Upload eat_accelerate_in_30_minites.md
Browse files- eat_accelerate_in_30_minites.md +595 -0
eat_accelerate_in_30_minites.md
ADDED
@@ -0,0 +1,595 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
# 30分钟吃掉accelerate模型训练加速工具
|
3 |
+
|
4 |
+
|
5 |
+
accelerate 是huggingface开源的一个方便将pytorch模型迁移到 GPU/multi-GPUs/TPU/fp16/bf16 模式下训练的小巧工具。
|
6 |
+
|
7 |
+
和标准的 pytorch 方法相比,使用accelerate 进行多GPU DDP模式/TPU/fp16/bf16 训练你的模型变得非常简单,而且训练速度非常快。
|
8 |
+
|
9 |
+
官方范例:https://github.com/huggingface/accelerate/tree/main/examples
|
10 |
+
|
11 |
+
本文将以一个图片分类模型为例,演示在accelerate的帮助下使用pytorch编写一套可以在 cpu/单GPU/多GPU(DDP)模式/TPU 下通用的训练代码。
|
12 |
+
|
13 |
+
在我们的演示范例中,在kaggle的双GPU环境下,双GPU(DDP)模式是单GPU训练速度的1.6倍,加速效果非常明显。
|
14 |
+
|
15 |
+
|
16 |
+
|
17 |
+
|
18 |
+
DP和DDP的区别
|
19 |
+
|
20 |
+
* DP(DataParallel):实现简单但更慢。只能单机多卡使用。GPU分成server节点和worker节点,有负载不均衡。
|
21 |
+
|
22 |
+
* DDP(DistributedDataParallel):更快但实现麻烦。可单机多卡也可多机多卡。各个GPU是平等的,无负载不均衡。
|
23 |
+
|
24 |
+
参考文章:《pytorch中的分布式训练之DP VS DDP》https://zhuanlan.zhihu.com/p/356967195
|
25 |
+
|
26 |
+
|
27 |
+
```python
|
28 |
+
#从git安装最新的accelerate仓库
|
29 |
+
!pip install git+https://github.com/huggingface/accelerate
|
30 |
+
```
|
31 |
+
|
32 |
+
|
33 |
+
|
34 |
+
kaggle 源码:https://www.kaggle.com/code/lyhue1991/kaggle-ddp-tpu-examples
|
35 |
+
|
36 |
+
|
37 |
+
|
38 |
+
## 一,使用 CPU/单GPU 训练你的pytorch模型
|
39 |
+
|
40 |
+
|
41 |
+
当系统存在GPU时,accelerate 会自动使用GPU训练你的pytorch模型,否则会使用CPU训练模型。
|
42 |
+
|
43 |
+
```python
|
44 |
+
import os,PIL
|
45 |
+
import numpy as np
|
46 |
+
from torch.utils.data import DataLoader, Dataset
|
47 |
+
import torch
|
48 |
+
from torch import nn
|
49 |
+
|
50 |
+
import torchvision
|
51 |
+
from torchvision import transforms
|
52 |
+
import datetime
|
53 |
+
|
54 |
+
#======================================================================
|
55 |
+
# import accelerate
|
56 |
+
from accelerate import Accelerator
|
57 |
+
from accelerate.utils import set_seed
|
58 |
+
#======================================================================
|
59 |
+
|
60 |
+
|
61 |
+
def create_dataloaders(batch_size=64):
|
62 |
+
transform = transforms.Compose([transforms.ToTensor()])
|
63 |
+
|
64 |
+
ds_train = torchvision.datasets.MNIST(root="./mnist/",train=True,download=True,transform=transform)
|
65 |
+
ds_val = torchvision.datasets.MNIST(root="./mnist/",train=False,download=True,transform=transform)
|
66 |
+
|
67 |
+
dl_train = torch.utils.data.DataLoader(ds_train, batch_size=batch_size, shuffle=True,
|
68 |
+
num_workers=2,drop_last=True)
|
69 |
+
dl_val = torch.utils.data.DataLoader(ds_val, batch_size=batch_size, shuffle=False,
|
70 |
+
num_workers=2,drop_last=True)
|
71 |
+
return dl_train,dl_val
|
72 |
+
|
73 |
+
|
74 |
+
def create_net():
|
75 |
+
net = nn.Sequential()
|
76 |
+
net.add_module("conv1",nn.Conv2d(in_channels=1,out_channels=512,kernel_size = 3))
|
77 |
+
net.add_module("pool1",nn.MaxPool2d(kernel_size = 2,stride = 2))
|
78 |
+
net.add_module("conv2",nn.Conv2d(in_channels=512,out_channels=256,kernel_size = 5))
|
79 |
+
net.add_module("pool2",nn.MaxPool2d(kernel_size = 2,stride = 2))
|
80 |
+
net.add_module("dropout",nn.Dropout2d(p = 0.1))
|
81 |
+
net.add_module("adaptive_pool",nn.AdaptiveMaxPool2d((1,1)))
|
82 |
+
net.add_module("flatten",nn.Flatten())
|
83 |
+
net.add_module("linear1",nn.Linear(256,128))
|
84 |
+
net.add_module("relu",nn.ReLU())
|
85 |
+
net.add_module("linear2",nn.Linear(128,10))
|
86 |
+
return net
|
87 |
+
|
88 |
+
|
89 |
+
|
90 |
+
def training_loop(epochs = 5,
|
91 |
+
lr = 1e-3,
|
92 |
+
batch_size= 1024,
|
93 |
+
ckpt_path = "checkpoint.pt",
|
94 |
+
mixed_precision="no", #'fp16' or 'bf16'
|
95 |
+
):
|
96 |
+
|
97 |
+
train_dataloader, eval_dataloader = create_dataloaders(batch_size)
|
98 |
+
model = create_net()
|
99 |
+
|
100 |
+
|
101 |
+
optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
|
102 |
+
lr_scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer=optimizer, max_lr=25*lr,
|
103 |
+
epochs=epochs, steps_per_epoch=len(train_dataloader))
|
104 |
+
|
105 |
+
#======================================================================
|
106 |
+
# initialize accelerator and auto move data/model to accelerator.device
|
107 |
+
set_seed(42)
|
108 |
+
accelerator = Accelerator(mixed_precision=mixed_precision)
|
109 |
+
accelerator.print(f'device {str(accelerator.device)} is used!')
|
110 |
+
model, optimizer,lr_scheduler, train_dataloader, eval_dataloader = accelerator.prepare(
|
111 |
+
model, optimizer,lr_scheduler, train_dataloader, eval_dataloader)
|
112 |
+
#======================================================================
|
113 |
+
|
114 |
+
|
115 |
+
for epoch in range(epochs):
|
116 |
+
model.train()
|
117 |
+
for step, batch in enumerate(train_dataloader):
|
118 |
+
features,labels = batch
|
119 |
+
preds = model(features)
|
120 |
+
loss = nn.CrossEntropyLoss()(preds,labels)
|
121 |
+
|
122 |
+
#======================================================================
|
123 |
+
#attention here!
|
124 |
+
accelerator.backward(loss) #loss.backward()
|
125 |
+
#======================================================================
|
126 |
+
|
127 |
+
optimizer.step()
|
128 |
+
lr_scheduler.step()
|
129 |
+
optimizer.zero_grad()
|
130 |
+
|
131 |
+
|
132 |
+
model.eval()
|
133 |
+
accurate = 0
|
134 |
+
num_elems = 0
|
135 |
+
|
136 |
+
for _, batch in enumerate(eval_dataloader):
|
137 |
+
features,labels = batch
|
138 |
+
with torch.no_grad():
|
139 |
+
preds = model(features)
|
140 |
+
predictions = preds.argmax(dim=-1)
|
141 |
+
|
142 |
+
#======================================================================
|
143 |
+
#gather data from multi-gpus (used when in ddp mode)
|
144 |
+
predictions = accelerator.gather_for_metrics(predictions)
|
145 |
+
labels = accelerator.gather_for_metrics(labels)
|
146 |
+
#======================================================================
|
147 |
+
|
148 |
+
accurate_preds = (predictions==labels)
|
149 |
+
num_elems += accurate_preds.shape[0]
|
150 |
+
accurate += accurate_preds.long().sum()
|
151 |
+
|
152 |
+
eval_metric = accurate.item() / num_elems
|
153 |
+
|
154 |
+
#======================================================================
|
155 |
+
#print logs and save ckpt
|
156 |
+
accelerator.wait_for_everyone()
|
157 |
+
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
|
158 |
+
accelerator.print(f"epoch【{epoch}】@{nowtime} --> eval_metric= {100 * eval_metric:.2f}%")
|
159 |
+
net_dict = accelerator.get_state_dict(model)
|
160 |
+
accelerator.save(net_dict,ckpt_path+"_"+str(epoch))
|
161 |
+
#======================================================================
|
162 |
+
|
163 |
+
#training_loop(epochs = 5,lr = 1e-3,batch_size= 1024,ckpt_path = "checkpoint.pt",
|
164 |
+
# mixed_precision="no")
|
165 |
+
|
166 |
+
```
|
167 |
+
|
168 |
+
```python
|
169 |
+
training_loop(epochs = 5,lr = 1e-4,batch_size= 1024,
|
170 |
+
ckpt_path = "checkpoint.pt",
|
171 |
+
mixed_precision="no") #mixed_precision='fp16' or 'bf16'
|
172 |
+
```
|
173 |
+
|
174 |
+
```
|
175 |
+
|
176 |
+
device cuda is used!
|
177 |
+
epoch【0】@2023-01-15 12:06:45 --> eval_metric= 95.20%
|
178 |
+
epoch【1】@2023-01-15 12:07:01 --> eval_metric= 96.79%
|
179 |
+
epoch【2】@2023-01-15 12:07:17 --> eval_metric= 98.47%
|
180 |
+
epoch【3】@2023-01-15 12:07:34 --> eval_metric= 98.78%
|
181 |
+
epoch【4】@2023-01-15 12:07:51 --> eval_metric= 98.87%
|
182 |
+
|
183 |
+
```
|
184 |
+
|
185 |
+
|
186 |
+
|
187 |
+
## 二,使用多GPU DDP模式训练你的pytorch模型
|
188 |
+
|
189 |
+
|
190 |
+
Kaggle中右边settings 中的 ACCELERATOR选择 GPU T4x2。
|
191 |
+
|
192 |
+
|
193 |
+
### 1,设置config
|
194 |
+
|
195 |
+
```python
|
196 |
+
import os
|
197 |
+
from accelerate.utils import write_basic_config
|
198 |
+
write_basic_config() # Write a config file
|
199 |
+
os._exit(0) # Restart the notebook to reload info from the latest config file
|
200 |
+
|
201 |
+
```
|
202 |
+
|
203 |
+
```python
|
204 |
+
# or answer some question to create a config
|
205 |
+
#!accelerate config
|
206 |
+
```
|
207 |
+
|
208 |
+
```python
|
209 |
+
# %load /root/.cache/huggingface/accelerate/default_config.yaml
|
210 |
+
{
|
211 |
+
"compute_environment": "LOCAL_MACHINE",
|
212 |
+
"deepspeed_config": {},
|
213 |
+
"distributed_type": "MULTI_GPU",
|
214 |
+
"downcast_bf16": false,
|
215 |
+
"dynamo_backend": "NO",
|
216 |
+
"fsdp_config": {},
|
217 |
+
"machine_rank": 0,
|
218 |
+
"main_training_function": "main",
|
219 |
+
"megatron_lm_config": {},
|
220 |
+
"mixed_precision": "no",
|
221 |
+
"num_machines": 1,
|
222 |
+
"num_processes": 2,
|
223 |
+
"rdzv_backend": "static",
|
224 |
+
"same_network": false,
|
225 |
+
"use_cpu": false
|
226 |
+
}
|
227 |
+
|
228 |
+
```
|
229 |
+
|
230 |
+
### 2,训练代码
|
231 |
+
|
232 |
+
|
233 |
+
与之前代码完全一致。
|
234 |
+
|
235 |
+
如果是脚本方式启动,需要将训练代码写入到脚本文件中,如cv_example.py
|
236 |
+
|
237 |
+
```python
|
238 |
+
%%writefile cv_example.py
|
239 |
+
import os,PIL
|
240 |
+
import numpy as np
|
241 |
+
from torch.utils.data import DataLoader, Dataset
|
242 |
+
import torch
|
243 |
+
from torch import nn
|
244 |
+
|
245 |
+
import torchvision
|
246 |
+
from torchvision import transforms
|
247 |
+
import datetime
|
248 |
+
|
249 |
+
#======================================================================
|
250 |
+
# import accelerate
|
251 |
+
from accelerate import Accelerator
|
252 |
+
from accelerate.utils import set_seed
|
253 |
+
#======================================================================
|
254 |
+
|
255 |
+
|
256 |
+
def create_dataloaders(batch_size=64):
|
257 |
+
transform = transforms.Compose([transforms.ToTensor()])
|
258 |
+
|
259 |
+
ds_train = torchvision.datasets.MNIST(root="./mnist/",train=True,download=True,transform=transform)
|
260 |
+
ds_val = torchvision.datasets.MNIST(root="./mnist/",train=False,download=True,transform=transform)
|
261 |
+
|
262 |
+
dl_train = torch.utils.data.DataLoader(ds_train, batch_size=batch_size, shuffle=True,
|
263 |
+
num_workers=2,drop_last=True)
|
264 |
+
dl_val = torch.utils.data.DataLoader(ds_val, batch_size=batch_size, shuffle=False,
|
265 |
+
num_workers=2,drop_last=True)
|
266 |
+
return dl_train,dl_val
|
267 |
+
|
268 |
+
|
269 |
+
def create_net():
|
270 |
+
net = nn.Sequential()
|
271 |
+
net.add_module("conv1",nn.Conv2d(in_channels=1,out_channels=512,kernel_size = 3))
|
272 |
+
net.add_module("pool1",nn.MaxPool2d(kernel_size = 2,stride = 2))
|
273 |
+
net.add_module("conv2",nn.Conv2d(in_channels=512,out_channels=256,kernel_size = 5))
|
274 |
+
net.add_module("pool2",nn.MaxPool2d(kernel_size = 2,stride = 2))
|
275 |
+
net.add_module("dropout",nn.Dropout2d(p = 0.1))
|
276 |
+
net.add_module("adaptive_pool",nn.AdaptiveMaxPool2d((1,1)))
|
277 |
+
net.add_module("flatten",nn.Flatten())
|
278 |
+
net.add_module("linear1",nn.Linear(256,128))
|
279 |
+
net.add_module("relu",nn.ReLU())
|
280 |
+
net.add_module("linear2",nn.Linear(128,10))
|
281 |
+
return net
|
282 |
+
|
283 |
+
|
284 |
+
|
285 |
+
def training_loop(epochs = 5,
|
286 |
+
lr = 1e-3,
|
287 |
+
batch_size= 1024,
|
288 |
+
ckpt_path = "checkpoint.pt",
|
289 |
+
mixed_precision="no", #'fp16' or 'bf16'
|
290 |
+
):
|
291 |
+
|
292 |
+
train_dataloader, eval_dataloader = create_dataloaders(batch_size)
|
293 |
+
model = create_net()
|
294 |
+
|
295 |
+
|
296 |
+
optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
|
297 |
+
lr_scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer=optimizer, max_lr=25*lr,
|
298 |
+
epochs=epochs, steps_per_epoch=len(train_dataloader))
|
299 |
+
|
300 |
+
#======================================================================
|
301 |
+
# initialize accelerator and auto move data/model to accelerator.device
|
302 |
+
set_seed(42)
|
303 |
+
accelerator = Accelerator(mixed_precision=mixed_precision)
|
304 |
+
accelerator.print(f'device {str(accelerator.device)} is used!')
|
305 |
+
model, optimizer,lr_scheduler, train_dataloader, eval_dataloader = accelerator.prepare(
|
306 |
+
model, optimizer,lr_scheduler, train_dataloader, eval_dataloader)
|
307 |
+
#======================================================================
|
308 |
+
|
309 |
+
|
310 |
+
for epoch in range(epochs):
|
311 |
+
model.train()
|
312 |
+
for step, batch in enumerate(train_dataloader):
|
313 |
+
features,labels = batch
|
314 |
+
preds = model(features)
|
315 |
+
loss = nn.CrossEntropyLoss()(preds,labels)
|
316 |
+
|
317 |
+
#======================================================================
|
318 |
+
#attention here!
|
319 |
+
accelerator.backward(loss) #loss.backward()
|
320 |
+
#======================================================================
|
321 |
+
|
322 |
+
optimizer.step()
|
323 |
+
lr_scheduler.step()
|
324 |
+
optimizer.zero_grad()
|
325 |
+
|
326 |
+
|
327 |
+
model.eval()
|
328 |
+
accurate = 0
|
329 |
+
num_elems = 0
|
330 |
+
|
331 |
+
for _, batch in enumerate(eval_dataloader):
|
332 |
+
features,labels = batch
|
333 |
+
with torch.no_grad():
|
334 |
+
preds = model(features)
|
335 |
+
predictions = preds.argmax(dim=-1)
|
336 |
+
|
337 |
+
#======================================================================
|
338 |
+
#gather data from multi-gpus (used when in ddp mode)
|
339 |
+
predictions = accelerator.gather_for_metrics(predictions)
|
340 |
+
labels = accelerator.gather_for_metrics(labels)
|
341 |
+
#======================================================================
|
342 |
+
|
343 |
+
accurate_preds = (predictions==labels)
|
344 |
+
num_elems += accurate_preds.shape[0]
|
345 |
+
accurate += accurate_preds.long().sum()
|
346 |
+
|
347 |
+
eval_metric = accurate.item() / num_elems
|
348 |
+
|
349 |
+
#======================================================================
|
350 |
+
#print logs and save ckpt
|
351 |
+
accelerator.wait_for_everyone()
|
352 |
+
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
|
353 |
+
accelerator.print(f"epoch【{epoch}】@{nowtime} --> eval_metric= {100 * eval_metric:.2f}%")
|
354 |
+
net_dict = accelerator.get_state_dict(model)
|
355 |
+
accelerator.save(net_dict,ckpt_path+"_"+str(epoch))
|
356 |
+
#======================================================================
|
357 |
+
|
358 |
+
training_loop(epochs = 5,lr = 1e-4,batch_size= 1024,ckpt_path = "checkpoint.pt",
|
359 |
+
mixed_precision="no") #mixed_precision='fp16' or 'bf16'
|
360 |
+
|
361 |
+
```
|
362 |
+
|
363 |
+
### 3,执行代码
|
364 |
+
|
365 |
+
|
366 |
+
**方式1,在notebook中启动**
|
367 |
+
|
368 |
+
```python
|
369 |
+
from accelerate import notebook_launcher
|
370 |
+
#args = (5,1e-4,1024,'checkpoint.pt','no')
|
371 |
+
args = dict(epochs = 5,
|
372 |
+
lr = 1e-4,
|
373 |
+
batch_size= 1024,
|
374 |
+
ckpt_path = "checkpoint.pt",
|
375 |
+
mixed_precision="no").values()
|
376 |
+
notebook_launcher(training_loop, args, num_processes=2)
|
377 |
+
|
378 |
+
|
379 |
+
```
|
380 |
+
|
381 |
+
```
|
382 |
+
Launching training on 2 GPUs.
|
383 |
+
device cuda:0 is used!
|
384 |
+
epoch【0】@2023-01-15 12:10:48 --> eval_metric= 89.18%
|
385 |
+
epoch【1】@2023-01-15 12:10:58 --> eval_metric= 97.20%
|
386 |
+
epoch【2】@2023-01-15 12:11:08 --> eval_metric= 98.03%
|
387 |
+
epoch【3】@2023-01-15 12:11:19 --> eval_metric= 98.16%
|
388 |
+
epoch【4】@2023-01-15 12:11:30 --> eval_metric= 98.32%
|
389 |
+
```
|
390 |
+
|
391 |
+
|
392 |
+
|
393 |
+
**方式2,accelerate方式执行脚本**
|
394 |
+
|
395 |
+
```python
|
396 |
+
!accelerate launch ./cv_example.py
|
397 |
+
```
|
398 |
+
|
399 |
+
```
|
400 |
+
device cuda:0 is used!
|
401 |
+
epoch【0】@2023-02-03 11:38:02 --> eval_metric= 91.79%
|
402 |
+
epoch【1】@2023-02-03 11:38:13 --> eval_metric= 97.22%
|
403 |
+
epoch【2】@2023-02-03 11:38:22 --> eval_metric= 98.18%
|
404 |
+
epoch【3】@2023-02-03 11:38:32 --> eval_metric= 98.33%
|
405 |
+
epoch【4】@2023-02-03 11:38:43 --> eval_metric= 98.38%
|
406 |
+
```
|
407 |
+
|
408 |
+
|
409 |
+
**方式3,torch方式执行脚本**
|
410 |
+
|
411 |
+
```python
|
412 |
+
# or traditional pytorch style
|
413 |
+
!python -m torch.distributed.launch --nproc_per_node 2 --use_env ./cv_example.py
|
414 |
+
```
|
415 |
+
|
416 |
+
```
|
417 |
+
device cuda:0 is used!
|
418 |
+
epoch【0】@2023-01-15 12:18:26 --> eval_metric= 94.79%
|
419 |
+
epoch【1】@2023-01-15 12:18:37 --> eval_metric= 96.44%
|
420 |
+
epoch【2】@2023-01-15 12:18:48 --> eval_metric= 98.34%
|
421 |
+
epoch【3】@2023-01-15 12:18:59 --> eval_metric= 98.41%
|
422 |
+
epoch【4】@2023-01-15 12:19:10 --> eval_metric= 98.51%
|
423 |
+
```
|
424 |
+
|
425 |
+
|
426 |
+
|
427 |
+
## 三,使用TPU加速你的pytorch模型
|
428 |
+
|
429 |
+
|
430 |
+
Kaggle中右边settings 中的 ACCELERATOR选择 TPU v3-8。
|
431 |
+
|
432 |
+
|
433 |
+
### 1,安装torch_xla
|
434 |
+
|
435 |
+
```python
|
436 |
+
#安装torch_xla支持
|
437 |
+
!pip uninstall -y torch torch_xla
|
438 |
+
!pip install torch==1.8.2+cpu -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
|
439 |
+
!pip install cloud-tpu-client==0.10 https://storage.googleapis.com/tpu-pytorch/wheels/torch_xla-1.8-cp37-cp37m-linux_x86_64.whl
|
440 |
+
```
|
441 |
+
|
442 |
+
```python
|
443 |
+
#从git安装最新的accelerate仓库
|
444 |
+
!pip install git+https://github.com/huggingface/accelerate
|
445 |
+
```
|
446 |
+
|
447 |
+
```python
|
448 |
+
#检查是否成功安装 torch_xla
|
449 |
+
import torch_xla
|
450 |
+
```
|
451 |
+
|
452 |
+
### 2,训练代码
|
453 |
+
|
454 |
+
|
455 |
+
和之前代码完全一样。
|
456 |
+
|
457 |
+
```python
|
458 |
+
import os,PIL
|
459 |
+
import numpy as np
|
460 |
+
from torch.utils.data import DataLoader, Dataset
|
461 |
+
import torch
|
462 |
+
from torch import nn
|
463 |
+
|
464 |
+
import torchvision
|
465 |
+
from torchvision import transforms
|
466 |
+
import datetime
|
467 |
+
|
468 |
+
#======================================================================
|
469 |
+
# import accelerate
|
470 |
+
from accelerate import Accelerator
|
471 |
+
from accelerate.utils import set_seed
|
472 |
+
#======================================================================
|
473 |
+
|
474 |
+
|
475 |
+
def create_dataloaders(batch_size=64):
|
476 |
+
transform = transforms.Compose([transforms.ToTensor()])
|
477 |
+
|
478 |
+
ds_train = torchvision.datasets.MNIST(root="./mnist/",train=True,download=True,transform=transform)
|
479 |
+
ds_val = torchvision.datasets.MNIST(root="./mnist/",train=False,download=True,transform=transform)
|
480 |
+
|
481 |
+
dl_train = torch.utils.data.DataLoader(ds_train, batch_size=batch_size, shuffle=True,
|
482 |
+
num_workers=2,drop_last=True)
|
483 |
+
dl_val = torch.utils.data.DataLoader(ds_val, batch_size=batch_size, shuffle=False,
|
484 |
+
num_workers=2,drop_last=True)
|
485 |
+
return dl_train,dl_val
|
486 |
+
|
487 |
+
|
488 |
+
def create_net():
|
489 |
+
net = nn.Sequential()
|
490 |
+
net.add_module("conv1",nn.Conv2d(in_channels=1,out_channels=512,kernel_size = 3))
|
491 |
+
net.add_module("pool1",nn.MaxPool2d(kernel_size = 2,stride = 2))
|
492 |
+
net.add_module("conv2",nn.Conv2d(in_channels=512,out_channels=256,kernel_size = 5))
|
493 |
+
net.add_module("pool2",nn.MaxPool2d(kernel_size = 2,stride = 2))
|
494 |
+
net.add_module("dropout",nn.Dropout2d(p = 0.1))
|
495 |
+
net.add_module("adaptive_pool",nn.AdaptiveMaxPool2d((1,1)))
|
496 |
+
net.add_module("flatten",nn.Flatten())
|
497 |
+
net.add_module("linear1",nn.Linear(256,128))
|
498 |
+
net.add_module("relu",nn.ReLU())
|
499 |
+
net.add_module("linear2",nn.Linear(128,10))
|
500 |
+
return net
|
501 |
+
|
502 |
+
|
503 |
+
|
504 |
+
def training_loop(epochs = 5,
|
505 |
+
lr = 1e-3,
|
506 |
+
batch_size= 1024,
|
507 |
+
ckpt_path = "checkpoint.pt",
|
508 |
+
mixed_precision="no", #fp16' or 'bf16'
|
509 |
+
):
|
510 |
+
|
511 |
+
train_dataloader, eval_dataloader = create_dataloaders(batch_size)
|
512 |
+
model = create_net()
|
513 |
+
|
514 |
+
|
515 |
+
optimizer = torch.optim.AdamW(params=model.parameters(), lr=lr)
|
516 |
+
lr_scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer=optimizer, max_lr=25*lr,
|
517 |
+
epochs=epochs, steps_per_epoch=len(train_dataloader))
|
518 |
+
|
519 |
+
#======================================================================
|
520 |
+
# initialize accelerator and auto move data/model to accelerator.device
|
521 |
+
set_seed(42)
|
522 |
+
accelerator = Accelerator(mixed_precision=mixed_precision)
|
523 |
+
accelerator.print(f'device {str(accelerator.device)} is used!')
|
524 |
+
model, optimizer,lr_scheduler, train_dataloader, eval_dataloader = accelerator.prepare(
|
525 |
+
model, optimizer,lr_scheduler, train_dataloader, eval_dataloader)
|
526 |
+
#======================================================================
|
527 |
+
|
528 |
+
|
529 |
+
for epoch in range(epochs):
|
530 |
+
model.train()
|
531 |
+
for step, batch in enumerate(train_dataloader):
|
532 |
+
features,labels = batch
|
533 |
+
preds = model(features)
|
534 |
+
loss = nn.CrossEntropyLoss()(preds,labels)
|
535 |
+
|
536 |
+
#======================================================================
|
537 |
+
#attention here!
|
538 |
+
accelerator.backward(loss) #loss.backward()
|
539 |
+
#======================================================================
|
540 |
+
|
541 |
+
optimizer.step()
|
542 |
+
lr_scheduler.step()
|
543 |
+
optimizer.zero_grad()
|
544 |
+
|
545 |
+
|
546 |
+
model.eval()
|
547 |
+
accurate = 0
|
548 |
+
num_elems = 0
|
549 |
+
|
550 |
+
for _, batch in enumerate(eval_dataloader):
|
551 |
+
features,labels = batch
|
552 |
+
with torch.no_grad():
|
553 |
+
preds = model(features)
|
554 |
+
predictions = preds.argmax(dim=-1)
|
555 |
+
|
556 |
+
#======================================================================
|
557 |
+
#gather data from multi-gpus (used when in ddp mode)
|
558 |
+
predictions = accelerator.gather_for_metrics(predictions)
|
559 |
+
labels = accelerator.gather_for_metrics(labels)
|
560 |
+
#======================================================================
|
561 |
+
|
562 |
+
accurate_preds = (predictions==labels)
|
563 |
+
num_elems += accurate_preds.shape[0]
|
564 |
+
accurate += accurate_preds.long().sum()
|
565 |
+
|
566 |
+
eval_metric = accurate.item() / num_elems
|
567 |
+
|
568 |
+
#======================================================================
|
569 |
+
#print logs and save ckpt
|
570 |
+
accelerator.wait_for_everyone()
|
571 |
+
nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
|
572 |
+
accelerator.print(f"epoch【{epoch}】@{nowtime} --> eval_metric= {100 * eval_metric:.2f}%")
|
573 |
+
net_dict = accelerator.get_state_dict(model)
|
574 |
+
accelerator.save(net_dict,ckpt_path+"_"+str(epoch))
|
575 |
+
#======================================================================
|
576 |
+
|
577 |
+
#training_loop(epochs = 5,lr = 1e-3,batch_size= 1024,ckpt_path = "checkpoint.pt",
|
578 |
+
# mixed_precision="no") #mixed_precision='fp16' or 'bf16'
|
579 |
+
|
580 |
+
```
|
581 |
+
|
582 |
+
### 3,启动训练
|
583 |
+
|
584 |
+
```python
|
585 |
+
from accelerate import notebook_launcher
|
586 |
+
#args = (5,1e-4,1024,'checkpoint.pt','no')
|
587 |
+
args = dict(epochs = 5,
|
588 |
+
lr = 1e-4,
|
589 |
+
batch_size= 1024,
|
590 |
+
ckpt_path = "checkpoint.pt",
|
591 |
+
mixed_precision="no").values()
|
592 |
+
notebook_launcher(training_loop, args, num_processes=8)
|
593 |
+
|
594 |
+
```
|
595 |
+
|