Tests

[ardila]

We need tests to make sure our implementation gives the same results as NYs version.

[ardila]

I didn't have time to finish this tonight since I don't know how to call scripts and then check their standard out, but here is something you can test your convnet.py against (gives the same results every time)

python convnet.py --data-path=/home/ardila/archconvnets/dropnn-release/cifar-10-py-colmajor --save-path=/home/ardila/test_run_diego/ --test-range=6 --train-range=1-5 --layer-def=./cifar-layers/layers-80sec.cfg --layer-params=./cifar-layers/layer-params-80sec.cfg --data-provider=cifar-cropped-rand --test-freq=13 --epochs=1

Should give the following output:

Initialized data layer 'data', producing 1728 outputs
Initialized data layer 'labels', producing 1 outputs
Initialized convolutional layer 'conv1', producing 24x24 32-channel output
Initialized max-pooling layer 'pool1', producing 12x12 32-channel output
Initialized convolutional layer 'conv2', producing 12x12 32-channel output
Initialized avg-pooling layer 'pool2', producing 6x6 32-channel output
Initialized convolutional layer 'conv3', producing 6x6 64-channel output
Initialized avg-pooling layer 'pool3', producing 3x3 64-channel output
Initialized fully-connected layer 'fc128', producing 64 outputs
Initialized fully-connected layer 'fc10', producing 10 outputs
Initialized softmax layer 'probs', producing 10 outputs
Initialized logistic regression cost 'logprob'
Initialized neuron layer 'pool1_neuron', producing 4608 outputs
Initialized neuron layer 'conv2_neuron', producing 4608 outputs
Initialized neuron layer 'conv3_neuron', producing 2304 outputs
Initialized neuron layer 'fc128_neuron', producing 64 outputs
=========================
Importing _ConvNet C++ module
============================================
learning rate scale     :  1
Reset Momentum          :  False
Image Rotation & Scaling:  False
============================================
=========================
Training ConvNet
Adaptive Drop Training                              : False [DEFAULT]
Check gradients and quit?                           : 0     [DEFAULT]
Compress checkpoints?                               : 0     [DEFAULT]
Conserve GPU memory (slower)?                       : 0     [DEFAULT]
Convert given conv layers to unshared local         :       
Cropped DP: crop border size                        : 4     [DEFAULT]
Cropped DP: logreg layer name (for --multiview-test):       [DEFAULT]
Cropped DP: test on multiple patches?               : 0     [DEFAULT]
Data batch range: testing                           : 6-6   
Data batch range: training                          : 1-5   
Data path                                           : /home/ardila/archconvnets/dropnn-release/cifar-10-py-colmajor 
Data provider                                       : cifar-cropped-rand 
Enable image rotation and scaling transformation    : False [DEFAULT]
GPU override                                        : -1    [DEFAULT]
Image Size                                          : 0     [DEFAULT]
Layer definition file                               : ./cifar-layers/layers-80sec.cfg 
Layer parameter file                                : ./cifar-layers/layer-params-80sec.cfg 
Learning Rate Scale Factor                          : 1     [DEFAULT]
Load file                                           :       [DEFAULT]
Maximum save file size (MB)                         : 0     [DEFAULT]
Minibatch size                                      : 128   [DEFAULT]
Model File Name                                     :       [DEFAULT]
Number of GPUs                                      : 1     [DEFAULT]
Number of channels in image                         : 3     [DEFAULT]
Number of epochs                                    : 1     
Reset layer momentum                                : False [DEFAULT]
Save path                                           : /home/ardila/test_run_diego/ 
Test and quit?                                      : 0     [DEFAULT]
Test on one batch at a time?                        : 1     [DEFAULT]
Testing frequency                                   : 13    
Unshare weight matrices in given layers             :       
Whether filp training image                         : True  [DEFAULT]
=========================
Running on CUDA device(s) -2
Current time: Wed Oct  9 21:15:27 2013
Saving checkpoints to /home/ardila/test_run_diego/ConvNet__2013-10-09_21.15.25
=========================
1.1... logprob:  2.176140, 0.815800 (1.355 sec)
1.2... logprob:  1.884615, 0.698000 (1.352 sec)
1.3... logprob:  1.761647, 0.646700 (1.349 sec)
1.4... logprob:  1.726771, 0.632500 (1.352 sec)
1.5... logprob:  1.657396, 0.607800 (1.343 sec)
epoch_cost: 9.20657008266
2.1... logprob:  1.612479, 0.583400 (1.350 sec)

[yamins81]

Working on this now.

On Wed, Oct 9, 2013 at 9:16 PM, Diego Ardila [email protected]:

I didn't have time to finish this tonight since I don't know how to call
scripts and then check their standard out, but here is how something you
can test your convnet.py against (gives the same results every time)

python convnet.py --data-path=/home/ardila/archconvnets/dropnn-release/cifar-10-py-colmajor --save-path=/home/ardila/test_run_diego/ --test-range=6 --train-range=1-5 --layer-def=./cifar-layers/layers-80sec.cfg --layer-params=./cifar-layers/layer-params-80sec.cfg --data-provider=cifar-cropped-rand --test-freq=13 --epochs=1

Should give the following output:

Initialized data layer 'data', producing 1728 outputs
Initialized data layer 'labels', producing 1 outputs
Initialized convolutional layer 'conv1', producing 24x24 32-channel output
Initialized max-pooling layer 'pool1', producing 12x12 32-channel output
Initialized convolutional layer 'conv2', producing 12x12 32-channel output
Initialized avg-pooling layer 'pool2', producing 6x6 32-channel output
Initialized convolutional layer 'conv3', producing 6x6 64-channel output
Initialized avg-pooling layer 'pool3', producing 3x3 64-channel output
Initialized fully-connected layer 'fc128', producing 64 outputs
Initialized fully-connected layer 'fc10', producing 10 outputs
Initialized softmax layer 'probs', producing 10 outputs
Initialized logistic regression cost 'logprob'
Initialized neuron layer 'pool1_neuron', producing 4608 outputs
Initialized neuron layer 'conv2_neuron', producing 4608 outputs
Initialized neuron layer 'conv3_neuron', producing 2304 outputs
Initialized neuron layer 'fc128_neuron', producing 64 outputs Importing
_ConvNet C++ module

learning rate scale : 1
Reset Momentum : False
Image Rotation & Scaling: False

Training ConvNet
Adaptive Drop Training : False [DEFAULT]
Check gradients and quit? : 0 [DEFAULT]
Compress checkpoints? : 0 [DEFAULT]
Conserve GPU memory (slower)? : 0 [DEFAULT]
Convert given conv layers to unshared local :

Cropped DP: crop border size : 4 [DEFAULT]
Cropped DP: logreg layer name (for --multiview-test): [DEFAULT]
Cropped DP: test on multiple patches? : 0 [DEFAULT]
Data batch range: testing : 6-6

Data batch range: training : 1-5

Data path : /home/ardila/archconvnets/dropnn-release/cifar-10-py-colmajor
Data provider : cifar-cropped-rand
Enable image rotation and scaling transformation : False [DEFAULT]
GPU override : -1 [DEFAULT]
Image Size : 0 [DEFAULT]
Layer definition file : ./cifar-layers/layers-80sec.cfg
Layer parameter file : ./cifar-layers/layer-params-80sec.cfg
Learning Rate Scale Factor : 1 [DEFAULT]
Load file : [DEFAULT]
Maximum save file size (MB) : 0 [DEFAULT]
Minibatch size : 128 [DEFAULT]
Model File Name : [DEFAULT]
Number of GPUs : 1 [DEFAULT]
Number of channels in image : 3 [DEFAULT]
Number of epochs : 1

Reset layer momentum : False [DEFAULT]
Save path : /home/ardila/test_run_diego/
Test and quit? : 0 [DEFAULT]
Test on one batch at a time? : 1 [DEFAULT]
Testing frequency : 13

Unshare weight matrices in given layers :
Whether filp training image : True [DEFAULT]

Running on CUDA device(s) -2
Current time: Wed Oct 9 21:15:27 2013
Saving checkpoints to
/home/ardila/test_run_diego/ConvNet__2013-10-09_21.15.25

1.1... logprob: 2.176140, 0.815800 (1.355 sec)
1.2... logprob: 1.884615, 0.698000 (1.352 sec)
1.3... logprob: 1.761647, 0.646700 (1.349 sec)
1.4... logprob: 1.726771, 0.632500 (1.352 sec)
1.5... logprob: 1.657396, 0.607800 (1.343 sec)
epoch_cost: 9.20657008266
2.1... logprob: 1.612479, 0.583400 (1.350 sec)

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/5#issuecomment-26022412
.

[ardila]

Maybe we want to test on this model which is trained on CIFAR:
home/ardila/test_run_diego/ConvNet__2013-10-09_21.23.05

[ardila]

This model was trained for longer and posted online by the nyu group
/home/ardila/model_fc128-dcf-50/model_fc128-dcf-50_run12/1500.5