-
Notifications
You must be signed in to change notification settings - Fork 2
/
demo_glove.py
183 lines (142 loc) · 5.74 KB
/
demo_glove.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
import time
import torch
import sys
from KDEformer import KDEformer
from Performer import PerformerAttention
from Reformer import ReformerAttention
try:
from src.models.attention.sblocal_attention import SBLocalAttention
SB_INSTALLED = True
except:
print("ScatterBrain is not installed.")
SB_INSTALLED = False
from functools import partial
def set_random_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
def get_glove(seq_len=8192, seed=1):
data = open("glove.twitter.27B.100d.txt", "r").read().split("\n")
batch_size, head_size = 8, 1
max_seq_len = 16384
data_all = []
for i in range(batch_size * head_size * max_seq_len * 3):
data_i = [float(_x) for _x in data[i].split()[1:]]
if len(data_i) != 100:
data_i = [float(_x) for _x in data[i].split()]
data_all.append(data_i)
data_all = torch.tensor(data_all).double()
query = data_all[:batch_size*head_size*max_seq_len]
query = query.reshape(batch_size, max_seq_len, head_size, -1)
value = data_all[2*batch_size*head_size*max_seq_len:]
value = value.reshape(batch_size, max_seq_len, head_size, -1)
del data, data_all
query = query[:,:seq_len,:,:]
value = value[:,:seq_len,:,:]
normalizer = 10**0.25
query /= normalizer
key = query.clone()
return query, key, value
class ExactAttention(torch.nn.Module):
def __init__(self):
super().__init__()
def forward(self, query, key, value):
qk = torch.einsum("bhnd,bhmd->bhnm", query.transpose(1,2), key.transpose(1,2))
attn = torch.softmax(qk, dim=-1)
return torch.einsum('bhnm,bhmd->bhnd', attn, value.transpose(1,2))
@torch.no_grad()
def run(seq_len=8192, device='cpu', seed=1, num_feats=256):
query, key, value = get_glove(seq_len)
query, key, value = map(lambda _x: _x.to(device).double(), (query, key, value))
tic0 = time.time()
set_random_seed(seed)
batch_size = query.shape[0]
head_size = query.shape[2]
exact_attn = ExactAttention()
inputs = (query, key, value)
# 1. Exact attention
method = 'Exact'
tic = time.time()
out_exact = exact_attn(query, key, value)
tim_exact = time.time() - tic
print(f"[TIM] computing exact attention | {tim_exact:<12.8f} sec | ")
out_exact = out_exact.to('cpu')
err_normalizer = torch.linalg.norm(out_exact, ord=2, dim=(2, 3))
def compute_rel_err(A):
return torch.mean(torch.linalg.norm(A - out_exact, ord=2, dim=(2,3)) / err_normalizer)
# 2. Reformer attention
method = 'Reformer'
if query.shape[1] != value.shape[1]:
print(f"Reformer is not available for this input.")
tim_rfm, err_rfm, reformer_flops, reformer_mems = -1, -1, -1, -1
else:
reformer = ReformerAttention(bucket_size=num_feats//2).to(device)
reformer_call = partial(reformer, qk=query, k=None, v=value)
out_rfm, _ = reformer_call()
tim_rfm = time.time() - tic
out_rfm = out_rfm.unsqueeze(1).to('cpu')
err_rfm = compute_rel_err(out_rfm).item()
print(f"{method:<16} | {num_feats} | {err_rfm:.10f} | {tim_rfm:<12.8f} sec | ")
del out_rfm
# 3. Performer attetion
method = 'Performer'
performer = PerformerAttention(num_feats=num_feats).to(device)
performer_call = partial(performer, query=query, key=key, value=value)
tic = time.time()
out_pfm = performer_call()
tim_pfm = time.time() - tic
out_pfm = out_pfm.to('cpu')
err_pfm = compute_rel_err(out_pfm).item()
print(f"{method:<16} | {num_feats} | {err_pfm:.10f} | {tim_pfm:<12.8f} sec | ")
del out_pfm
# 4. ScatterBrain (sblocal)
if SB_INSTALLED:
method = 'ScatterBrain'
local_context = num_feats // 2
tic = time.time()
sblocal_attn = SBLocalAttention(local_context=local_context, dim_heads=query.shape[-1], nb_features=num_feats, softmax_temp=1.).to(device)
sblocal_call = partial(sblocal_attn, query=query, key=key, value=value)
out_sbn, _ = sblocal_call()
tim_sbn = time.time() - tic
out_sbn = out_sbn.transpose(1,2)
out_sbn = out_sbn.to('cpu')
err_sbn = compute_rel_err(out_sbn).item()
print(f"{method:<16} | {local_context + num_feats} | {err_sbn:.10f} | {tim_sbn:<12.8f} sec | ")
del out_sbn
# 5. Our KDEformer
method = 'KDEformer'
sample_size = num_feats
mask_size = batch_size * head_size * sample_size * query.shape[1]
kde_attn = KDEformer(num_projs=7, Bucket_size=num_feats//2 , sample_size=num_feats, mask_size=mask_size).to(device)
kde_call = partial(kde_attn, query=query, key=key, value=value)
tic = time.time()
out_our = kde_call()
tim_our = time.time() - tic
out_our = out_our.to('cpu')
err_our = compute_rel_err(out_our).item()
del out_our
print(f"{method:<16} | {int(1.5 * num_feats)} | {err_our:.10f} | {tim_our:<12.8f} sec | ")
print()
def varying_num_feats(
seq_len=8192,
num_iters=1,
num_feats_all=[64, 128, 256, 512, 1024, 2048, 4096],
calc_flops=False,
calc_memory=False,
debug=False
):
it = 0
tic00 = time.time()
for seed in range(num_iters):
print(f"iter: {it} | ", end='')
print(f" seed: {seed}")
results = {}
for bs in num_feats_all:
print(f"seq_len: {seq_len}, bucket_size: {bs}, seed: {seed}")
# results[bs] = main(seq_len=seq_len, bucket_size=bs, seed=seed)
results[bs] = run(seq_len=seq_len, seed=seed, num_feats=bs, query_key_same=True)
print(f"elapsed time: {time.time() - tic00:.4f} sec")
print(f"[iter {it} is done.]")
it += 1
if __name__ == "__main__":
run()
# varying_num_feats()