Update ip_adapter/attention_processor.py

ip_adapter/attention_processor.py

@@ -216,51 +216,11 @@ class IPAttnProcessor(nn.Module):
         ip_key = attn.head_to_batch_dim(ip_key)
         ip_value = attn.head_to_batch_dim(ip_value)
 
-        # ip_key.shape=[16, 4, 40]
-        # query = [16,6025,40]
-        # target = [16,6025,4]
-        # print("**************************************************")
-        # print(query)
-        # print("**************************************************")
-        # threshold = 5
-        # tensor_from_data = torch.tensor(query).to("cuda")
-        # binary_mask = torch.where(tensor_from_data > threshold, torch.tensor(0).to("cuda"), torch.tensor(1).to("cuda"))
-        # binary_mask = binary_mask.to(torch.float16)
-        # print("**************************************************")
-        # print(binary_mask)
-        # print("**************************************************")
-
-
-        # query.shape=[16,6205,40]
         ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
-        ##########################################
-        # attention_probs
-        #ip_attention_probs = attn.get_attention_scores(keyforIPADAPTER, ip_key, None)
-        ##########################################
-
         ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
-        ##########################################
-        # ip_hidden_states = ip_hidden_states*binary_mask +(1-binary_mask)*query
-        ##########################################
         ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
-
-        # hidden_states.shape=【2,6205,320】s
-        # ip_hidden_states.shape=【2,3835,320】
-        # hidden_states = hidden_states + self.scale* ip_hidden_states
-        #print("Control_factor:{}".format(self.Control_factor))
-        #print("IP_factor:{}".format(self.IP_factor))
         hidden_states = self.Control_factor * hidden_states + self.IP_factor * self.scale * ip_hidden_states
 
-
-
-        #hidden_states = 2*hidden_states +0.6*self.scale*ip_hidden_states
-        # if self.roundNumber < 12:
-        #     hidden_states = hidden_states
-        # else:
-        #     hidden_states = 1.2*hidden_states +0.6*self.scale*ip_hidden_states
-        # self.roundNumber = self.roundNumber + 1
-
-
         # linear proj
         hidden_states = attn.to_out[0](hidden_states)
         # dropout