Skip to content

M
models

PaddlePaddle / models
接近 2 年 前同步成功

通知 230
Star 6828
Fork 2962
M
models
未验证 提交 53723856 编写于 作者: O overlordmax 提交者： GitHub
浏览文件
操作

Mmoe 04162304 (#4539) 

* fix some bugs

* edit README.md
上级 282e4890
隐藏空白更改
内联 并排
Showing with 127 addition and 121 deletion
PaddleRec/multi-task/MMoE/README.md
浏览文件 @ 53723856
......@@ -4,6 +4,7 @@
```
├── README.md # 文档
├── requirements.txt # 需要的安装包
├── mmoe_train.py # mmoe模型脚本
├── utils # 通用函数
├── args # 参数脚本
......@@ -21,15 +22,15 @@
1.income
> best：0.94856
> max_mmoe_test_auc_income：0.94937
>
> mean：0.944105
> mean_mmoe_test_auc_income：0.94465
2.marital
> best：0.99403
> max_mmoe_test_auc_marital：0.99419
>
> mean：0.99324
> mean_mmoe_test_auc_marital：0.99324
本项目支持GPU和CPU两种单机训练环境。
......@@ -74,15 +75,16 @@ GPU环境
```sh
CUDA_VISIBLE_DEVICES=0 python train_mmoe.py --use_gpu 1 \ #使用gpu训练
--train_path data/data24913/train_data/\ #训练数据路径
--test_path data/data24913/test_data/\ #测试数据路径
--feature_size 499\ #设置特征的维度
--batch_size 32\ #设置batch_size大小
--expert_num 8\ #设置expert数量
--gate_num 2\ #设置gate数量
--expert_size 16\ #设置expert网络大小
--tower_size 8\ #设置tower网络大小
--epochs 400 #设置epoch轮次
--train_data_path 'train_data'\ #训练数据路径
--test_data_path 'test_data'\ #测试数据路径
--model_dir 'model_dir'\ #模型保存地址
--feature_size 499\ #设置特征的维度
--batch_size 32\ #设置batch_size大小
--expert_num 8\ #设置expert数量
--gate_num 2\ #设置gate数量
--expert_size 16\ #设置expert网络大小
--tower_size 8\ #设置tower网络大小
--epochs 400 #设置epoch轮次
```
修改脚本的可执行权限并运行
......@@ -97,15 +99,16 @@ CPU环境
```sh
python train_mmoe.py --use_gpu 0 \ #使用cpu训练
--train_path data/data24913/train_data/\ #训练数据路径
--test_path data/data24913/test_data/\ #测试数据路径
--feature_size 499\ #设置特征的维度
--batch_size 32\ #设置batch_size大小
--expert_num 8\ #设置expert数量
--gate_num 2\ #设置gate数量
--expert_size 16\ #设置expert网络大小
--tower_size 8\ #设置tower网络大小
--epochs 400 #设置epoch轮次
--train_data_path 'train_data'\ #训练数据路径
--test_data_path 'test_data'\ #测试数据路径
--model_dir 'model_dir'\ #模型保存地址
--feature_size 499\ #设置特征的维度
--batch_size 32\ #设置batch_size大小
--expert_num 8\ #设置expert数量
--gate_num 2\ #设置gate数量
--expert_size 16\ #设置expert网络大小
--tower_size 8\ #设置tower网络大小
--epochs 400 #设置epoch轮次
```
修改脚本的可执行权限并运行
......@@ -124,6 +127,13 @@ python train_mmoe.py --use_gpu 0 \ #使用cpu训练
epoch设置为100的训练和测试效果如下：
![](./image/mmoe.png)
![](./image/mmoe2.png)
\ No newline at end of file
```text
batch_size:[32],feature_size:[499],expert_num:[8],gate_num[2],expert_size[16],tower_size[8],epochs:[100]
2020-04-16 11:28:06,- INFO - epoch_id: 0,epoch_time: 129.17434 s,loss: 0.62215,train_auc_income: 0.86302,train_auc_marital: 0.92316,test_auc_income: 0.84525,test_auc_marital: 0.98269
2020-04-16 11:30:36,- INFO - epoch_id: 1,epoch_time: 149.79017 s,loss: 0.42484,train_auc_income: 0.90634,train_auc_marital: 0.98418,test_auc_income:
......
2020-04-16 15:31:23,- INFO - epoch_id: 97,epoch_time: 147.07304 s,loss: 0.30267,train_auc_income: 0.94743,train_auc_marital: 0.99430,test_auc_income: 0.94905,test_auc_marital: 0.99414
2020-04-16 15:33:51,- INFO - epoch_id: 98,epoch_time: 148.34412 s,loss: 0.29688,train_auc_income: 0.94736,train_auc_marital: 0.99433,test_auc_income: 0.94846,test_auc_marital: 0.99409
2020-04-16 15:36:21,- INFO - epoch_id: 99,epoch_time: 149.91047 s,loss: 0.31330,train_auc_income: 0.94732,train_auc_marital: 0.99403,test_auc_income: 0.94881,test_auc_marital: 0.99386
2020-04-16 15:36:21,- INFO - mean_mmoe_test_auc_income: 0.94465,mean_mmoe_test_auc_marital 0.99324,max_mmoe_test_auc_income: 0.94937,max_mmoe_test_auc_marital 0.99419
```
PaddleRec/multi-task/MMoE/args.py
浏览文件 @ 53723856
......@@ -30,8 +30,9 @@ def parse_args():
parser.add_argument("--epochs", type=int, default=400, help="epochs")
parser.add_argument("--batch_size", type=int, default=32, help="batch_size")
parser.add_argument('--use_gpu', type=int, default=0, help='whether using gpu')
parser.add_argument('--train_data_path',type=str, default='./data/data24913/train_data/', help="train_data_path")
parser.add_argument('--test_data_path',type=str, default='./data/data24913/test_data/', help="test_data_path")
parser.add_argument('--model_dir',type=str, default='./model_dir', help="model_dir")
parser.add_argument('--train_data_path',type=str, default='./train_data', help="train_data_path")
parser.add_argument('--test_data_path',type=str, default='./test_data', help="test_data_path")
args = parser.parse_args()
return args
......
PaddleRec/multi-task/MMoE/create_data.sh
浏览文件 @ 53723856
......@@ -6,6 +6,8 @@ test_path="data/census-income.test"
train_data_path="train_data/"
test_data_path="test_data/"
pip install -r requirements.txt
wget -P data/ https://archive.ics.uci.edu/ml/machine-learning-databases/census-income-mld/census.tar.gz
tar -zxvf data/census.tar.gz -C data/
......
PaddleRec/multi-task/MMoE/data_preparation.py
浏览文件 @ 53723856
import pandas as pd
import numpy as np
import paddle.fluid as fluid
from sklearn.preprocessing import MinMaxScaler
from args import *
......
PaddleRec/multi-task/MMoE/mmoe_train.py
浏览文件 @ 53723856
......@@ -3,14 +3,10 @@ import pandas as pd
import numpy as np
import paddle
import time
import datetime
import os
import utils
from sklearn.metrics import roc_auc_score
from sklearn.preprocessing import MinMaxScaler
from args import *
import warnings
warnings.filterwarnings("ignore")
#显示所有列
pd.set_option('display.max_columns', None)
def set_zero(var_name,scope=fluid.global_scope(),place=fluid.CPUPlace(),param_type="int64"):
"""
......@@ -118,9 +114,9 @@ loader.set_sample_list_generator(train_reader, places=place)
test_loader = fluid.io.DataLoader.from_generator(feed_list=data_list, capacity=batch_size, iterable=True)
test_loader.set_sample_list_generator(test_reader, places=place)
mean_auc_income = []
mean_auc_marital = []
auc_income_list = []
auc_marital_list = []
mmoe_res_file = open('mmoe_res.txt', 'w',encoding='utf-8')
for epoch in range(epochs):
for var in auc_states_1: # reset auc states
set_zero(var.name,place=place)
......@@ -150,12 +146,21 @@ for epoch in range(epochs):
feed=test_data,
fetch_list=[out_1,out_2,label_1,label_2,auc_income,auc_marital],
return_numpy=True)
mean_auc_income.append(test_auc_1_p)
mean_auc_marital.append(test_auc_2_p)
model_dir = os.path.join(args.model_dir,'epoch_' + str(epoch + 1), "checkpoint")
main_program = fluid.default_main_program()
fluid.io.save(main_program,model_dir)
auc_income_list.append(test_auc_1_p)
auc_marital_list.append(test_auc_2_p)
end = time.time()
print("epoch_id:[%d],epoch_time:[%.5f s],loss:[%.5f],train_auc_income:[%.5f],train_auc_marital:[%.5f],test_auc_income:[%.5f],test_auc_marital:[%.5f]"%
(epoch,end - begin,loss_data,auc_1_p,auc_2_p,test_auc_1_p,test_auc_2_p))
print("mean_auc_income:[%.5f],mean_auc_marital[%.5f]"%(np.mean(mean_auc_income),np.mean(mean_auc_marital)))
time_stamp = datetime.datetime.now()
print("%s,- INFO - epoch_id: %d,epoch_time: %.5f s,loss: %.5f,train_auc_income: %.5f,train_auc_marital: %.5f,test_auc_income: %.5f,test_auc_marital: %.5f"%
(time_stamp.strftime('%Y-%m-%d %H:%M:%S'),epoch,end - begin,loss_data,auc_1_p,auc_2_p,test_auc_1_p,test_auc_2_p))
time_stamp = datetime.datetime.now()
print("%s,- INFO - mean_mmoe_test_auc_income: %.5f,mean_mmoe_test_auc_marital %.5f,max_mmoe_test_auc_income: %.5f,max_mmoe_test_auc_marital %.5f"%(
time_stamp.strftime('%Y-%m-%d %H:%M:%S'),np.mean(auc_income_list),np.mean(auc_marital_list),np.max(auc_income_list),np.max(auc_marital_list)))
......
PaddleRec/multi-task/MMoE/requirements.txt 0 → 100644
浏览文件 @ 53723856
pandas==0.23.4
\ No newline at end of file
PaddleRec/multi-task/MMoE/train_cpu.sh
浏览文件 @ 53723856
python mmoe_train.py --use_gpu 0\
--train_data_path 'train_data'\
--test_data_path 'test_data'\
--feature_size 499\
--batch_size 32\
--expert_num 8\
--gate_num 2\
--expert_size 16\
--tower_size 8\
--epochs 100
--train_data_path 'train_data'\
--test_data_path 'test_data'\
--model_dir 'model_dir'\
--feature_size 499\
--batch_size 32\
--expert_num 8\
--gate_num 2\
--expert_size 16\
--tower_size 8\
--epochs 100
PaddleRec/multi-task/MMoE/train_gpu.sh
浏览文件 @ 53723856
CUDA_VISIBLE_DEVICES=0 python mmoe_train.py --use_gpu 1\
--train_data_path 'train_data'\
--test_data_path 'test_data'\
--model_dir 'model_dir'\
--feature_size 499\
--batch_size 32\
--expert_num 8\
......
PaddleRec/multi-task/MMoE/utils.py
浏览文件 @ 53723856
......@@ -5,18 +5,14 @@ import os
import paddle.fluid as fluid
import io
from itertools import islice
from sklearn.preprocessing import MinMaxScaler
import warnings
##按行读取文件
def reader_creator(file_dir):
def reader():
files = os.listdir(file_dir)
for fi in files:
with io.open(
os.path.join(file_dir, fi), "r", encoding='utf-8') as f:
for l in islice(f, 1, None): ##忽略第一行
for l in islice(f, 1, None):
l = l.strip().split(',')
l = list(map(float, l))
label_income = []
......@@ -36,8 +32,6 @@ def reader_creator(file_dir):
return reader
##读取一个batch
def batch_reader(reader, batch_size):
def batch_reader():
r = reader()
......@@ -47,14 +41,8 @@ def batch_reader(reader, batch_size):
if (len(b) == batch_size):
yield b
b = []
#if len(b) != 0:
# yield b
#
return batch_reader
##准备数据
def prepare_reader(data_path, batch_size):
data_set = reader_creator(data_path)
#random.shuffle(data_set)
return batch_reader(data_set, batch_size)
PaddleRec/multi-task/Share_bottom/README.md
浏览文件 @ 53723856
......@@ -4,6 +4,7 @@
```
├── README.md # 文档
├── requirements.txt # 需要的安装包
├── share_bottom.py # mmoe模型脚本
├── utils # 通用函数
├── args # 参数脚本
......@@ -21,15 +22,15 @@ share_bottom是多任务学习的基本框架，其特点是对于不同的任
1.income
>best：0.94899
>max_sb_test_auc_income：0.94993
>
>mean：0.94402
>mean_sb_test_auc_income： 0.93120
2.marital
> best：0.99394
> max_sb_test_auc_marital：0.99384
>
> mean：0.99311
> mean_sb_test_auc_marital：0.99256
本项目支持GPU和CPU两种单机训练环境。
......@@ -74,8 +75,9 @@ GPU环境
```sh
python share_bottom.py --use_gpu 1\ #使用gpu训练
--train_path data/data24913/train_data/\ #训练数据路径
--test_path data/data24913/test_data/\ #测试数据路径
--train_path 'train_data'\ #训练数据路径
--test_path 'test_data'\ #测试数据路径
--model_dir 'model_dir'\ #模型保存地址
--batch_size 32\ #设置batch_size大小
--feature_size 499\ #设置特征维度
--bottom_size 117\ #设置bottom网络大小
......@@ -96,8 +98,9 @@ CPU环境
```sh
python share_bottom.py --use_gpu 0\ #使用cpu训练
--train_path data/data24913/train_data/\ #训练数据路径
--test_path data/data24913/test_data/\ #测试数据路径
--train_path 'train_data'\ #训练数据路径
--test_path 'test_data'\ #测试数据路径
--model_dir 'model_dir'\ #模型保存地址
--batch_size 32\ #设置batch_size大小
--feature_size 499\ #设置特征维度
--bottom_size 117\ #设置bottom网络大小
......@@ -122,6 +125,14 @@ python share_bottom.py --use_gpu 0\ #使用cpu训练
epoch设置为100的训练和测试效果如下：
![](./image/share_bottom.png)
```text
batch_size:[32],epochs:[100],feature_size:[499],bottom_size:[117],tower_nums:[2],tower_size:[8]
2020-04-16 16:01:04,- INFO - epoch_id: 0,epoch_time: 77.17624 s,loss: 0.62643,train_auc_income: 0.49442,train_auc_marital: 0.93509,test_auc_income: 0.50000,test_auc_marital: 0.93920
2020-04-16 16:02:23,- INFO - epoch_id: 1,epoch_time: 78.84795 s,loss: 0.47955,train_auc_income: 0.49721,train_auc_marital: 0.98118,test_auc_income: 0.50000,test_auc_marital: 0.98804
2020-04-16 16:03:43,- INFO - epoch_id: 2,epoch_time: 79.67485 s,loss:
......
2020-04-16 18:22:36,- INFO - epoch_id: 98,epoch_time: 85.56907 s,loss: 0.30696,train_auc_income: 0.94701,train_auc_marital: 0.99425,test_auc_income: 0.94919,test_auc_marital: 0.99376
2020-04-16 18:24:02,- INFO - epoch_id: 99,epoch_time: 86.08858 s,loss: 0.29395,train_auc_income: 0.94736,train_auc_marital: 0.99422,test_auc_income: 0.94908,test_auc_marital: 0.99383
2020-04-16 18:24:02,- INFO - mean_sb_test_auc_income: 0.93120,mean_sb_test_auc_marital 0.99256,max_sb_test_auc_income: 0.94993,max_sb_test_auc_marital 0.99384
```
![](./image/share_bottom2.png)
\ No newline at end of file
PaddleRec/multi-task/Share_bottom/args.py
浏览文件 @ 53723856
......@@ -29,16 +29,9 @@ def parse_args():
parser.add_argument("--epochs", type=int, default=400, help="epochs")
parser.add_argument("--batch_size", type=int, default=32, help="batch_size")
parser.add_argument('--use_gpu', type=int, default=0, help='whether using gpu')
parser.add_argument(
'--train_data_path',
type=str,
default=' ',
help="train_data_path")
parser.add_argument(
'--test_data_path',
type=str,
default=' ',
help="test_data_path")
parser.add_argument('--train_data_path',type=str,default='train_data',help="train_data_path")
parser.add_argument('--test_data_path',type=str,default='test_data',help="test_data_path")
parser.add_argument('--model_dir',type=str, default='model_dir', help="model_dir")
args = parser.parse_args()
return args
......@@ -49,9 +42,9 @@ def data_preparation_args():
parser.add_argument("--test_path", type=str, default='', help="test_path")
parser.add_argument(
'--train_data_path', type=str, default='', help="train_data_path")
'--train_data_path', type=str, default='train_data', help="train_data_path")
parser.add_argument(
'--test_data_path', type=str, default='', help="test_data_path")
'--test_data_path', type=str, default='test_data', help="test_data_path")
parser.add_argument(
'--validation_data_path',
type=str,
......
PaddleRec/multi-task/Share_bottom/create_data.sh
浏览文件 @ 53723856
......@@ -5,6 +5,7 @@ train_path="data/census-income.data"
test_path="data/census-income.test"
train_data_path="train_data/"
test_data_path="test_data/"
pip install -r requirements.txt
wget -P data/ https://archive.ics.uci.edu/ml/machine-learning-databases/census-income-mld/census.tar.gz
tar -zxvf data/census.tar.gz -C data/
......
PaddleRec/multi-task/Share_bottom/data_preparation.py
浏览文件 @ 53723856
import pandas as pd
import numpy as np
import paddle.fluid as fluid
from sklearn.preprocessing import MinMaxScaler
from args import *
......
PaddleRec/multi-task/Share_bottom/requirements.txt 0 → 100644
浏览文件 @ 53723856
pandas==0.23.4
\ No newline at end of file
PaddleRec/multi-task/Share_bottom/share_bottom.py
浏览文件 @ 53723856
import paddle.fluid as fluid
import pandas as pd
import numpy as np
import paddle
import os
import time
import datetime
import utils
from sklearn.metrics import roc_auc_score
from sklearn.preprocessing import MinMaxScaler
from args import *
import warnings
warnings.filterwarnings("ignore")
#显示所有列
pd.set_option('display.max_columns', None)
def set_zero(var_name,scope=fluid.global_scope(),place=fluid.CPUPlace(),param_type="int64"):
"""
......@@ -106,8 +99,8 @@ loader.set_sample_list_generator(train_reader, places=place)
test_loader = fluid.io.DataLoader.from_generator(feed_list=data_list, capacity=batch_size, iterable=True)
test_loader.set_sample_list_generator(test_reader, places=place)
mean_auc_income = []
mean_auc_marital = []
auc_income_list = []
auc_marital_list = []
for epoch in range(epochs):
begin = time.time()
......@@ -139,13 +132,20 @@ for epoch in range(epochs):
feed=test_data,
fetch_list=[out_1,out_2,label_1,label_2,auc_income,auc_marital],
return_numpy=True)
mean_auc_income.append(test_auc_1_p)
mean_auc_marital.append(test_auc_2_p)
model_dir = os.path.join(args.model_dir,'epoch_' + str(epoch + 1), "checkpoint")
main_program = fluid.default_main_program()
fluid.io.save(main_program,model_dir)
auc_income_list.append(test_auc_1_p)
auc_marital_list.append(test_auc_2_p)
end = time.time()
print("epoch_id:[%d],epoch_time:[%.5f s],loss:[%.5f],train_auc_income:[%.5f],train_auc_marital:[%.5f],test_auc_income:[%.5f],test_auc_marital:[%.5f]"%
(epoch,end - begin,loss_data,auc_1_p,auc_2_p,test_auc_1_p,test_auc_2_p))
print("mean_auc_income:[%.5f],mean_auc_marital[%.5f]"%(np.mean(mean_auc_income),np.mean(mean_auc_marital)))
time_stamp = datetime.datetime.now()
print("%s,- INFO - epoch_id: %d,epoch_time: %.5f s,loss: %.5f,train_auc_income: %.5f,train_auc_marital: %.5f,test_auc_income: %.5f,test_auc_marital: %.5f"%
(time_stamp.strftime('%Y-%m-%d %H:%M:%S'),epoch,end - begin,loss_data,auc_1_p,auc_2_p,test_auc_1_p,test_auc_2_p))
time_stamp = datetime.datetime.now()
print("%s,- INFO - mean_sb_test_auc_income: %.5f,mean_sb_test_auc_marital %.5f,max_sb_test_auc_income: %.5f,max_sb_test_auc_marital %.5f"%(
time_stamp.strftime('%Y-%m-%d %H:%M:%S'),np.mean(auc_income_list),np.mean(auc_marital_list),np.max(auc_income_list),np.max(auc_marital_list)))
......
PaddleRec/multi-task/Share_bottom/train_cpu.sh
浏览文件 @ 53723856
python share_bottom.py --use_gpu 0 \
--epochs 100 \
--train_data_path '../train_data' \
--test_data_path '../test_data' \
--batch_size 16 \
--train_data_path 'train_data' \
--test_data_path 'test_data' \
--model_dir 'model_dir' \
--batch_size 32 \
--feature_size 499 \
--bottom_size 117 \
--tower_nums 2 \
......
PaddleRec/multi-task/Share_bottom/train_gpu.sh
浏览文件 @ 53723856
python share_bottom.py --use_gpu 1 \
--epochs 100 \
--train_data_path '../train_data' \
--test_data_path '../test_data' \
--batch_size 16 \
--train_data_path '.train_data' \
--test_data_path '.test_data' \
--model_dir 'model_dir' \
--batch_size 32 \
--feature_size 499 \
--bottom_size 117 \
--tower_nums 2 \
......
PaddleRec/multi-task/Share_bottom/utils.py
浏览文件 @ 53723856
......@@ -5,18 +5,14 @@ import os
import paddle.fluid as fluid
import io
from itertools import islice
from sklearn.preprocessing import MinMaxScaler
import warnings
##按行读取文件
def reader_creator(file_dir):
def reader():
files = os.listdir(file_dir)
for fi in files:
with io.open(
os.path.join(file_dir, fi), "r", encoding='utf-8') as f:
for l in islice(f, 1, None): ##忽略第一行
for l in islice(f, 1, None):
l = l.strip().split(',')
l = list(map(float, l))
label_income = []
......@@ -36,8 +32,6 @@ def reader_creator(file_dir):
return reader
##读取一个batch
def batch_reader(reader, batch_size):
def batch_reader():
r = reader()
......@@ -49,9 +43,6 @@ def batch_reader(reader, batch_size):
b = []
return batch_reader
##准备数据
def prepare_reader(data_path, batch_size):
data_set = reader_creator(data_path)
#random.shuffle(data_set)
return batch_reader(data_set, batch_size)
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或