david-klindt / nips2017 Goto Github PK

In the fig5.ipynb,there is one line code:"Data=np.load('/gpfs01/bethge/home/dklindt/David/publish/data.npz')"
Obviously,this is a local path,so when I run the code,I can't load the data file,and neither can I find corresponding data file in the github repositories.Can you please fix the code,or add the data file needed for reproducing fig 5?

Hi,

I'm trying to do reproduce fig 3 using your code, with Tensorflow 1.4.1 and Python 3.6. However, there are several problems.

1. fig3.ipynb is not generated using the same training.py as on the GitHub. For example, in the GitHub version of fig3.ipynb, the return value of training.train has 10 elements. However, in the GitHub version of training.py, there are 13 elements on return. There are also other differences, such as out of range error for dropout. etc #3
2. Even with all these fixed, I cannot get the results on GitHub. attached is my result of running the code. The train/val curves look very different than those on GitHub. any idea why?

Can I reproduce Table 1 using data without using a database? I used the populate.py file you provided to import data into the database, but most of the data is empty.

In the paper you have mentioned:
We set this pixel to the standard deviation of the neuron’s response (because
the output of the convolutional layer has unit variance) and initialized the rest of the mask randomly
from a Gaussian N(0,0.001). We initialized the convolution kernels randomly from N(0,0.01) and the feature weights from N(1/K,0.01).

In the figure3 python notebook however:
init_scales = np.array([[0,.01],[0,.001],[1,.1]])#Mean/SD for Kernel,Mask,Weights
for K = 1.

There is a difference between the feature Weight inits mentioned in paper and code. Which is the correct one?

What is the intuition for using dropout before every conv layer in fig3?
Also the dropout rate is set to reg[3] while reg is defined as reg = [0,.001,0]#regularization parameters: L2 Kernels, L1 Mask, L1 Weights
Please clarify this. Thanks.