vgg16.npy instead of vgg16.pkl

[ksd777]

Hello,
Do you have an updated training.py file where you demonstrate how to correctly load and use parameters from npy pretrained file? Or maybe you can provide path to download vgg16.pkl?

[ccny-ros-pkg]

Please check demo:
with np.load('/home/robolab/project/deepProjectBridge/YL_50iterV2.npz') as f:
param_values = [f['arr_%d' % i] for i in range(len(f.files))]
lasagne.layers.set_all_param_values(output_layer, param_values)

Just use this one to load the pre-trained the model.

In order to train, in the training.py, please comment this line:

Load model weights and metadata

d = pickle.load(open('vgg16.pkl'))

Then everything should work

[ksd777]

Ok, thanks!
There were necessary to these lines to the build_model function:
net['fc7_2048'] = DenseLayer(net['fc6'], num_units=2048)
net['fc8_7'] = DenseLayer(net['fc7_2048'], num_units=2, nonlinearity=None)
net['prob'] = NonlinearityLayer(net['fc8_7'], softmax)

#net['prob'] = NonlinearityLayer(net['fc8'], softmax)

Also, there are three broken image files into the training list:
spallSubImageForTraining/imageClustersV130B130/accepted/accept186007.jpg
spallSubImageForTraining/imageClustersV130B130/accepted/accept186008.jpg
spallSubImageForTraining/imageClustersV130B130/accepted/accept186009.jpg

Also could you please let me know what parameters I have to use in demo.py to reproduce your result on example image?
i.e. I found out that with regionSize=25 result is following:

regionSize=55 gives:

the original 200 gives:

Also the result does not seem to be stable - it is easy to get different result with the same parameters.

My console output from demo.py:

Using gpu device 0: GeForce GTX 1080 (CNMeM is disabled, cuDNN 5005)
/usr/local/lib/python2.7/dist-packages/theano/tensor/signal/downsample.py:6: UserWarning: downsample module has been moved to the theano.tensor.signal.pool module.
"downsample module has been moved to the theano.tensor.signal.pool module.")
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
32
successfully..., from new model 500
direct testing

[ccny-ros-pkg]

Please try 60 or 50, also, for egionSize=55, it looks really weird. I will update a repeatable version using pytorch lately and let your know.

[ccny-ros-pkg]

Please check the updated pytorch version, it should be much easier for you. https://github.com/ccny-ros-pkg/pytorch_Concrete_Inspection.git