The paper mentions three augmentation strategies, but I only found two in the code.

The paper mentions three data augmentation strategies: Flip Response, Drop Item, and Swap Adjacent Items. However, I could only find code for Flip Response ：
s_flip[b, i] = 1 - s_flip[b, i]
and Swap Adjacent Items：
q_[b, i], q_[b, i + 1] = q_[b, i + 1], q_[b, i]
s_[b, i], s_[b, i + 1] = s_[b, i + 1], s_[b, i].
I couldn't locate code for Drop Item. Can you provide information to help me find the code or provide the core code as in the previous examples?

Drop Item strategy is achieved by applying random mask here. Pos/neg sequences will be assigned different masking to achieve the effect of the contrastive drop operation.

DTransformer/DTransformer/model.py

Lines 320 to 333 in 479175b

    
           # mask manipulation 
        
           if self.training: 
        
               mask = mask.expand(query.size(0), -1, -1, -1).contiguous() 
        
               for b in range(query.size(0)): 
        
                   # sample for each batch 
        
                   if lens[b] < MIN_SEQ_LEN: 
        
                       # skip for short sequences 
        
                       continue 
        
                   idx = random.sample( 
        
                       range(lens[b] - 1), max(1, int(lens[b] * self.dropout_rate)) 
        
                   ) 
        
                   for i in idx: 
        
                       mask[b, :, i + 1 :, i] = 0

Thank you very much!

	# mask manipulation
	if self.training:
	mask = mask.expand(query.size(0), -1, -1, -1).contiguous()

	for b in range(query.size(0)):
	# sample for each batch
	if lens[b] < MIN_SEQ_LEN:
	# skip for short sequences
	continue
	idx = random.sample(
	range(lens[b] - 1), max(1, int(lens[b] * self.dropout_rate))
	)
	for i in idx:
	mask[b, :, i + 1 :, i] = 0