-
Notifications
You must be signed in to change notification settings - Fork 0
/
SparseMOE.py
295 lines (222 loc) · 8.53 KB
/
SparseMOE.py
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
286
287
288
289
290
291
292
293
294
295
import torch
import torch.nn as nn
from torch.nn import functional as F
from torch.nn import init
##############################
torch.manual_seed(42)
###dataset
batch_size=16
block_size=32
###train
max_iters=500
learning_rate=1e-3
device='cuda'
###evel
eval_iters=400
eval_interval=100
###model
head_size=16
n_embed=128
n_head=8
n_layer=8
dropout=0.1
num_experts=8
topk=2
##############################
with open('input.txt','r',encoding='utf-8') as f:
text=f.read()
chars=sorted(list(set(text)))
vocab_size=len(chars)
stoi={ch:i for i,ch in enumerate(chars)}
itos={i:ch for i,ch in enumerate(chars)}
encode=lambda s:[stoi[c] for c in s]
decode=lambda l:''.join(itos[i] for i in l)
#[46, 48, 40, 1, 46, 43, 56, 43]
#hjb here
data=torch.tensor(encode(text),dtype=torch.long)
#torch.Size([1115389])
###################划分数据集
n=int(0.9*len(data))
train_data=data[:n]
val_data=data[n:]
'''
tensor([258140, 252540, 992416, 387428])
tensor([[39, 52, 58, 1, 58, 46, 43, 47],
[56, 42, 2, 0, 20, 43, 1, 47],
[41, 46, 1, 51, 63, 1, 44, 56],
[39, 63, 1, 63, 53, 59, 1, 52]])
'''
def get_batch(split):
data=train_data if split=='train' else val_data
ix = torch.randint(len(data) - block_size, (batch_size,))
x = torch.stack([data[i:i + block_size] for i in ix])
y = torch.stack([data[i + 1:i + block_size + 1] for i in ix])
return x,y
'''
torch.Size([4, 8])
tensor([[ 1, 58, 46, 47, 52, 45, 1, 21],
[43, 42, 6, 1, 39, 52, 42, 1],
[51, 43, 52, 58, 0, 32, 46, 39],
[10, 1, 57, 46, 43, 1, 50, 47]])
'''
#torch.Size([4, 8, 16])
class Head(nn.Module):
def __init__(self,head_size,n_embed):
super(Head, self).__init__()
self.key = nn.Linear(n_embed, head_size, bias=False)
self.query = nn.Linear(n_embed, head_size, bias=False)
self.value = nn.Linear(n_embed, head_size, bias=False)
self.register_buffer('tril',torch.tril(torch.ones(block_size,block_size)))
self.dropout=nn.Dropout(0.1)
def forward(self,x):
B,T,C=x.shape
k=self.key(x)
q=self.query(x)
v=self.value(x)
score=q@k.transpose(-2,-1)*C**-0.5
score=score.masked_fill(self.tril[:T,:T]==0,float('-inf'))#[B,T,T]
score=F.softmax(score,dim=-1)
score=self.dropout(score)
output=score@v
return output
class MultiHeadAttention(nn.Module):
def __init__(self,num_heads,head_size,n_embed):
super(MultiHeadAttention, self).__init__()
self.heads=nn.ModuleList([Head(head_size,n_embed) for _ in range(num_heads)])
self.proj=nn.Linear(n_embed,n_embed)
self.dropout=nn.Dropout(0.1)
def forward(self,x):
output=torch.cat([h(x) for h in self.heads],dim=-1)
output=self.dropout(self.proj(output))
return output
class Expert(nn.Module):
def __init__(self,n_embed):
super(Expert, self).__init__()
self.net=nn.Sequential(
nn.Linear(n_embed,4*n_embed),
nn.ReLU(),
nn.Linear(4*n_embed,n_embed),
nn.Dropout(0.1),
)
def forward(self,x):
return self.net(x)
class NoisyTopkRouter(nn.Module):
def __init__(self,n_embed,num_experts,top_k):
super(NoisyTopkRouter, self).__init__()
self.topk=top_k
self.topk_router=nn.Linear(n_embed,num_experts)
self.noisy_linear=nn.Linear(n_embed,num_experts)
def forward(self,x):
#torch.Size([4, 8, 16])
logits=self.topk_router(x)
noisy_logits=self.noisy_linear(x)
noisy=torch.randn_like(logits)*F.softplus(noisy_logits)
noisy_logits=logits+noisy
topk_logits,indices=noisy_logits.topk(self.topk,dim=-1)
zeros=torch.full_like(noisy_logits,float('-inf'))
sparse_logits=zeros.scatter(-1,indices,topk_logits)
router_output=F.softmax(sparse_logits,dim=-1)
return router_output,indices
class SparseMoE(nn.Module):
def __init__(self,n_embed,num_experts,top_k):
super(SparseMoE, self).__init__()
self.router=NoisyTopkRouter(n_embed,num_experts,top_k)
self.experts=nn.ModuleList([Expert(n_embed) for _ in range(num_experts)])
self.top_k=top_k
def forward(self,x):
#torch.Size([4, 8, 16])
gating_output,indices=self.router(x)
#torch.Size([4, 8, 8]) 8个专家
#torch.Size([4, 8, 2])
final_output=torch.zeros_like(x)
#torch.Size([4, 8, 16])
flat_x=x.view(-1,x.size(-1))
#torch.Size([32, 16])
flat_gating_output=gating_output.view(-1,gating_output.size(-1))
#torch.Size([32, 8])
for i,expert in enumerate(self.experts):
#当前batch size中,对于第i个专家,对原始输入滤出最大表征为编号i的那部分表征
expert_mask=(indices==i).any(dim=-1)
'''
tensor([[False, False, False, False, True, False, True, False],
[False, False, False, False, False, False, False, False],
[False, False, False, False, False, False, False, False],
[ True, True, False, False, True, False, False, False]])
'''
flat_mask=expert_mask.view(-1)
#torch.Size([32])
if flat_mask.any():
expert_input=flat_x[flat_mask]
#torch.Size([5, 16])
expert_output=expert(expert_input)
#torch.Size([5, 16])
gating_scores=flat_gating_output[flat_mask,i].unsqueeze(1)#[32,8]->[5,8]->[5]->[5,1]
#torch.Size([5, 1])
weighted_output=expert_output*gating_scores
#torch.Size([5, 16])
final_output[expert_mask]+=weighted_output.squeeze(1)
#torch.Size([4, 8, 16])
return final_output
class Block(nn.Module):
def __init__(self,n_embed,n_head,head_size,num_experts,top_k):
super(Block, self).__init__()
self.sa=MultiHeadAttention(n_head,head_size,n_embed)
self.smoe=SparseMoE(n_embed,num_experts,top_k)
self.ln1=nn.LayerNorm(n_embed)
self.ln2=nn.LayerNorm(n_embed)
def forward(self,x):
x=x+self.sa(self.ln1(x))
x=x+self.smoe(self.ln2(x))
return x
class SparseMoELM(nn.Module):
def __init__(self):
super(SparseMoELM, self).__init__()
self.token_embedding_table=nn.Embedding(vocab_size,n_embed)
self.position_embedding_table=nn.Embedding(block_size,n_embed)
self.blocks=nn.Sequential(*[Block(n_embed,n_head,head_size,num_experts,topk) for _ in range(n_layer)])
self.ln_f=nn.LayerNorm(n_embed)
self.lm_head=nn.Linear(n_embed,vocab_size)
def forward(self,idx,target=None):
B,T=idx.shape
tok_emb=self.token_embedding_table(idx)
pos_emb=self.position_embedding_table(torch.arange(T,device=device))
x=tok_emb+pos_emb
x=self.blocks(x)
x=self.ln_f(x)
logits=self.lm_head(x)
if target is None:
loss=None
else:
B,T,C=logits.shape
logits=logits.view(B*T,C)
target=target.view(B*T)
loss=F.cross_entropy(logits,target)
return logits,loss
def generate(self,idx,max_new_token):
for _ in range(max_new_token):
idx_cond=idx[:,-block_size:]
logits,loss=self(idx_cond)
logits=logits[:,-1,:]#[B,C],最后一个位置上
probs=F.softmax(logits,dim=-1)#[B,C]
idx_next=torch.multinomial(probs,num_samples=1) #[B,1]
idx=torch.cat((idx,idx_next),dim=1)
return idx
def kaiming_init_weights(m):
if isinstance(m,(nn.Linear)):
init.kaiming_normal_(m.weight)
model=SparseMoELM().to(device)
model.apply(kaiming_init_weights)
print(sum(p.numel() for p in model.parameters())/1e6,'M parameters')
optimizer=torch.optim.AdamW(model.parameters(),lr=learning_rate)
for iter in range(max_iters):
xb,yb =get_batch('train')
xb=xb.to(device)
yb=yb.to(device)
logits,loss=model(xb,yb)
optimizer.zero_grad(set_to_none=True)
loss.backward()
optimizer.step()
if iter % 20 ==0:
print('iter: %d, loss: %3f' % (iter,loss))
context=torch.zeros((1,1),dtype=torch.long,device=device)
print(decode(model.generate(context,2000)[0].tolist()))