-
Notifications
You must be signed in to change notification settings - Fork 2
/
pretrain_MaskCTR_ddp.py
229 lines (198 loc) · 10.1 KB
/
pretrain_MaskCTR_ddp.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
import warnings
warnings.filterwarnings("ignore")
import numpy as np
import pandas as pd
from pretrain_config import create_parser
import torch
from sklearn.model_selection import train_test_split
from sklearn.metrics import log_loss, roc_auc_score
from preprocessing.inputs import SparseFeat, get_feature_names
from torch.cuda.amp import autocast,GradScaler
from dataset import CtrDataset3
import random
from transformers import AutoModel, AutoTokenizer
from tqdm import tqdm
from model.MaskCTR_ddp import MaskCTR
from utils import AvgMeter, get_lr, MetricLogger, SmoothedValue, is_main_process, get_rank
import torch.backends.cudnn as cudnn
import torch.distributed as dist
from torch.utils.data import Dataset, DataLoader
from torch.utils.data.distributed import DistributedSampler
import time
import datetime
import json
def make_train_valid_dfs(struct_data_path, text_data_path, seed, data_source):
struct_data = pd.read_csv(struct_data_path)
text_data = pd.read_table(text_data_path,names = ["content"],header=None)
if cfg.sample:
struct_data,_ = train_test_split(struct_data,test_size= (1-cfg.sample_ration) ,random_state= seed)
text_data,_ = train_test_split(text_data,test_size= (1-cfg.sample_ration) ,random_state= seed)
text_data['label'] = struct_data['label']
train_struct, test_struct = struct_data.iloc[:int(len(struct_data) * 0.9)].copy(),\
struct_data.iloc[int(len(struct_data) * 0.9):].copy()
train_text, test_text = text_data.iloc[:int(len(text_data) * 0.9)].copy(), text_data.iloc[int(len(
text_data) * 0.9):].copy()
return train_struct,test_struct,train_text,test_text, struct_data
def process_struct_data(data_source, train, test, data):
embedding_dim = 32
if (data_source == 'movielens'):
sparse_features = ['user_id', 'gender', 'age', 'occupation', 'zipcode','movie_id', 'title','genre']
elif (data_source == 'bookcrossing'):
sparse_features = ['User ID', 'Location', 'Age', 'ISBN', 'Book title', 'Author', 'Publication year', 'Publisher']
elif (data_source == 'goodreads'):
sparse_features = ['User ID','Book ID', 'Book title', 'Book genres' ,'Average rating', 'Number of book reviews', 'Author ID', 'Author name',
'Number of pages','eBook flag', 'Format', 'Publisher', 'Publication year', 'Work ID', 'Media type']
sparse_feature_columns = [SparseFeat(feat, data[feat].nunique(), embedding_dim=embedding_dim)
for i, feat in enumerate(sparse_features)]
linear_feature_columns = sparse_feature_columns
dnn_feature_columns = sparse_feature_columns
train_model_input = {name: train[name] for name in sparse_features }
test_model_input = {name: test[name] for name in sparse_features }
return linear_feature_columns, dnn_feature_columns, train_model_input, test_model_input
def build_loaders(struct_input, text_input, linear_feature_columns,dnn_feature_columns,tokenizer, mode):
dataset = CtrDataset3(
struct_input,
text_input["content"].values,
text_input["label"].values,
linear_feature_columns=linear_feature_columns,
dnn_feature_columns=dnn_feature_columns,
tokenizer=tokenizer,
max_length=cfg.max_length,
mask_ratio=cfg.mask_ratio
)
sampler = DistributedSampler(dataset, shuffle=True if mode == "train" else False)
dataloader = DataLoader(
dataset,
batch_size=cfg.batch_size,
num_workers=cfg.num_workers,
sampler=sampler,
pin_memory = True,
)
return dataloader, sampler
def dynamic_mask(inputs, mask_ratio, same_column):
batch_size = inputs['rec_data'].shape[0]
num_field = inputs['rec_data'].shape[1]
if same_column: # mask same column
mask_index = inputs['mask_text_index']
else:
mask_num = int(num_field * mask_ratio)
mask_index = torch.rand((batch_size, num_field), device=inputs['rec_data'].device)
mask_index = torch.argsort(mask_index, dim=-1)[:, :mask_num]
inputs['mask_rec_index'] = mask_index
inputs['mask_rec_label'] = torch.gather(inputs['rec_data'], 1, mask_index)
inputs['mask_rec_data'] = torch.scatter(inputs['rec_data'], 1, mask_index, 0)
return inputs
def train_epoch(model, train_loader, optimizer, lr_scheduler, step):
metric_logger = MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', SmoothedValue(window_size=50, fmt='{value:.6f}'))
for i in range(loss_num):
metric_logger.add_meter(f'loss_{i}', SmoothedValue(window_size=50, fmt='{value:.4f}'))
header = 'Train Epoch: [{}]'.format(epoch)
print_freq = 50
logger_data = {}
scaler = GradScaler()
if cfg.mixed_precision:
for i, batch in enumerate(metric_logger.log_every(train_loader, print_freq, header)):
batch = dynamic_mask(batch, cfg.mask_ratio, cfg.same_column)
batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
optimizer.zero_grad()
# loss_ita, loss_mfm, loss_mlm = model(batch)
with autocast():
loss, loss_list = model(batch)
assert len(loss_list) == loss_num
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
if step == "batch":
lr_scheduler.step()
for i in range(loss_num):
logger_data[f'loss_{i}'] = loss_list[i].item()
logger_data['lr'] = optimizer.param_groups[0]["lr"]
metric_logger.update(logger_data)
else:
for i, batch in enumerate(metric_logger.log_every(train_loader, print_freq, header)):
batch = dynamic_mask(batch, cfg.mask_ratio, cfg.same_column)
batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
optimizer.zero_grad()
# loss_ita, loss_mfm, loss_mlm = model(batch)
loss, loss_list = model(batch)
assert len(loss_list) == loss_num
loss.backward()
optimizer.step()
if step == "batch":
lr_scheduler.step()
for i in range(loss_num):
logger_data[f'loss_{i}'] = loss_list[i].item()
logger_data['lr'] = optimizer.param_groups[0]["lr"]
metric_logger.update(logger_data)
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger.global_avg())
return {k: "{:.3f}".format(meter.global_avg) for k, meter in metric_logger.meters.items()}
if __name__ == '__main__':
cfg = create_parser()
# fix seed
seed = cfg.seed + get_rank()
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
cudnn.benchmark = True
torch.cuda.set_device(cfg.local_rank)
device = torch.device("cuda", cfg.local_rank)
dist.init_process_group(backend='nccl')
data_type = cfg.dataset
model = None
if not cfg.use_mask_loss:
loss_num = 1
elif cfg.use_mfm and cfg.use_mlm:
loss_num = 3
else:
loss_num =2
train_struct,valid_struct,train_text,valid_text, struct_data = make_train_valid_dfs(cfg.struct_path, cfg.text_path, seed, data_type)
linear_feature_columns, dnn_feature_columns, train_struct_input, valid_struct_input = \
process_struct_data(data_type,train_struct,valid_struct,struct_data)
tokenizer = AutoTokenizer.from_pretrained(cfg.text_tokenizer, local_files_only=True)
num_added_tokens = tokenizer.add_tokens('[val]')
train_loader, train_sampler = build_loaders(train_struct_input, train_text,
linear_feature_columns, dnn_feature_columns,tokenizer, mode='train')
valid_loader, _ = build_loaders(valid_struct_input, valid_text,
linear_feature_columns,dnn_feature_columns,tokenizer, mode ='valid')
with open(cfg.meta_path) as fh:
meta_data = json.load(fh)
total_feature_num = meta_data['feature_num']
if is_main_process():
print('world size', dist.get_world_size())
print('train len', len(train_struct), ', valid_len', len(valid_struct))
print("len_tokenizer:", len(tokenizer))
print("total feature num: ", total_feature_num)
model = MaskCTR(cfg, cfg.rec_embedding_dim, cfg.text_embedding_dim, cfg.text_encoder_model,
struct_linear_feature_columns = linear_feature_columns,
struct_dnn_feature_columns= dnn_feature_columns,
struct_feature_num=total_feature_num,
text_tokenizer_num=len(tokenizer)-1).to(device)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[cfg.local_rank], find_unused_parameters=True)
optimizer = torch.optim.AdamW(
model.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay
)
# lr_scheduler = torch.optim.lr_scheduler.LinearLR(optimizer,start_factor=1.0, end_factor=.0, total_iters=cfg.epochs)
lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,T_max=cfg.epochs, eta_min=0)
step = "epoch"
best_loss = float('inf')
restore_path = cfg.output_prefix_path+ str(data_type)+ f'{cfg.temperature}_{cfg.use_mfm}_{cfg.use_mlm}_{cfg.epochs}_{cfg.lr}_'
for epoch in range(cfg.epochs):
# shuffle
train_sampler.set_epoch(epoch)
model.train()
train_stats = train_epoch(model, train_loader, optimizer, lr_scheduler, step)
lr_scheduler.step()
if is_main_process():
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
'epoch': epoch,
}
torch.save(model.module, restore_path +"best.pt")
torch.save(model.module.rec_encoder,restore_path +"rec_best.pt")
torch.save(model.module.text_encoder, restore_path +"text_best.pt")
torch.save(model.module.rec_projection, restore_path +"rec_projection.pt")
torch.save(model.module.text_projection, restore_path +"text_projection.pt")
with open(restore_path+ "log.txt","a") as f:
f.write(json.dumps(log_stats) + "\n")
dist.barrier()