Spaces:
Running
Running
File size: 11,973 Bytes
1b51bf6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 |
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
from .text_embedding import TextEmbeddingModel
class Tree():
def __init__(self,path):
self.name = {}
self.childs = {}
self.father = {}
self.dep = {}
self.root = None
self.max_dep = 0
self.subtree = {}
self.grad_fa = {} # the node closest to the root for each leaf
with open(path, 'r') as f:
lines = f.readlines()
for line in lines:
parts = line.strip().split()
assert len(parts) == 3, "Each line must have exactly three parts"
now,fa,name = parts
now,fa = int(now),int(fa)
if name != 'none':
self.name[now] = name.split(',')
if fa != -1:
self.childs[fa] = self.childs.get(fa, []) + [now]
self.father[now] = fa
else:
self.root = now
self.fa_pos = torch.zeros((len(self.father),len(self.father)),dtype=torch.bool)
self.dfs(self.root)
#max_dep,N,N+K 0/1
self.pos_down2up = torch.zeros((self.max_dep,len(self.name),len(self.father)),dtype=torch.bool)
self.neg_down2up = torch.zeros((self.max_dep,len(self.name),len(self.father)),dtype=torch.bool)
self.pos_up2down = torch.zeros((self.max_dep,len(self.name),len(self.father)),dtype=torch.bool)
self.neg_up2down = torch.zeros((self.max_dep,len(self.name),len(self.father)),dtype=torch.bool)
self.pos_center = torch.zeros((self.max_dep,len(self.name)),dtype=torch.long)
self.mask_center = torch.zeros((self.max_dep,len(self.name),len(self.father)),dtype=torch.bool)
#max_dep,N 0/1
self.mask = torch.zeros((self.max_dep,len(self.name)),dtype=torch.bool)
self.depth = torch.zeros(len(self.name))
self.labels = torch.zeros(len(self.name),dtype=torch.long)
self.vis_leaf()
label_value = list(set(self.grad_fa.values()))
for key, value in self.grad_fa.items():
self.labels[key] = label_value.index(value)
def dfs(self, node, depth=0,grfa=-1):
self.dep[node] = depth
self.max_dep = max(self.max_dep, depth)
if node!=self.root:
self.subtree[node] = torch.zeros(len(self.father),dtype=torch.bool)
self.subtree[node][node] = 1
# if self.fa_pos.get(node) is None:
if self.father[node] != self.root:
self.fa_pos[node] = self.fa_pos[self.father[node]].clone()
self.fa_pos[node][node] = 1
if grfa == -1:
grfa = node
if self.childs.get(node) is None:
self.grad_fa[node] = grfa
for child in self.childs.get(node, []):
self.dfs(child, depth + 1,grfa)
if node!=self.root:
self.subtree[node] = torch.logical_or(self.subtree[node], self.subtree[child])
def gen_leaf_item(self,node):
last_node = -1
leaf_id = node
self.depth[node] = self.dep[node]
while node != self.root:
now_dep=self.dep[node]-1
self.mask[now_dep,leaf_id] = 1
self.pos_center[now_dep,leaf_id] = node
self.mask_center[now_dep,leaf_id] = torch.logical_not(torch.logical_or(self.fa_pos[node],self.subtree[node]))
self.mask_center[now_dep,leaf_id,node] = 1
if last_node == -1:
self.pos_down2up[now_dep,leaf_id] = self.subtree[node]
else:
self.pos_down2up[now_dep,leaf_id]=torch.logical_xor(self.subtree[node],self.subtree[last_node])
self.neg_down2up[now_dep,leaf_id]=torch.logical_not(self.subtree[node])
if self.father[node] == self.root:
self.neg_up2down[now_dep,leaf_id] = torch.logical_not(self.subtree[node])
else:
self.neg_up2down[now_dep,leaf_id] = torch.logical_xor(self.subtree[node],self.subtree[self.father[node]])
self.pos_up2down[now_dep,leaf_id] = self.subtree[node]
last_node = node
node = self.father[node]
def vis_leaf(self):
for node, name in self.name.items():
self.gen_leaf_item(node)
def display(self):
for node, name in self.name.items():
depth = self.dep[node]
print(f"{depth}- {name} {self.father[node]}")
class SimCLR_Tree(nn.Module):
def __init__(self, opt, fabric):
super(SimCLR_Tree, self).__init__()
self.temperature = opt.temperature
self.opt = opt
self.fabric = fabric
adapter_path = getattr(opt, "adapter_path", None)
self.model = TextEmbeddingModel(
opt.model_name,
lora=opt.lora,
use_pooling=opt.pooling,
lora_r=opt.lora_r,
lora_alpha=opt.lora_alpha,
lora_dropout=opt.lora_dropout,
adapter_path=adapter_path,
)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.tree = Tree(opt.tree_txt)
self.pos_down2up = self.tree.pos_down2up.to(self.device)
self.neg_down2up = self.tree.neg_down2up.to(self.device)
self.pos_up2down = self.tree.pos_up2down.to(self.device)
self.neg_up2down = self.tree.neg_up2down.to(self.device)
self.pos_center = self.tree.pos_center.to(self.device)
self.mask_center = self.tree.mask_center.to(self.device)
self.K = self.pos_down2up.shape[0]
self.mask = self.tree.mask.to(self.device)
self.depth = self.tree.depth.to(self.device)
self.root_labels = self.tree.labels.to(self.device)
self.esp = torch.tensor(1e-6, device=self.device)
self.max_dep = self.tree.max_dep
self.leaf_cnt = len(self.tree.name)
self.names2id = {}
for key, value in self.tree.name.items():
for item in value:
self.names2id[item] = key
self.vitual_center = nn.Parameter(
torch.randn((len(self.tree.father), opt.projection_size), device=self.device),
requires_grad=True,
)
nn.init.xavier_uniform_(self.vitual_center)
self.center_labels = torch.arange(len(self.tree.father), dtype=torch.long, device=self.device)
if adapter_path is not None:
self.load_tree_state(adapter_path)
def get_encoder(self):
return self.model
def save_pretrained(self, save_directory: str, save_tokenizer: bool = True):
os.makedirs(save_directory, exist_ok=True)
self.model.save_pretrained(save_directory, save_tokenizer=save_tokenizer)
torch.save(
{"vitual_center": self.vitual_center.detach().cpu()},
os.path.join(save_directory, "tree_state.pt"),
)
def load_tree_state(self, directory: str):
state_path = os.path.join(directory, "tree_state.pt")
if not os.path.exists(state_path):
return
state = torch.load(state_path, map_location=self.vitual_center.device)
self.vitual_center.data.copy_(state["vitual_center"].to(self.vitual_center.device))
def load_from_directory(self, directory: str, is_trainable: bool = True):
if getattr(self.opt, "lora", False):
self.model.load_adapter(directory, is_trainable=is_trainable)
else:
self.model = TextEmbeddingModel(
directory,
lora=False,
use_pooling=self.opt.pooling,
output_hidden_states=False,
)
self.load_tree_state(directory)
def _compute_logits(self, q,q_labels,k,k_labels,pos_mask,neg_mask):
def cosine_similarity_matrix(q, k):
q_norm = F.normalize(q,dim=-1)
k_norm = F.normalize(k,dim=-1)
cosine_similarity = q_norm@k_norm.T
return cosine_similarity
def gen_label_mask(relation_matrix,q_labels, k_labels):
N1 = q_labels.shape[0]
N2 = k_labels.shape[0]
q_labels_expanded = q_labels.unsqueeze(1).expand(-1, N2) # N1 x N2
k_labels_expanded = k_labels.unsqueeze(0).expand(N1, -1) # N1 x N2
result_matrix = relation_matrix[:,q_labels_expanded, k_labels_expanded]
return result_matrix
logits=cosine_similarity_matrix(q,k)
logits=logits/self.temperature
logits = logits.unsqueeze(0).expand(self.K,-1,-1) #K,N1,N2
pos_mask = gen_label_mask(pos_mask,q_labels, k_labels)
neg_mask = gen_label_mask(neg_mask,q_labels, k_labels) #K,N1,N2
pos_logits = torch.sum(logits*pos_mask,dim=-1)/torch.max(torch.sum(pos_mask,dim=-1),self.esp)#K,N1
pos_logits = pos_logits.unsqueeze(-1)#K,N1,1
neg_logits = logits*neg_mask#K,N1,N2
logits = torch.cat((pos_logits, neg_logits), dim=-1)#K,N1,N2+1
#model:model set
# pos_logits_model = torch.sum(logits*same_model,dim=1)/torch.max(torch.sum(same_model,dim=1),self.esp)# N
# neg_logits_model=logits*torch.logical_not(same_model)# N,N+K
# logits_model=torch.cat((pos_logits_model.unsqueeze(1), neg_logits_model), dim=1)
return logits
def forward(self, encoded_batch, labels):
q = self.model(encoded_batch)
N1 = q.shape[0]
k = q.clone().detach()
k = self.fabric.all_gather(k).view(-1, k.size(1))
k_labels = self.fabric.all_gather(labels).view(-1)
now_depth = self.depth[labels].unsqueeze(0).expand(self.K,-1)
now_mask = self.mask[:,labels]
# leaf_labels = self.root_labels[labels]
k = torch.concat((k,self.vitual_center),dim=0)
k_labels = torch.concat((k_labels,self.center_labels),dim=0)
logits_sample = self._compute_logits(q,labels,k,k_labels,self.pos_down2up,self.neg_down2up)#K,N1,N2+1
gt_sample = torch.zeros(logits_sample.shape[:-1], dtype=torch.long,device=logits_sample.device)
logits_sample = logits_sample.permute(0,2,1)
loss_smaple1 = F.cross_entropy(logits_sample, gt_sample, reduction='none') #K,N1
loss_smaple1 = torch.sum((loss_smaple1/now_depth)*now_mask)/N1*self.max_dep
# out = self.root_classfier(q)
# loss_classfiy = F.cross_entropy(out, leaf_labels)
loss = loss_smaple1
return loss,loss_smaple1
# def forward(self, encoded_batch, labels):
# q = self.model(encoded_batch)
# # N1 = q.shape[0]
# # k = q.clone().detach()
# # k = self.fabric.all_gather(k).view(-1, k.size(1))
# # k_labels = self.fabric.all_gather(labels).view(-1)
# # now_depth = self.depth[labels].unsqueeze(0).expand(self.K,-1)
# # now_mask = self.mask[:,labels]
# leaf_labels = self.root_labels[labels]
# # k = torch.concat((k,self.vitual_center),dim=0)
# # k_labels = torch.concat((k_labels,self.center_labels),dim=0)
# # logits_sample = self._compute_logits(q,labels,k,k_labels,self.pos_down2up,self.neg_down2up)#K,N1,N2+1
# # gt_sample = torch.zeros(logits_sample.shape[:-1], dtype=torch.long,device=logits_sample.device)
# # logits_sample = logits_sample.permute(0,2,1)
# # loss_smaple1 = F.cross_entropy(logits_sample, gt_sample, reduction='none') #K,N1
# # loss_smaple1 = torch.sum((loss_smaple1/now_depth)*now_mask)/N1*self.max_dep
# out = self.root_classfier(q)
# loss_classfiy = F.cross_entropy(out, leaf_labels)
# loss = loss_classfiy
# return loss,loss_classfiy
|