-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathtrain.py
More file actions
191 lines (158 loc) · 7.19 KB
/
train.py
File metadata and controls
191 lines (158 loc) · 7.19 KB
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
import pickle
from tqdm import tqdm
import numpy as np
import torch
from matplotlib import pyplot as plt
from models.tokenizer import Tokenizer, Encoder, Decoder, EncoderDecoderConfig
from models.world_model import WorldModel, TransformerConfig
from torch.utils.data.distributed import DistributedSampler
from torch.nn.parallel import DistributedDataParallel as DDP
import torch.distributed as dist
from torch.utils.data import TensorDataset, DataLoader
import torch.multiprocessing as mp
from utils import compute_loss
class Trainer:
def __init__(self, vocab_size, rank, world_size, dir_name) -> None:
self.rank = rank
self.world_size = world_size
self.device = torch.device(f"cuda:{rank}")
self.vocab_size = vocab_size
self.dir_name = dir_name
self.batch_size = 32
self.epochs = 100000
self.learning_rate = 0.0001
self.max_grad_norm = 10.0
self.test_num = 32
self.scaler = torch.cuda.amp.GradScaler()
self.train_size = 26000
self.num_gpus = torch.cuda.device_count()
self.setup(rank, world_size)
def setup(self, rank, world_size):
dist.init_process_group(
backend='nccl',
init_method='tcp://127.0.0.1:3456',
world_size=world_size,
rank=rank)
def cleanup(self):
dist.destroy_process_group()
def load_dataset(self, train_size):
obs = np.load("X_obs_resized.npy")
obs = torch.from_numpy(obs).permute(0, 1, 4, 2, 3)
act = np.load("X_act.npy")
print(obs.shape)
print(act.shape)
obs_test = obs[train_size:]
act_test = act[train_size:]
obs_train = obs[:train_size]
act_train = act[:train_size]
train_dataset = TensorDataset(torch.tensor(obs_train), torch.tensor(act_train))
test_dataset = TensorDataset(torch.tensor(obs_test), torch.tensor(act_test))
train_sampler = DistributedSampler(train_dataset, num_replicas=self.world_size, rank=self.rank)
test_sampler = DistributedSampler(test_dataset, num_replicas=self.world_size, rank=self.rank)
train_loader = DataLoader(train_dataset, sampler=train_sampler, pin_memory=True, num_workers=4 * self.num_gpus, batch_size=self.batch_size)
test_loader = DataLoader(test_dataset, sampler=test_sampler, pin_memory=True, num_workers=4 * self.num_gpus, batch_size=self.batch_size)
return train_loader, test_loader
def build_tokenizer(self):
config = EncoderDecoderConfig(
resolution=256,
in_channels=3,
z_channels=256,
ch=128,
ch_mult=[1, 1, 1, 2, 2, 4],
num_res_blocks=2,
out_ch=3,
dropout=0.0,
attn_resolutions=[16],
)
encoder = Encoder(config)
decoder = Decoder(config)
tokenizer = Tokenizer(
vocab_size=self.vocab_size, embed_dim=1024, encoder=encoder, decoder=decoder
)
tokenizer_state_dict = torch.load(
"checkpoint/tokenizer.pt", map_location=self.device
)
tokenizer.load_state_dict(tokenizer_state_dict)
tokenizer.eval()
return tokenizer
def build_worldmodel(self):
t = TransformerConfig(
tokens_per_block=17,
max_blocks=20,
attention="causal",
num_layers=10,
num_heads=4,
embed_dim=256,
embed_pdrop=0.1,
resid_pdrop=0.1,
attn_pdrop=0.1,
)
vocab_size = 512
world_model = WorldModel(vocab_size, 7, t)
world_model.train()
return world_model
def run(self):
train_loader, test_loader = self.load_dataset(self.train_size)
train_losses = []
test_losses = []
mask_fill = torch.logical_not(torch.cat((torch.ones(self.batch_size, 20)), dim=-1,)).to(self.device, non_blocking=True)
tokenizer = self.build_tokenizer().to(self.device)
tokenizer = torch.compile(tokenizer)
worldmodel = self.build_worldmodel().to(self.device)
worldmodel = DDP(worldmodel, device_ids=[self.rank])
worldmodel = torch.compile(worldmodel)
optimizer = torch.optim.Adam(worldmodel.parameters(), lr=self.learning_rate)
for epoch in range(self.epochs):
print(epoch, "Training.")
loss_total_epoch = []
test_error_total_epoch = []
tqdm_leave = False
if self.rank == 1:
tqdm_leave = True
for x, x_act in tqdm(train_loader, position=self.rank, leave=tqdm_leave, desc=f"{epoch} Training (Rank : {self.rank})"):
optimizer.zero_grad(set_to_none=True)
x = x.to(self.device, non_blocking=True).float() / 255.
x_act = x_act.to(self.device, non_blocking=True)
with torch.cuda.amp.autocast():
losses = compute_loss(worldmodel, x, x_act, mask_fill, tokenizer)
loss_total_step = losses
self.scaler.scale(loss_total_step).backward()
loss_total_epoch.append(loss_total_step.item())
torch.nn.utils.clip_grad_norm_(worldmodel.parameters(), self.max_grad_norm)
self.scaler.step(optimizer)
self.scaler.update()
epoch_train_loss = np.mean(loss_total_epoch)
train_losses.append(epoch_train_loss)
# Wait
torch.distributed.barrier()
print("############## Epoch: ", epoch, ", Train loss: ", epoch_train_loss, ", Test error: ", epoch_test_error, " ##############")
for x, x_act in test_loader:
x = x.to(self.device, non_blocking=True).float() / 255.
x_act = x_act.to(self.device, non_blocking=True)
with torch.cuda.amp.autocast(), torch.no_grad():
losses = compute_loss(x, x_act, mask_fill, tokenizer)
test_error_total_epoch.append(losses.item())
epoch_test_error = np.mean(test_error_total_epoch)
test_losses.append(epoch_test_error)
print("############## Epoch: ", epoch, ", Test loss: ", epoch_test_error, ", Test error: ", epoch_test_error, " ##############")
torch.distributed.barrier()
if epoch % 100 == 0 and self.rank == 0:
plt.close()
plt.plot(train_losses)
plt.savefig("{}/train_losses.png".format(self.dir_name))
plt.close()
plt.plot(test_losses)
plt.savefig("{}/test_losses.png".format(self.dir_name))
torch.save(worldmodel.state_dict(), "{}/world_model.pt".format(self.dir_name))
pickle.dump(train_losses, open("{}/train_losses.pkl".format(self.dir_name), "wb"),)
self.cleanup()
def main_worker(gpu, ngpus_per_node, vocab_size, dir_name):
trainer = Trainer(vocab_size=vocab_size, rank=gpu, world_size=ngpus_per_node, dir_name=dir_name)
trainer.run()
def main():
vocab_size = 512
dir_name = "output"
ngpus_per_node = torch.cuda.device_count()
mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, vocab_size, dir_name))
if __name__ == "__main__":
main()