-
Notifications
You must be signed in to change notification settings - Fork 3
/
latency_predictor.py
303 lines (255 loc) · 15 KB
/
latency_predictor.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
296
297
298
299
300
301
302
303
# HAT: Hardware-Aware Transformers for Efficient Natural Language Processing
# Hanrui Wang, Zhanghao Wu, Zhijian Liu, Han Cai, Ligeng Zhu, Chuang Gan and Song Han
# The 58th Annual Meeting of the Association for Computational Linguistics (ACL), 2020.
# Paper: https://arxiv.org/abs/2005.14187
# Project page: https://hanruiwang.me/project_pages/hat/
# AutoMoE: Neural Architecture Search for Efficient Sparsely Activated Transformers
# Ganesh Jawahar, Subhabrata Mukherjee, Xiaodong Liu, Young Jin Kim, Muhammad Abdul-Mageed, Laks V. S. Lakshmanan, Ahmed Hassan Awadallah, Sebastien Bubeck, Jianfeng Gao
# Paper: https://arxiv.org/abs/2210.07535
import random
import configargparse
import numpy as np
import json
import torch
import torch.nn as nn
import torch.nn.functional as F
from sklearn import linear_model, ensemble
from sklearn.metrics import mean_squared_error
import sys
from fairseq import utils
import xgboost as xgb
class Net(nn.Module):
def __init__(self, feature_dim, hidden_dim, hidden_layer_num):
super(Net, self).__init__()
self.first_layer = nn.Linear(feature_dim, hidden_dim)
self.layers = nn.ModuleList()
for i in range(hidden_layer_num):
self.layers.append(nn.Linear(hidden_dim, hidden_dim))
self.predict = nn.Linear(hidden_dim, 1)
def forward(self, x):
x = F.relu(self.first_layer(x))
for i in range(len(self.layers)):
x = F.relu(self.layers[i](x))
x = self.predict(x)
return x
class LatencyPredictor(object):
def __init__(self, feature_norm, lat_norm, ckpt_path, lat_dataset_path='./latency_dataset/lat.tmp', feature_dim=10, hidden_dim=400, hidden_layer_num=3, train_steps=5000, bsz=128, lr=1e-5, use_sklearn_predictor=None):
self.dataset_path = lat_dataset_path
self.feature_norm = np.array(feature_norm)
self.lat_norm = lat_norm
self.feature_dim = feature_dim
self.hidden_dim = hidden_dim
self.hidden_layer_num = hidden_layer_num
self.ckpt_path = ckpt_path
self.dataset = None
self.train_x = None
self.train_y = None
self.valid_x = None
self.valid_y = None
self.test_x = None
self.test_y = None
self.model = Net(self.feature_dim, self.hidden_dim, self.hidden_layer_num)
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=lr)
self.criterion = torch.nn.MSELoss()
self.train_steps = train_steps
self.bsz = bsz
def train(self):
for i in range(self.train_steps):
sample_ind = random.sample(range(len(self.train_x)), k=self.bsz)
sample_x = [self.train_x[sample_ind[k]] for k in range(self.bsz)]
sample_y = [self.train_y[sample_ind[k]] for k in range(self.bsz)]
sample_x_tensor = torch.Tensor(sample_x)
sample_y_tensor = torch.Tensor(sample_y)
prediction = self.model(sample_x_tensor).squeeze()
loss = self.criterion(prediction, sample_y_tensor)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# validation
if i % 100 == 0:
with torch.no_grad():
sample_x_tensor = torch.Tensor(self.valid_x)
sample_y_tensor = torch.Tensor(self.valid_y)
prediction = self.model(sample_x_tensor).squeeze()
loss = self.criterion(prediction, sample_y_tensor)
print(f"Validation loss at {i} steps: {loss}")
# test
with torch.no_grad():
sample_x_tensor = torch.Tensor(self.test_x)
sample_y_tensor = torch.Tensor(self.test_y)
prediction = self.model(sample_x_tensor).squeeze()
print(prediction.size(), sample_y_tensor.size())
loss = self.criterion(prediction, sample_y_tensor)
print(loss)
print(f"Predicted latency: {prediction}")
print(f"Real latency: {self.test_y}")
print(f"Loss: {loss}")
print(f"RMSE: {np.sqrt(self.criterion(self.lat_norm*sample_y_tensor, self.lat_norm*prediction))}")
print(f"MAPD: {torch.mean(torch.abs((sample_y_tensor - prediction) / sample_y_tensor))}")
torch.save(self.model.state_dict(), self.ckpt_path)
def load_ckpt(self):
self.model.load_state_dict(torch.load(self.ckpt_path))
def predict_lat(self, config):
with torch.no_grad():
print(config)
features = utils.get_config_features(config, None)
features = features[0:len(self.feature_norm)] # TODO: proper fix needed. introduced when adding number of experts to route to dimension
features_norm = np.array(features) / self.feature_norm
features_norm = np.array(features) / self.feature_norm
prediction = self.model(torch.Tensor(features_norm)).item() * self.lat_norm
return prediction
def split(self, train_size=None, fixed_val_test_size=None):
sample_num = len(self.dataset['x'])
if not fixed_val_test_size:
train_num = int(np.floor(0.8 * sample_num))
valid_num = int(np.floor(0.1 * sample_num))
else:
train_num = train_size
valid_num = fixed_val_test_size
self.train_x = self.dataset['x'][:train_num]
self.train_y = self.dataset['y'][:train_num]
self.valid_x = self.dataset['x'][train_num:(train_num+valid_num)]
self.valid_y = self.dataset['y'][train_num:(train_num+valid_num)]
self.test_x = self.dataset['x'][(train_num+valid_num):]
self.test_y = self.dataset['y'][(train_num+valid_num):]
def read_dataset(self, lat_features=None, add_overall_model_size=None, add_active_model_size=None):
features_norm_all = []
lats_all = []
with open(self.dataset_path, 'r') as fid:
next(fid) # skip first line of CSV
for line in fid:
if not lat_features:
features = line.split(',')[:self.feature_dim]
features_eval = list(map(eval, features))
features_norm = np.array(features_eval) / self.feature_norm
features_norm_all.append(features_norm)
lats = line.split(',')[self.feature_dim:]
total_lat = eval(lats[0]) + eval(lats[1])
lats_all.append(total_lat / self.lat_norm)
else:
features = json.loads(line.strip())
features_norm = []
# {"encoder_encoder_embed_dim": 640, "encoder_encoder_layer_num": 6, "encoder_encoder_ffn_embed_dim": [2048, 3072, 1024, 1024, 512, 512], "encoder_encoder_self_attention_heads": [4, 2, 4, 8, 2, 2], "decoder_decoder_embed_dim": 640, "decoder_decoder_layer_num": 3, "decoder_decoder_ffn_embed_dim": [512, 3072, 3072], "decoder_decoder_self_attention_heads": [8, 2, 2], "decoder_decoder_ende_attention_heads": [8, 8, 4], "encoder_encoder_n_experts": [1, 1, 1], "decoder_decoder_n_experts": [1, 1, 1], "encoder_latencies": 6.844480037689209, "decoder_latencies": 124.20365142822266}
# 640 6 2048 6 640 6 2048 6 6 2 2 2
# encoder_embed_dim,encoder_layer_num,encoder_ffn_embed_dim_avg,encoder_self_attention_heads_avg,decoder_embed_dim,decoder_layer_num,decoder_ffn_embed_dim_avg,decoder_self_attention_heads_avg,decoder_ende_attention_heads_avg,decoder_arbitrary_ende_attn_avg,encoder_n_experts_avg,decoder_n_experts_avg,latency_mean_encoder,latency_mean_decoder,latency_std_encoder,latency_std_decoder
for feat in sorted(features):
if feat == "encoder_latencies" or feat == "decoder_latencies" or "size" in feat:
continue
if feat.endswith("encoder_embed_dim") or feat.endswith("layer_num") or feat.endswith("decoder_embed_dim"):
feat_val = features[feat]
feat_norm = None
if feat.endswith("embed_dim"):
feat_norm = 640
elif feat.endswith("layer_num"):
feat_norm = 6
features_norm.append(feat_val/feat_norm)
else:
for i in range(6):
feat_val = features[feat][i] if i < len(features[feat]) else 0
feat_norm = None
if feat.endswith("ffn_embed_dim"):
feat_norm = 2048
elif feat.endswith("self_attention_heads") or feat.endswith("ende_attention_heads"):
feat_norm = 6
elif feat.endswith("n_experts"):
feat_norm = 2
features_norm.append(feat_val/feat_norm)
if add_overall_model_size:
features_norm.append(features["model_size"]/200000000)
if add_active_model_size:
features_norm.append(features["active_model_size"]/200000000)
features_norm_all.append(features_norm)
total_lat = features["encoder_latencies"] + features["decoder_latencies"]
lats_all.append(total_lat / self.lat_norm)
tmp = list(zip(features_norm_all, lats_all))
random.shuffle(tmp)
features_norm_all, lats_all = zip(*tmp)
self.dataset = {'x': features_norm_all, 'y': lats_all}
if __name__=='__main__':
parser = configargparse.ArgumentParser()
parser.add_argument('--configs', required=True, is_config_file=True)
parser.add_argument('--dataset-path')
parser.add_argument('--lat-dataset-path', type=str, default='./latency_dataset/lat.tmp', help='the path to read latency dataset')
parser.add_argument('--feature-norm', type=float, nargs='+', default=[640, 6, 2048, 6, 640, 6, 2048, 6, 6, 2], help='normalizing factor for each feature')
parser.add_argument('--lat-norm', type=float, default=200, help='normalizing factor for latency')
parser.add_argument('--feature-dim', type=int, default=10, help='dimension of feature vector')
parser.add_argument('--hidden-dim', type=int, default=400, help='hidden dimension of FC layers in latency predictor')
parser.add_argument('--hidden-layer-num', type=int, default=3, help='number of FC layers')
parser.add_argument('--ckpt-path', type=str, default='latency_dataset/ckpts/tmp.pt', help='path to save latency predictor weights')
parser.add_argument('--train-steps', type=int, default=5000, help='latency predictor training steps')
parser.add_argument('--bsz', type=int, default=128, help='latency predictor training batch size')
parser.add_argument('--lr', type=float, default=1e-5, help='latency predictor training learning rate')
parser.add_argument('--fixed-val-test-size', type=int, default=-1, help='control for the size of validation and test set')
parser.add_argument('--lat-features', type=str, default=None, help="latency predictor features")
parser.add_argument('--add-overall-model-size', type=str, default=None, help="add overall model size")
parser.add_argument('--add-active-model-size', type=str, default=None, help="add active model size")
parser.add_argument('--use-sklearn-predictor', type=str, default=None, help="use classifier from sklearn")
args = parser.parse_args()
print(args)
predictor = LatencyPredictor(lat_dataset_path=args.lat_dataset_path,
feature_norm=args.feature_norm,
lat_norm=args.lat_norm,
feature_dim=args.feature_dim,
hidden_dim=args.hidden_dim,
hidden_layer_num=args.hidden_layer_num,
ckpt_path=args.ckpt_path,
train_steps=args.train_steps,
bsz=args.bsz,
lr=args.lr)
predictor.read_dataset(lat_features=args.lat_features, add_overall_model_size=args.add_overall_model_size, add_active_model_size=args.add_active_model_size)
if args.use_sklearn_predictor is not None:
model = None
if args.use_sklearn_predictor == "LinearRegression":
model = linear_model.LinearRegression()
elif args.use_sklearn_predictor == "GradientBoostingRegressor":
params = {"n_estimators": 500, "max_depth": 4, "min_samples_split": 5, "learning_rate": 0.01, "loss": "squared_error"}
model = ensemble.GradientBoostingRegressor(**params)
elif args.use_sklearn_predictor == "XGBRegressor":
model = xgb.XGBRegressor(objective="reg:linear", random_state=42, n_estimators=1000)
predictor.split()
model.fit(predictor.train_x, predictor.train_y)
y_pred = model.predict(predictor.test_x)
print(y_pred[0:10])
print(predictor.test_y[0:10])
print("Loss = %.2f"%(mean_squared_error(predictor.lat_norm*np.array(predictor.test_y), predictor.lat_norm*np.array(y_pred))))
print("RMSE = %.2f"%(np.sqrt(mean_squared_error(predictor.lat_norm*np.array(predictor.test_y), predictor.lat_norm*np.array(y_pred)))))
sample_x_tensor = torch.Tensor(predictor.test_x)
sample_y_tensor = torch.Tensor(predictor.test_y)
prediction = torch.Tensor(y_pred)
loss = predictor.criterion(prediction, sample_y_tensor)
print(f"Loss: {loss}")
print(f"RMSE: {np.sqrt(predictor.criterion(predictor.lat_norm*sample_y_tensor, predictor.lat_norm*prediction))}")
sys.exit(0)
if args.fixed_val_test_size == -1:
predictor.split()
predictor.train()
else:
for train_size in [1600, 4600, 9600]:
predictor.model = Net(args.feature_dim, args.hidden_dim, args.hidden_layer_num)
predictor.optimizer = torch.optim.Adam(predictor.model.parameters(), lr=args.lr)
predictor.criterion = torch.nn.MSELoss()
predictor.split(train_size=train_size, fixed_val_test_size=args.fixed_val_test_size)
predictor.train()
print('Latency predictor training finished')
predictor.load_ckpt()
config_example = {
'encoder': {
'encoder_embed_dim': 512,
'encoder_layer_num': 6,
'encoder_ffn_embed_dim': [3072, 3072, 3072, 3072, 3072, 3072],
'encoder_self_attention_heads': [8, 8, 8, 8, 8, 4],
'encoder_n_experts': [1,2, 1, 1, 1, 1]
},
'decoder': {
'decoder_embed_dim': 512,
'decoder_layer_num': 5,
'decoder_ffn_embed_dim': [2048, 3072, 3072, 3072, 1024],
'decoder_self_attention_heads': [4, 8, 8, 4, 4],
'decoder_ende_attention_heads': [4, 8, 8, 4, 4],
'decoder_arbitrary_ende_attn': [-1, 1, 1, 1, 1],
'decoder_n_experts': [1, 2, 2, 2, 2]
}
}
#predict = predictor.predict_lat(config_example)
#print(f'Example config: {config_example}')
#print(f'Example latency: {predict}')