- Python 3.6.4 |Anaconda
- numpy = 1.14.0
- tensorflow = 1.7.0
- sublime text3
- pycharm = 2018.1.3
- Windows10
- Python = 3.6.4
-
MNIST training: 4 workers: 1. 4 workers: 13750 pictures for each one
2. compared group: group1 (using single data set): 13750 pictures group2 (using full data set): 55000 pictures 10 workers: 1. 10 workers: 5500 pictures for each one 2. compared group: group1 (using single data set): 5500 pictures group2 (using full data set): 55000 pictures validation: 5000 pictures test: 10000 pictures
training model:
Input -> Conv layer(3x3)-> Maxpool layer(2x2)-> Fully_connected layer -> Output layer
input image shape: (1,1,28,28)(1: picture num,1: deepth,28: length ,28: width)
affer conv image shape: (10,1,28,28)
after max pool image shape: (10,1,14,14)
fully connected layer: (1960,128)
output layer: (128,10)
iteration : 1500
epoch : 1
learning rate : 0.5
Worker_NUM : 4
mini_batch_size : 64
optimizer function: stochastic gradient descent
conv layer : W1 = (10,1,3,3) b1 =(10,1)
fully connected layer : W2 = (1960,128) b2 = (1,128)
output layer : W3 = (128,10) b3 = (1,10）
worker_num: 10
accuracy:
training(validation):
worker1: 0.968000
worker2: 0.969800
worker3: 0.973400
worker4: 0.972200
worker5: 0.974200
worker6: 0.969200
worker7: 0.966800
worker8: 0.971400
worker9: 0.971400
worker10: 0.970200
compared groups:
group1 (using single data set) : 0.966000
group2 (using full data set) : 0.975800
testing:
worker1: 0.9677
worker2: 0.9687
worker3: 0.9701
worker4: 0.9706
worker5: 0.9703
worker6: 0.9654
worker7: 0.9658
worker8: 0.9705
worker9: 0.9728
worker10: 0.9681
compared groups :
group1 (using single data set) : 0.9644
group2 (using full data set) : 0.9762
worker_num: 4
accuracy:
training(validation):
worker1: 0.970800
worker2: 0.969600
worker3: 0.972000
worker4: 0.973200
compared groups:
group1 (using single data set) : 0.963600
group2 (using full data set) : 0.981800
testing:
worker1: 0.9674
worker2: 0.9733
worker3: 0.9710
worker4: 0.9720
compared groups :
group1 (using single data set) : 0.9621
group2 (using full data set) : 0.9837
-
If you are on Linux,you can run run_new.sh on the terminal,and then the computer will run the file in folder mnist_4,mnist_5,mnist_6,mnist_7,mnist_8,mnist_9,mnist_10,of cource you go to the specific folder and run the file (run_mnist*.py) seperately.(use command python run_mnist*.py)
-
If you are on Windows,you can run run_new.sh on the gitShell,but it maybe slowly,and you can also go to the specific folder and run the file (run_mnist*.py)(use command python run_mnist*.py) seperately in the folder.(Recommend)
-
The time we record is not the same in the two platform,windows is slower than linux.
-
To get the time :you can copy one file's whole print result and paste it to the file './1.txt'(in the same path with sum_time.py file),and run the file sum_time.py,and the last result is the whole time .
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent