-
Notifications
You must be signed in to change notification settings - Fork 0
/
Trainer.py
184 lines (148 loc) · 7.83 KB
/
Trainer.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
import os
import math
import json
import logging
from tqdm import tqdm
import numpy as np
import pandas as pd
import torch
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import _LRScheduler
import torch.optim.lr_scheduler as lr_scheduler
import wandb
logger = logging.getLogger(__name__)
class TrainerConfig:
def __init__(self, **kwargs):
for k, v in kwargs.items():
setattr(self, k, v)
class WarmupThenCosineAnnealingLR(_LRScheduler):
def __init__(self, optimizer, total_steps, warmup_steps, last_epoch=-1):
self.total_steps = total_steps
self.warmup_steps = warmup_steps
self.cosine_annealing_steps = total_steps - warmup_steps
super(WarmupThenCosineAnnealingLR, self).__init__(optimizer, last_epoch)
def get_lr(self):
if self.last_epoch < self.warmup_steps:
# Linear warmup
return [base_lr * self.last_epoch / self.warmup_steps for base_lr in self.base_lrs]
else:
# Cosine annealing
elapsed_steps = self.last_epoch - self.warmup_steps
return [base_lr * (1 + math.cos(math.pi * elapsed_steps / self.cosine_annealing_steps)) / 2
for base_lr in self.base_lrs]
class Trainer:
def __init__(self, model, trainset, evalset, train_config):
self.model = model
self.trainset = trainset
self.evalset = evalset
self.config = train_config
wandb.init(project="UT-runs", name=f'Train_{"UT" if train_config.act else "Vanilla"}_model_\
{train_config.max_epoch}_epoch_{train_config.train_batch_size}_batch_{train_config.learning_rate:.0e}_LR\
{"_"+str(train_config.ponder_penalty)+"_ponder_penalty" if train_config.act else ""}_{train_config.seed}_seed')
self.device = torch.cuda.current_device() if torch.cuda.is_available() else torch.device("cpu")
self.model = self.model.to(self.device)
def save_checkpoints(self, ckpt_id):
model = self.model
ckpt_folder = self.config.ckpt_path
torch.save(model.state_dict(), f"{ckpt_folder}/{ckpt_id}.pth")
def generate_text(self, model, num_tokens):
idx = torch.zeros((1,1), dtype=torch.long).to(self.device)
if self.config.act:
token_ids, ponder_time, token_ponder_dict = model.generate(idx, num_tokens)
decoded_token_ponder_dict = {generate_index: (self.config.tokenizer.decode(token_id), ponder_time)
for generate_index, (token_id, ponder_time) in token_ponder_dict.items()}
else:
token_ids = model.generate(idx, num_tokens)
text = self.config.tokenizer.decode(token_ids.squeeze())
return (text, ponder_time, decoded_token_ponder_dict) if self.config.act else text
def train(self):
config = self.config
model = self.model
optimizer = model.UT_optimizer(config)
lr_steps = int(len(self.trainset) / config.train_batch_size * config.max_epoch)
#scheduler = lr_scheduler.CosineAnnealingLR(optimizer, lr_steps)
scheduler = WarmupThenCosineAnnealingLR(optimizer, total_steps=lr_steps, warmup_steps=int(0.1*lr_steps))
def train_loop(train_dataloader, epoch_idx=1):
model.train()
for itr, (x,y) in tqdm(enumerate(train_dataloader), total=len(train_dataloader), desc='Train'):
x = x.to(self.device)
y = y.to(self.device)
optimizer.zero_grad()
if config.act:
_, loss, (_, n_updates)= model(x, y)
if float(config.ponder_penalty) != 0.0:
ponder_cost = n_updates.mean() * config.ponder_penalty
loss = loss + ponder_cost
else:
_, loss = model(x, y)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
scheduler.step()
train_metrics = {"train/train_loss": loss, "train/train_lr": scheduler.get_last_lr()[0]}
if config.act:
train_metrics.update({"train/train_ponder": n_updates.mean().item()})
wandb.log(train_metrics)
if itr%1000 == 0:
if config.act:
generated_text, ponder_time, token_ponder_dict = self.generate_text(model, num_tokens=config.num_generated_tokens)
else:
generated_text = self.generate_text(model, num_tokens=config.num_generated_tokens)
state_generated_text = {"epoch":epoch_idx,
"model": "UT" if config.act else "Vanilla",
"ponder_penalty" : config.ponder_penalty if config.act else 0,
"ponder_time" : ponder_time if config.act else config.num_layers,
"generated_text": generated_text,
"token_ponder_dict": token_ponder_dict if config.act else {"all_tokens":config.num_layers},
"train_itr": itr}
try:
if os.path.exists(config.generatation_save_path):
with open(config.generatation_save_path, 'r') as file:
data = json.load(file)
data.append(state_generated_text)
else:
data = [state_generated_text]
with open(config.generatation_save_path, 'w') as file:
json.dump(data, file, indent=4)
except IOError as e:
print(f"Error writing to JSON file: {e}")
def eval_loop(eval_dataloader):
model.eval()
losses = []
for _, (x, y) in tqdm(enumerate(eval_dataloader), total=len(eval_dataloader), desc='Eval'):
x = x.to(self.device)
y = y.to(self.device)
if config.act:
_, loss, (_, n_updates)= model(x, y)
else:
_, loss = model(x, y)
losses.append(loss.item())
val_metrics = {"val/val_loss": loss}
if config.act:
val_metrics.update({"val/val_ponder": n_updates.mean().item()})
wandb.log(val_metrics)
return float(np.mean(losses))
train_dataloader = DataLoader(
self.trainset,
batch_size = config.train_batch_size,
num_workers = config.num_workers,
drop_last = True,
)
eval_dataloader = DataLoader(
self.evalset,
batch_size = config.eval_batch_size,
num_workers = config.num_workers,
drop_last= True
)
best_loss = float('inf')
for epoch in range(config.max_epoch):
logger.info(f"===============Epoch:{epoch+1}/{config.max_epoch}=============")
epoch_idx = (epoch+1)
train_loop(train_dataloader, epoch_idx=epoch_idx)
eval_loss = eval_loop(eval_dataloader)
goodModel = eval_loss < best_loss
if config.ckpt_path is not None and goodModel:
best_loss = eval_loss
self.save_checkpoints(f"{config.max_epoch}epoch_best_model_{config.train_batch_size}batch{'_'+str(config.ponder_penalty)+'ponder_penalty' if config.act else ''}_{config.learning_rate:.0e}LR_{config.seed}Seed")
self.save_checkpoints(f"{config.max_epoch}epoch_last_model_{config.train_batch_size}batch{'_'+str(config.ponder_penalty)+'ponder_penalty' if config.act else ''}_{config.learning_rate:.0e}LR_{config.seed}Seed")
wandb.finish()