('--victim', default='RN50', choices=['ViT-L/14', 'ViT-B/16', 'ViT-B/32', 'RN50', 'RN101'])
     parser.add_argument("--text_encoder", type=str, default="bert-base-uncased")
     parser.add_argument("--topk", type=int, default=10)
     parser.add_argument("--num-perturbation", type=int, default=3)
     parser.add_argument("--txt-dim", type=int, default=1024)
-    parser.add_argument("--output-dim", type=int, default=512)
+    parser.add_argument("--output-dim", type=int, default=1024)
     parser.add_argument("--epochs", type=int, default=100)
     parser.add_argument("--max-words", type=int, default=77)
+    parser.add_argument("--max-candidate", type=int, default=7)
+    parser.add_argument("--enable-bpe", type=bool, default=False)
     parser.add_argument("--resolution", type=int, default=224)
     parser.add_argument("--batch-size", type=int, default=8)
     parser.add_argument("--num-workers", type=int, default=4)