MNIST without training the weights

[zengqg]

Is the code

"python model.py mnist256test -e 1"

performed to test WANN on MNIST without training the weights?

But I just get an accuracy of 0.0568.

[agaier]

Are you sure everything is setup right? When I call this in the WANN_tool directory I get

But that command is to use the trained weights. If you are interested in untrained weights, you could use the sweep command to look at accuracies of different single values:

So even with a single weight of 0 you should be getting 18%

[zengqg]

I run the code:

To my knowledge, this code will load the model the weights from "champions/mnist.out".
If adding the arugment -f zoo/mnist256.wann.json, it will load the weights from "zoo/mnist256.wann.json".
So, what is the difference between the weights from the out file and json file?

After reading your response, I also try:

which means the models architecture with or without adding"-f zoo/mnist256.wann.json" is the same, the only difference is the weights.
Therefore, could you tell me why the result of using the weights from "champions/mnist.out" is so bad?

[agaier]

"champions/mnist.out" is only the topology, in a giant square weight matrix. every weight is set to 0. Well...almost. it looks like the initial weights were still set, though they are normally rewritten before evaluation (every non-nan weight is set to the shared weight). This is why it is so bad, being a bit worse than chance, it looks like an almost completely inactive ANN -- except for a couple pixels that are still hooked up to the output layer.