Quick cheatsheet
CIFAR-10 in Nolearn using CNN[1]
Sourse code:
https://app.dominodatalab.com/LeJit/GPU_Example/browse
Download the whole GPU_exmple folder, and follow the tutorial here.
If you are a user at elsa, you can download the modified version (higher accuracy) from server:
$ git clone ssh://140.114.75.138/var/git/nolearn_cifar10_example.git/
The system will start to build ssh-connection, just login.
After downloading, try to run ConvolutionNN.py,
$ cd nolearn_cifar10_example/
$ python ConvloutionNN.py
or try another network MultiLayerPerceptron.py.
$ python MultiLayerPerceptron.py
ConcolutionNN.py
Result (ConvolutionNN.py):
Finished setting up Theano
# Neural Network with 128194 learnable parameters
## Layer information
# name size
--- ------- --------
0 input 3x32x32
1 conv1 16x30x30
2 pool1 16x15x15
3 conv2 32x14x14
4 pool2 32x7x7
5 conv3 64x6x6
6 pool3 64x3x3
7 hidden4 200
8 output 10
epoch trn loss val loss trn/val valid acc dur
------- ---------- ---------- --------- ----------- -----
1 2.12073 1.86161 1.13920 0.32490 2.93s
2 1.71640 1.59117 1.07870 0.42770 2.92s
3 1.52466 1.44750 1.05330 0.47690 2.92s
4 1.40519 1.35891 1.03406 0.51230 2.92s
5 1.31972 1.28742 1.02509 0.54010 2.93s
6 1.24873 1.22473 1.01960 0.56470 2.93s
7 1.18727 1.17548 1.01003 0.58480 2.93s
(etc.)
495 0.00002 4.31504 0.00000 0.69960 2.93s
496 0.00002 4.31551 0.00000 0.69960 2.93s
497 0.00002 4.31599 0.00000 0.69960 2.93s
498 0.00002 4.31646 0.00000 0.69960 2.93s
499 0.00002 4.31693 0.00000 0.69960 2.93s
500 0.00001 4.31740 0.00000 0.69960 2.93s
The accuracy of this network is: 0.69
[1] Tutorial: Faster deep learning with GPUs and Theano (Chinese: 使用GPU和Theano加速深度学习)